diff --git "a/parse/dev/wtqb7pNL4e/wtqb7pNL4e_middle.json" "b/parse/dev/wtqb7pNL4e/wtqb7pNL4e_middle.json" new file mode 100644--- /dev/null +++ "b/parse/dev/wtqb7pNL4e/wtqb7pNL4e_middle.json" @@ -0,0 +1,48230 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 162, + 70, + 433, + 102 + ], + "lines": [ + { + "bbox": [ + 162, + 70, + 434, + 86 + ], + "spans": [ + { + "bbox": [ + 162, + 70, + 434, + 86 + ], + "score": 1.0, + "content": "Can ChatGPT Assess Human Personalities?", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 190, + 86, + 404, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 86, + 404, + 102 + ], + "score": 1.0, + "content": "A General Evaluation Framework", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 157, + 110, + 429, + 124 + ], + "lines": [ + { + "bbox": [ + 157, + 108, + 432, + 128 + ], + "spans": [ + { + "bbox": [ + 157, + 110, + 244, + 127 + ], + "score": 1.0, + "content": "Haocong Rao1,2", + "type": "text" + }, + { + "bbox": [ + 248, + 108, + 329, + 128 + ], + "score": 1.0, + "content": "Cyril Leung2,3", + "type": "text" + }, + { + "bbox": [ + 335, + 110, + 432, + 126 + ], + "score": 1.0, + "content": "Chunyan Miao1,2∗", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 75, + 119, + 525, + 195 + ], + "lines": [ + { + "bbox": [ + 73, + 123, + 525, + 142 + ], + "spans": [ + { + "bbox": [ + 73, + 126, + 78, + 137 + ], + "score": 0.35, + "content": "^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 78, + 123, + 525, + 142 + ], + "score": 1.0, + "content": "School of Computer Science and Engineering, Nanyang Technological University, Singapore", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 136, + 470, + 156 + ], + "spans": [ + { + "bbox": [ + 126, + 136, + 470, + 156 + ], + "score": 1.0, + "content": "2LILY Research Centre, Nanyang Technological University, Singapore", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 169, + 151, + 429, + 169 + ], + "spans": [ + { + "bbox": [ + 169, + 151, + 429, + 169 + ], + "score": 1.0, + "content": "3Department of Electrical and Computer Engineering", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 191, + 167, + 407, + 181 + ], + "spans": [ + { + "bbox": [ + 191, + 167, + 407, + 181 + ], + "score": 1.0, + "content": "The University of British Columbia, Canada", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 181, + 462, + 196 + ], + "spans": [ + { + "bbox": [ + 136, + 181, + 462, + 196 + ], + "score": 1.0, + "content": "{haocong001,ascymiao}@ntu.edu.sg {cleung}@ece.ubc.ca", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 157, + 213, + 202, + 226 + ], + "lines": [ + { + "bbox": [ + 155, + 212, + 204, + 228 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 204, + 228 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 87, + 238, + 273, + 620 + ], + "lines": [ + { + "bbox": [ + 86, + 238, + 273, + 250 + ], + "spans": [ + { + "bbox": [ + 86, + 238, + 273, + 250 + ], + "score": 1.0, + "content": "Large Language Models (LLMs) especially", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 250, + 273, + 262 + ], + "spans": [ + { + "bbox": [ + 86, + 250, + 273, + 262 + ], + "score": 1.0, + "content": "ChatGPT have produced impressive results in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 263, + 273, + 274 + ], + "spans": [ + { + "bbox": [ + 86, + 263, + 273, + 274 + ], + "score": 1.0, + "content": "various areas, but their potential human-like", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 86, + 275, + 274, + 286 + ], + "spans": [ + { + "bbox": [ + 86, + 275, + 274, + 286 + ], + "score": 1.0, + "content": "psychology is still largely unexplored. Ex-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 85, + 286, + 274, + 298 + ], + "spans": [ + { + "bbox": [ + 85, + 286, + 274, + 298 + ], + "score": 1.0, + "content": "isting works study the virtual personalities of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 85, + 297, + 274, + 311 + ], + "spans": [ + { + "bbox": [ + 85, + 297, + 274, + 311 + ], + "score": 1.0, + "content": "LLMs but rarely explore the possibility of an-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 86, + 310, + 274, + 321 + ], + "spans": [ + { + "bbox": [ + 86, + 310, + 274, + 321 + ], + "score": 1.0, + "content": "alyzing human personalities via LLMs. This", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 85, + 322, + 274, + 334 + ], + "spans": [ + { + "bbox": [ + 85, + 322, + 274, + 334 + ], + "score": 1.0, + "content": "paper presents a generic evaluation framework", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 86, + 334, + 274, + 346 + ], + "spans": [ + { + "bbox": [ + 86, + 334, + 274, + 346 + ], + "score": 1.0, + "content": "for LLMs to assess human personalities based", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 86, + 347, + 274, + 357 + ], + "spans": [ + { + "bbox": [ + 86, + 347, + 274, + 357 + ], + "score": 1.0, + "content": "on Myers–Briggs Type Indicator (MBTI) tests.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 86, + 358, + 273, + 370 + ], + "spans": [ + { + "bbox": [ + 86, + 358, + 273, + 370 + ], + "score": 1.0, + "content": "Specifically, we first devise unbiased prompts", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 85, + 369, + 275, + 383 + ], + "spans": [ + { + "bbox": [ + 85, + 369, + 275, + 383 + ], + "score": 1.0, + "content": "by randomly permuting options in MBTI ques-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 86, + 382, + 274, + 394 + ], + "spans": [ + { + "bbox": [ + 86, + 382, + 274, + 394 + ], + "score": 1.0, + "content": "tions and adopt the average testing result to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 86, + 394, + 275, + 406 + ], + "spans": [ + { + "bbox": [ + 86, + 394, + 275, + 406 + ], + "score": 1.0, + "content": "encourage more impartial answer generation.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 85, + 405, + 275, + 419 + ], + "spans": [ + { + "bbox": [ + 85, + 405, + 275, + 419 + ], + "score": 1.0, + "content": "Then, we propose to replace the subject in ques-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 86, + 417, + 274, + 430 + ], + "spans": [ + { + "bbox": [ + 86, + 417, + 274, + 430 + ], + "score": 1.0, + "content": "tion statements to enable flexible queries and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 85, + 430, + 275, + 442 + ], + "spans": [ + { + "bbox": [ + 85, + 430, + 275, + 442 + ], + "score": 1.0, + "content": "assessments on different subjects from LLMs.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 85, + 441, + 275, + 454 + ], + "spans": [ + { + "bbox": [ + 85, + 441, + 275, + 454 + ], + "score": 1.0, + "content": "Finally, we re-formulate the question instruc-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 86, + 453, + 273, + 465 + ], + "spans": [ + { + "bbox": [ + 86, + 453, + 273, + 465 + ], + "score": 1.0, + "content": "tions in a manner of correctness evaluation to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 85, + 464, + 275, + 478 + ], + "spans": [ + { + "bbox": [ + 85, + 464, + 275, + 478 + ], + "score": 1.0, + "content": "facilitate LLMs to generate clearer responses.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 85, + 477, + 275, + 489 + ], + "spans": [ + { + "bbox": [ + 85, + 477, + 275, + 489 + ], + "score": 1.0, + "content": "The proposed framework enables LLMs to flex-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 85, + 488, + 274, + 502 + ], + "spans": [ + { + "bbox": [ + 85, + 488, + 274, + 502 + ], + "score": 1.0, + "content": "ibly assess personalities of different groups of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 85, + 501, + 274, + 513 + ], + "spans": [ + { + "bbox": [ + 85, + 501, + 274, + 513 + ], + "score": 1.0, + "content": "people. We further propose three evaluation", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 86, + 514, + 274, + 525 + ], + "spans": [ + { + "bbox": [ + 86, + 514, + 274, + 525 + ], + "score": 1.0, + "content": "metrics to measure the consistency, robustness,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 86, + 525, + 274, + 536 + ], + "spans": [ + { + "bbox": [ + 86, + 525, + 274, + 536 + ], + "score": 1.0, + "content": "and fairness of assessment results from state-of-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 86, + 537, + 275, + 549 + ], + "spans": [ + { + "bbox": [ + 86, + 537, + 275, + 549 + ], + "score": 1.0, + "content": "the-art LLMs including ChatGPT and GPT-4.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 86, + 549, + 274, + 562 + ], + "spans": [ + { + "bbox": [ + 86, + 549, + 274, + 562 + ], + "score": 1.0, + "content": "Our experiments reveal ChatGPT’s ability to", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 86, + 561, + 274, + 574 + ], + "spans": [ + { + "bbox": [ + 86, + 561, + 274, + 574 + ], + "score": 1.0, + "content": "assess human personalities, and the average", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 85, + 573, + 273, + 585 + ], + "spans": [ + { + "bbox": [ + 85, + 573, + 273, + 585 + ], + "score": 1.0, + "content": "results demonstrate that it can achieve more", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 86, + 586, + 273, + 596 + ], + "spans": [ + { + "bbox": [ + 86, + 586, + 273, + 596 + ], + "score": 1.0, + "content": "consistent and fairer assessments in spite of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 86, + 597, + 275, + 609 + ], + "spans": [ + { + "bbox": [ + 86, + 597, + 275, + 609 + ], + "score": 1.0, + "content": "lower robustness against prompt biases com-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 86, + 608, + 189, + 621 + ], + "spans": [ + { + "bbox": [ + 86, + 608, + 189, + 621 + ], + "score": 1.0, + "content": "pared with InstructGPT†.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 70, + 643, + 153, + 657 + ], + "lines": [ + { + "bbox": [ + 67, + 642, + 155, + 659 + ], + "spans": [ + { + "bbox": [ + 67, + 642, + 155, + 659 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 69, + 666, + 290, + 733 + ], + "lines": [ + { + "bbox": [ + 69, + 666, + 290, + 680 + ], + "spans": [ + { + "bbox": [ + 69, + 666, + 290, + 680 + ], + "score": 1.0, + "content": "Pre-trained Large Language Models (LLMs) have", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 69, + 680, + 290, + 694 + ], + "spans": [ + { + "bbox": [ + 69, + 680, + 290, + 694 + ], + "score": 1.0, + "content": "been widely used in many applications including", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 69, + 694, + 291, + 706 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 291, + 706 + ], + "score": 1.0, + "content": "translation, storytelling, and chatbots (Devlin et al.,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 69, + 706, + 290, + 720 + ], + "spans": [ + { + "bbox": [ + 69, + 706, + 290, + 720 + ], + "score": 1.0, + "content": "2019; Raffel et al., 2020; Yang et al., 2022; Yuan", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 69, + 721, + 291, + 734 + ], + "spans": [ + { + "bbox": [ + 69, + 721, + 291, + 734 + ], + "score": 1.0, + "content": "et al., 2022; Ouyang et al., 2022; Bubeck et al.,", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 305, + 214, + 525, + 347 + ], + "lines": [ + { + "bbox": [ + 303, + 213, + 526, + 226 + ], + "spans": [ + { + "bbox": [ + 303, + 213, + 526, + 226 + ], + "score": 1.0, + "content": "2023). ChatGPT (Ouyang et al., 2022) and its", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 226, + 526, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 526, + 240 + ], + "score": 1.0, + "content": "enhanced version GPT-4 are currently recognized", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 241, + 525, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 525, + 253 + ], + "score": 1.0, + "content": "as the most capable chatbots, which can perform", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 255, + 525, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 525, + 267 + ], + "score": 1.0, + "content": "context-aware conversations, challenge incorrect", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 269, + 525, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 525, + 280 + ], + "score": 1.0, + "content": "premises, and reject inappropriate requests with", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 303, + 282, + 527, + 293 + ], + "spans": [ + { + "bbox": [ + 303, + 282, + 527, + 293 + ], + "score": 1.0, + "content": "a vast knowledge base and human-centered fine-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 304, + 295, + 526, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 526, + 307 + ], + "score": 1.0, + "content": "tuning. These advantages make them well-suited", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 303, + 308, + 527, + 321 + ], + "spans": [ + { + "bbox": [ + 303, + 308, + 527, + 321 + ], + "score": 1.0, + "content": "for a variety of real-world scenarios such as busi-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 303, + 322, + 527, + 334 + ], + "spans": [ + { + "bbox": [ + 303, + 322, + 527, + 334 + ], + "score": 1.0, + "content": "ness consultation and educational services (Zhai,", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 335, + 515, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 515, + 348 + ], + "score": 1.0, + "content": "2022; van Dis et al., 2023; Bubeck et al., 2023).", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 53.5 + }, + { + "type": "text", + "bbox": [ + 305, + 352, + 525, + 594 + ], + "lines": [ + { + "bbox": [ + 315, + 351, + 527, + 366 + ], + "spans": [ + { + "bbox": [ + 315, + 351, + 527, + 366 + ], + "score": 1.0, + "content": "Recent studies have revealed that LLMs may pos-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 365, + 526, + 379 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 526, + 379 + ], + "score": 1.0, + "content": "sess human-like self-improvement and reasoning", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 379, + 526, + 391 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 526, + 391 + ], + "score": 1.0, + "content": "characteristics (Huang et al., 2022; Bubeck et al.,", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 303, + 391, + 527, + 406 + ], + "spans": [ + { + "bbox": [ + 303, + 391, + 491, + 406 + ], + "score": 1.0, + "content": "2023). The latest GPT series can pass over", + "type": "text" + }, + { + "bbox": [ + 491, + 392, + 513, + 404 + ], + "score": 0.9, + "content": "90 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 513, + 391, + 527, + 406 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 406, + 525, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 525, + 419 + ], + "score": 1.0, + "content": "Theory of Mind (ToM) tasks with strong analysis", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 420, + 526, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 526, + 432 + ], + "score": 1.0, + "content": "and decision-making capabilities (Kosinski, 2023;", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 433, + 527, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 527, + 446 + ], + "score": 1.0, + "content": "Zhuo et al., 2023; Moghaddam and Honey, 2023).", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 446, + 525, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 525, + 460 + ], + "score": 1.0, + "content": "In this context, LLMs are increasingly assumed to", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 303, + 460, + 526, + 474 + ], + "spans": [ + { + "bbox": [ + 303, + 460, + 526, + 474 + ], + "score": 1.0, + "content": "have virtual personalities and psychologies, which", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 474, + 525, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 525, + 487 + ], + "score": 1.0, + "content": "plays an essential role in guiding their responses", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 488, + 525, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 525, + 500 + ], + "score": 1.0, + "content": "and interaction patterns (Jiang et al., 2022). Based", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 501, + 526, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 526, + 513 + ], + "score": 1.0, + "content": "on this assumption, a few works (Li et al., 2022;", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 303, + 515, + 525, + 527 + ], + "spans": [ + { + "bbox": [ + 303, + 515, + 525, + 527 + ], + "score": 1.0, + "content": "Jiang et al., 2022; Karra et al., 2022; Caron and", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 527, + 527, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 527, + 542 + ], + "score": 1.0, + "content": "Srivastava, 2022; Miotto et al., 2022) apply psy-", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 304, + 541, + 527, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 527, + 555 + ], + "score": 1.0, + "content": "chological tests such as Big Five Factors (Digman,", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 554, + 525, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 525, + 569 + ], + "score": 1.0, + "content": "1990) to evaluate their pseudo personalities (e.g.,", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 568, + 526, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 526, + 581 + ], + "score": 1.0, + "content": "behavior tendency), so as to detect societal and eth-", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 304, + 582, + 524, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 524, + 595 + ], + "score": 1.0, + "content": "ical risks (e.g., racial biases) in their applications.", + "type": "text" + } + ], + "index": 76 + } + ], + "index": 67.5 + }, + { + "type": "text", + "bbox": [ + 305, + 599, + 525, + 773 + ], + "lines": [ + { + "bbox": [ + 316, + 599, + 525, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 599, + 525, + 611 + ], + "score": 1.0, + "content": "Although existing works have investigated the", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 303, + 611, + 526, + 626 + ], + "spans": [ + { + "bbox": [ + 303, + 611, + 526, + 626 + ], + "score": 1.0, + "content": "personality traits of LLMs, they rarely explored", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 626, + 527, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 527, + 639 + ], + "score": 1.0, + "content": "whether LLMs can assess human personalities.", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 639, + 526, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 526, + 653 + ], + "score": 1.0, + "content": "This open problem can be the key to verifying the", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 652, + 527, + 667 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 527, + 667 + ], + "score": 1.0, + "content": "ability of LLMs to perform psychological (e.g., per-", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 303, + 666, + 526, + 680 + ], + "spans": [ + { + "bbox": [ + 303, + 666, + 526, + 680 + ], + "score": 1.0, + "content": "sonality psychology) analyses and revealing their", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 303, + 680, + 526, + 693 + ], + "spans": [ + { + "bbox": [ + 303, + 680, + 526, + 693 + ], + "score": 1.0, + "content": "potential understanding of humans, i.e., “How do", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 303, + 693, + 527, + 707 + ], + "spans": [ + { + "bbox": [ + 303, + 693, + 527, + 707 + ], + "score": 1.0, + "content": "LLMs think about humans?”. Specifically, assess-", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "score": 1.0, + "content": "ing human personalities from the point of LLMs", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 303, + 720, + 526, + 734 + ], + "spans": [ + { + "bbox": [ + 303, + 720, + 526, + 734 + ], + "score": 1.0, + "content": "(1) enables us to access the perception of LLMs", + "type": "text" + } + ], + "index": 86 + }, + { + "bbox": [ + 304, + 735, + 526, + 747 + ], + "spans": [ + { + "bbox": [ + 304, + 735, + 526, + 747 + ], + "score": 1.0, + "content": "on humans to better understand their potential re-", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 304, + 748, + 525, + 761 + ], + "spans": [ + { + "bbox": [ + 304, + 748, + 525, + 761 + ], + "score": 1.0, + "content": "sponse motivation and communication patterns", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 305, + 761, + 526, + 775 + ], + "spans": [ + { + "bbox": [ + 305, + 761, + 526, + 775 + ], + "score": 1.0, + "content": "(Jiang et al., 2020); (2) helps reveal whether LLMs", + "type": "text" + } + ], + "index": 89 + } + ], + "index": 83 + } + ], + "page_idx": 0, + "page_size": [ + 595, + 841 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 70, + 742, + 290, + 772 + ], + "lines": [ + { + "bbox": [ + 82, + 740, + 167, + 754 + ], + "spans": [ + { + "bbox": [ + 82, + 740, + 167, + 754 + ], + "score": 1.0, + "content": "*Corresponding author", + "type": "text" + } + ] + }, + { + "bbox": [ + 81, + 751, + 292, + 763 + ], + "spans": [ + { + "bbox": [ + 81, + 751, + 292, + 763 + ], + "score": 1.0, + "content": "†Our codes are available at https://github.com/Kali-", + "type": "text" + } + ] + }, + { + "bbox": [ + 69, + 762, + 150, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 762, + 150, + 772 + ], + "score": 1.0, + "content": "Hac/ChatGPT-MBTI.", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 162, + 70, + 433, + 102 + ], + "lines": [ + { + "bbox": [ + 162, + 70, + 434, + 86 + ], + "spans": [ + { + "bbox": [ + 162, + 70, + 434, + 86 + ], + "score": 1.0, + "content": "Can ChatGPT Assess Human Personalities?", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 190, + 86, + 404, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 86, + 404, + 102 + ], + "score": 1.0, + "content": "A General Evaluation Framework", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 157, + 110, + 429, + 124 + ], + "lines": [ + { + "bbox": [ + 157, + 108, + 432, + 128 + ], + "spans": [ + { + "bbox": [ + 157, + 110, + 244, + 127 + ], + "score": 1.0, + "content": "Haocong Rao1,2", + "type": "text" + }, + { + "bbox": [ + 248, + 108, + 329, + 128 + ], + "score": 1.0, + "content": "Cyril Leung2,3", + "type": "text" + }, + { + "bbox": [ + 335, + 110, + 432, + 126 + ], + "score": 1.0, + "content": "Chunyan Miao1,2∗", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 123, + 525, + 142 + ], + "spans": [ + { + "bbox": [ + 73, + 126, + 78, + 137 + ], + "score": 0.35, + "content": "^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 78, + 123, + 525, + 142 + ], + "score": 1.0, + "content": "School of Computer Science and Engineering, Nanyang Technological University, Singapore", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 136, + 470, + 156 + ], + "spans": [ + { + "bbox": [ + 126, + 136, + 470, + 156 + ], + "score": 1.0, + "content": "2LILY Research Centre, Nanyang Technological University, Singapore", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 169, + 151, + 429, + 169 + ], + "spans": [ + { + "bbox": [ + 169, + 151, + 429, + 169 + ], + "score": 1.0, + "content": "3Department of Electrical and Computer Engineering", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 191, + 167, + 407, + 181 + ], + "spans": [ + { + "bbox": [ + 191, + 167, + 407, + 181 + ], + "score": 1.0, + "content": "The University of British Columbia, Canada", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 181, + 462, + 196 + ], + "spans": [ + { + "bbox": [ + 136, + 181, + 462, + 196 + ], + "score": 1.0, + "content": "{haocong001,ascymiao}@ntu.edu.sg {cleung}@ece.ubc.ca", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 2, + "bbox_fs": [ + 157, + 108, + 432, + 128 + ] + }, + { + "type": "text", + "bbox": [ + 75, + 119, + 525, + 195 + ], + "lines": [], + "index": 5, + "bbox_fs": [ + 73, + 123, + 525, + 196 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 157, + 213, + 202, + 226 + ], + "lines": [ + { + "bbox": [ + 155, + 212, + 204, + 228 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 204, + 228 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 87, + 238, + 273, + 620 + ], + "lines": [ + { + "bbox": [ + 86, + 238, + 273, + 250 + ], + "spans": [ + { + "bbox": [ + 86, + 238, + 273, + 250 + ], + "score": 1.0, + "content": "Large Language Models (LLMs) especially", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 250, + 273, + 262 + ], + "spans": [ + { + "bbox": [ + 86, + 250, + 273, + 262 + ], + "score": 1.0, + "content": "ChatGPT have produced impressive results in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 263, + 273, + 274 + ], + "spans": [ + { + "bbox": [ + 86, + 263, + 273, + 274 + ], + "score": 1.0, + "content": "various areas, but their potential human-like", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 86, + 275, + 274, + 286 + ], + "spans": [ + { + "bbox": [ + 86, + 275, + 274, + 286 + ], + "score": 1.0, + "content": "psychology is still largely unexplored. Ex-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 85, + 286, + 274, + 298 + ], + "spans": [ + { + "bbox": [ + 85, + 286, + 274, + 298 + ], + "score": 1.0, + "content": "isting works study the virtual personalities of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 85, + 297, + 274, + 311 + ], + "spans": [ + { + "bbox": [ + 85, + 297, + 274, + 311 + ], + "score": 1.0, + "content": "LLMs but rarely explore the possibility of an-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 86, + 310, + 274, + 321 + ], + "spans": [ + { + "bbox": [ + 86, + 310, + 274, + 321 + ], + "score": 1.0, + "content": "alyzing human personalities via LLMs. This", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 85, + 322, + 274, + 334 + ], + "spans": [ + { + "bbox": [ + 85, + 322, + 274, + 334 + ], + "score": 1.0, + "content": "paper presents a generic evaluation framework", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 86, + 334, + 274, + 346 + ], + "spans": [ + { + "bbox": [ + 86, + 334, + 274, + 346 + ], + "score": 1.0, + "content": "for LLMs to assess human personalities based", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 86, + 347, + 274, + 357 + ], + "spans": [ + { + "bbox": [ + 86, + 347, + 274, + 357 + ], + "score": 1.0, + "content": "on Myers–Briggs Type Indicator (MBTI) tests.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 86, + 358, + 273, + 370 + ], + "spans": [ + { + "bbox": [ + 86, + 358, + 273, + 370 + ], + "score": 1.0, + "content": "Specifically, we first devise unbiased prompts", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 85, + 369, + 275, + 383 + ], + "spans": [ + { + "bbox": [ + 85, + 369, + 275, + 383 + ], + "score": 1.0, + "content": "by randomly permuting options in MBTI ques-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 86, + 382, + 274, + 394 + ], + "spans": [ + { + "bbox": [ + 86, + 382, + 274, + 394 + ], + "score": 1.0, + "content": "tions and adopt the average testing result to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 86, + 394, + 275, + 406 + ], + "spans": [ + { + "bbox": [ + 86, + 394, + 275, + 406 + ], + "score": 1.0, + "content": "encourage more impartial answer generation.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 85, + 405, + 275, + 419 + ], + "spans": [ + { + "bbox": [ + 85, + 405, + 275, + 419 + ], + "score": 1.0, + "content": "Then, we propose to replace the subject in ques-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 86, + 417, + 274, + 430 + ], + "spans": [ + { + "bbox": [ + 86, + 417, + 274, + 430 + ], + "score": 1.0, + "content": "tion statements to enable flexible queries and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 85, + 430, + 275, + 442 + ], + "spans": [ + { + "bbox": [ + 85, + 430, + 275, + 442 + ], + "score": 1.0, + "content": "assessments on different subjects from LLMs.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 85, + 441, + 275, + 454 + ], + "spans": [ + { + "bbox": [ + 85, + 441, + 275, + 454 + ], + "score": 1.0, + "content": "Finally, we re-formulate the question instruc-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 86, + 453, + 273, + 465 + ], + "spans": [ + { + "bbox": [ + 86, + 453, + 273, + 465 + ], + "score": 1.0, + "content": "tions in a manner of correctness evaluation to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 85, + 464, + 275, + 478 + ], + "spans": [ + { + "bbox": [ + 85, + 464, + 275, + 478 + ], + "score": 1.0, + "content": "facilitate LLMs to generate clearer responses.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 85, + 477, + 275, + 489 + ], + "spans": [ + { + "bbox": [ + 85, + 477, + 275, + 489 + ], + "score": 1.0, + "content": "The proposed framework enables LLMs to flex-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 85, + 488, + 274, + 502 + ], + "spans": [ + { + "bbox": [ + 85, + 488, + 274, + 502 + ], + "score": 1.0, + "content": "ibly assess personalities of different groups of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 85, + 501, + 274, + 513 + ], + "spans": [ + { + "bbox": [ + 85, + 501, + 274, + 513 + ], + "score": 1.0, + "content": "people. We further propose three evaluation", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 86, + 514, + 274, + 525 + ], + "spans": [ + { + "bbox": [ + 86, + 514, + 274, + 525 + ], + "score": 1.0, + "content": "metrics to measure the consistency, robustness,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 86, + 525, + 274, + 536 + ], + "spans": [ + { + "bbox": [ + 86, + 525, + 274, + 536 + ], + "score": 1.0, + "content": "and fairness of assessment results from state-of-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 86, + 537, + 275, + 549 + ], + "spans": [ + { + "bbox": [ + 86, + 537, + 275, + 549 + ], + "score": 1.0, + "content": "the-art LLMs including ChatGPT and GPT-4.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 86, + 549, + 274, + 562 + ], + "spans": [ + { + "bbox": [ + 86, + 549, + 274, + 562 + ], + "score": 1.0, + "content": "Our experiments reveal ChatGPT’s ability to", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 86, + 561, + 274, + 574 + ], + "spans": [ + { + "bbox": [ + 86, + 561, + 274, + 574 + ], + "score": 1.0, + "content": "assess human personalities, and the average", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 85, + 573, + 273, + 585 + ], + "spans": [ + { + "bbox": [ + 85, + 573, + 273, + 585 + ], + "score": 1.0, + "content": "results demonstrate that it can achieve more", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 86, + 586, + 273, + 596 + ], + "spans": [ + { + "bbox": [ + 86, + 586, + 273, + 596 + ], + "score": 1.0, + "content": "consistent and fairer assessments in spite of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 86, + 597, + 275, + 609 + ], + "spans": [ + { + "bbox": [ + 86, + 597, + 275, + 609 + ], + "score": 1.0, + "content": "lower robustness against prompt biases com-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 86, + 608, + 189, + 621 + ], + "spans": [ + { + "bbox": [ + 86, + 608, + 189, + 621 + ], + "score": 1.0, + "content": "pared with InstructGPT†.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 24.5, + "bbox_fs": [ + 85, + 238, + 275, + 621 + ] + }, + { + "type": "title", + "bbox": [ + 70, + 643, + 153, + 657 + ], + "lines": [ + { + "bbox": [ + 67, + 642, + 155, + 659 + ], + "spans": [ + { + "bbox": [ + 67, + 642, + 155, + 659 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 69, + 666, + 290, + 733 + ], + "lines": [ + { + "bbox": [ + 69, + 666, + 290, + 680 + ], + "spans": [ + { + "bbox": [ + 69, + 666, + 290, + 680 + ], + "score": 1.0, + "content": "Pre-trained Large Language Models (LLMs) have", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 69, + 680, + 290, + 694 + ], + "spans": [ + { + "bbox": [ + 69, + 680, + 290, + 694 + ], + "score": 1.0, + "content": "been widely used in many applications including", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 69, + 694, + 291, + 706 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 291, + 706 + ], + "score": 1.0, + "content": "translation, storytelling, and chatbots (Devlin et al.,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 69, + 706, + 290, + 720 + ], + "spans": [ + { + "bbox": [ + 69, + 706, + 290, + 720 + ], + "score": 1.0, + "content": "2019; Raffel et al., 2020; Yang et al., 2022; Yuan", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 69, + 721, + 291, + 734 + ], + "spans": [ + { + "bbox": [ + 69, + 721, + 291, + 734 + ], + "score": 1.0, + "content": "et al., 2022; Ouyang et al., 2022; Bubeck et al.,", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44, + "bbox_fs": [ + 69, + 666, + 291, + 734 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 214, + 525, + 347 + ], + "lines": [ + { + "bbox": [ + 303, + 213, + 526, + 226 + ], + "spans": [ + { + "bbox": [ + 303, + 213, + 526, + 226 + ], + "score": 1.0, + "content": "2023). ChatGPT (Ouyang et al., 2022) and its", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 226, + 526, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 526, + 240 + ], + "score": 1.0, + "content": "enhanced version GPT-4 are currently recognized", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 241, + 525, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 525, + 253 + ], + "score": 1.0, + "content": "as the most capable chatbots, which can perform", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 255, + 525, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 525, + 267 + ], + "score": 1.0, + "content": "context-aware conversations, challenge incorrect", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 269, + 525, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 525, + 280 + ], + "score": 1.0, + "content": "premises, and reject inappropriate requests with", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 303, + 282, + 527, + 293 + ], + "spans": [ + { + "bbox": [ + 303, + 282, + 527, + 293 + ], + "score": 1.0, + "content": "a vast knowledge base and human-centered fine-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 304, + 295, + 526, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 526, + 307 + ], + "score": 1.0, + "content": "tuning. These advantages make them well-suited", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 303, + 308, + 527, + 321 + ], + "spans": [ + { + "bbox": [ + 303, + 308, + 527, + 321 + ], + "score": 1.0, + "content": "for a variety of real-world scenarios such as busi-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 303, + 322, + 527, + 334 + ], + "spans": [ + { + "bbox": [ + 303, + 322, + 527, + 334 + ], + "score": 1.0, + "content": "ness consultation and educational services (Zhai,", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 335, + 515, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 515, + 348 + ], + "score": 1.0, + "content": "2022; van Dis et al., 2023; Bubeck et al., 2023).", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 53.5, + "bbox_fs": [ + 303, + 213, + 527, + 348 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 352, + 525, + 594 + ], + "lines": [ + { + "bbox": [ + 315, + 351, + 527, + 366 + ], + "spans": [ + { + "bbox": [ + 315, + 351, + 527, + 366 + ], + "score": 1.0, + "content": "Recent studies have revealed that LLMs may pos-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 365, + 526, + 379 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 526, + 379 + ], + "score": 1.0, + "content": "sess human-like self-improvement and reasoning", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 379, + 526, + 391 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 526, + 391 + ], + "score": 1.0, + "content": "characteristics (Huang et al., 2022; Bubeck et al.,", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 303, + 391, + 527, + 406 + ], + "spans": [ + { + "bbox": [ + 303, + 391, + 491, + 406 + ], + "score": 1.0, + "content": "2023). The latest GPT series can pass over", + "type": "text" + }, + { + "bbox": [ + 491, + 392, + 513, + 404 + ], + "score": 0.9, + "content": "90 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 513, + 391, + 527, + 406 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 406, + 525, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 525, + 419 + ], + "score": 1.0, + "content": "Theory of Mind (ToM) tasks with strong analysis", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 420, + 526, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 526, + 432 + ], + "score": 1.0, + "content": "and decision-making capabilities (Kosinski, 2023;", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 433, + 527, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 527, + 446 + ], + "score": 1.0, + "content": "Zhuo et al., 2023; Moghaddam and Honey, 2023).", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 446, + 525, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 525, + 460 + ], + "score": 1.0, + "content": "In this context, LLMs are increasingly assumed to", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 303, + 460, + 526, + 474 + ], + "spans": [ + { + "bbox": [ + 303, + 460, + 526, + 474 + ], + "score": 1.0, + "content": "have virtual personalities and psychologies, which", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 474, + 525, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 525, + 487 + ], + "score": 1.0, + "content": "plays an essential role in guiding their responses", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 488, + 525, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 525, + 500 + ], + "score": 1.0, + "content": "and interaction patterns (Jiang et al., 2022). Based", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 501, + 526, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 526, + 513 + ], + "score": 1.0, + "content": "on this assumption, a few works (Li et al., 2022;", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 303, + 515, + 525, + 527 + ], + "spans": [ + { + "bbox": [ + 303, + 515, + 525, + 527 + ], + "score": 1.0, + "content": "Jiang et al., 2022; Karra et al., 2022; Caron and", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 527, + 527, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 527, + 542 + ], + "score": 1.0, + "content": "Srivastava, 2022; Miotto et al., 2022) apply psy-", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 304, + 541, + 527, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 527, + 555 + ], + "score": 1.0, + "content": "chological tests such as Big Five Factors (Digman,", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 554, + 525, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 525, + 569 + ], + "score": 1.0, + "content": "1990) to evaluate their pseudo personalities (e.g.,", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 568, + 526, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 526, + 581 + ], + "score": 1.0, + "content": "behavior tendency), so as to detect societal and eth-", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 304, + 582, + 524, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 524, + 595 + ], + "score": 1.0, + "content": "ical risks (e.g., racial biases) in their applications.", + "type": "text" + } + ], + "index": 76 + } + ], + "index": 67.5, + "bbox_fs": [ + 303, + 351, + 527, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 599, + 525, + 773 + ], + "lines": [ + { + "bbox": [ + 316, + 599, + 525, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 599, + 525, + 611 + ], + "score": 1.0, + "content": "Although existing works have investigated the", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 303, + 611, + 526, + 626 + ], + "spans": [ + { + "bbox": [ + 303, + 611, + 526, + 626 + ], + "score": 1.0, + "content": "personality traits of LLMs, they rarely explored", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 626, + 527, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 527, + 639 + ], + "score": 1.0, + "content": "whether LLMs can assess human personalities.", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 639, + 526, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 526, + 653 + ], + "score": 1.0, + "content": "This open problem can be the key to verifying the", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 652, + 527, + 667 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 527, + 667 + ], + "score": 1.0, + "content": "ability of LLMs to perform psychological (e.g., per-", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 303, + 666, + 526, + 680 + ], + "spans": [ + { + "bbox": [ + 303, + 666, + 526, + 680 + ], + "score": 1.0, + "content": "sonality psychology) analyses and revealing their", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 303, + 680, + 526, + 693 + ], + "spans": [ + { + "bbox": [ + 303, + 680, + 526, + 693 + ], + "score": 1.0, + "content": "potential understanding of humans, i.e., “How do", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 303, + 693, + 527, + 707 + ], + "spans": [ + { + "bbox": [ + 303, + 693, + 527, + 707 + ], + "score": 1.0, + "content": "LLMs think about humans?”. Specifically, assess-", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "score": 1.0, + "content": "ing human personalities from the point of LLMs", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 303, + 720, + 526, + 734 + ], + "spans": [ + { + "bbox": [ + 303, + 720, + 526, + 734 + ], + "score": 1.0, + "content": "(1) enables us to access the perception of LLMs", + "type": "text" + } + ], + "index": 86 + }, + { + "bbox": [ + 304, + 735, + 526, + 747 + ], + "spans": [ + { + "bbox": [ + 304, + 735, + 526, + 747 + ], + "score": 1.0, + "content": "on humans to better understand their potential re-", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 304, + 748, + 525, + 761 + ], + "spans": [ + { + "bbox": [ + 304, + 748, + 525, + 761 + ], + "score": 1.0, + "content": "sponse motivation and communication patterns", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 305, + 761, + 526, + 775 + ], + "spans": [ + { + "bbox": [ + 305, + 761, + 526, + 775 + ], + "score": 1.0, + "content": "(Jiang et al., 2020); (2) helps reveal whether LLMs", + "type": "text" + } + ], + "index": 89 + }, + { + "bbox": [ + 68, + 72, + 290, + 86 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 290, + 86 + ], + "score": 1.0, + "content": "possess biases on people so that we can optimize", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 85, + 291, + 98 + ], + "spans": [ + { + "bbox": [ + 69, + 85, + 291, + 98 + ], + "score": 1.0, + "content": "them (e.g., add stricter rules) to generate fairer con-", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 99, + 290, + 111 + ], + "spans": [ + { + "bbox": [ + 68, + 99, + 290, + 111 + ], + "score": 1.0, + "content": "tents; (3) helps uncover potential ethical and social", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 112, + 291, + 126 + ], + "spans": [ + { + "bbox": [ + 68, + 112, + 291, + 126 + ], + "score": 1.0, + "content": "risks (e.g., misinformation) of LLMs (Weidinger", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 126, + 290, + 139 + ], + "spans": [ + { + "bbox": [ + 68, + 126, + 290, + 139 + ], + "score": 1.0, + "content": "et al., 2021) which can affect their reliability and", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "spans": [ + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "score": 1.0, + "content": "safety, thereby facilitating the development of more", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 153, + 244, + 166 + ], + "spans": [ + { + "bbox": [ + 68, + 153, + 244, + 166 + ], + "score": 1.0, + "content": "trustworthy and human-friendly LLMs.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 83, + "bbox_fs": [ + 303, + 599, + 527, + 775 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 70, + 72, + 290, + 165 + ], + "lines": [ + { + "bbox": [ + 68, + 72, + 290, + 86 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 290, + 86 + ], + "score": 1.0, + "content": "possess biases on people so that we can optimize", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 85, + 291, + 98 + ], + "spans": [ + { + "bbox": [ + 69, + 85, + 291, + 98 + ], + "score": 1.0, + "content": "them (e.g., add stricter rules) to generate fairer con-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 99, + 290, + 111 + ], + "spans": [ + { + "bbox": [ + 68, + 99, + 290, + 111 + ], + "score": 1.0, + "content": "tents; (3) helps uncover potential ethical and social", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 112, + 291, + 126 + ], + "spans": [ + { + "bbox": [ + 68, + 112, + 291, + 126 + ], + "score": 1.0, + "content": "risks (e.g., misinformation) of LLMs (Weidinger", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 126, + 290, + 139 + ], + "spans": [ + { + "bbox": [ + 68, + 126, + 290, + 139 + ], + "score": 1.0, + "content": "et al., 2021) which can affect their reliability and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "spans": [ + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "score": 1.0, + "content": "safety, thereby facilitating the development of more", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 153, + 244, + 166 + ], + "spans": [ + { + "bbox": [ + 68, + 153, + 244, + 166 + ], + "score": 1.0, + "content": "trustworthy and human-friendly LLMs.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 69, + 166, + 290, + 669 + ], + "lines": [ + { + "bbox": [ + 79, + 167, + 291, + 179 + ], + "spans": [ + { + "bbox": [ + 79, + 167, + 291, + 179 + ], + "score": 1.0, + "content": "To this end, we introduce the novel idea of let-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 179, + 291, + 194 + ], + "spans": [ + { + "bbox": [ + 68, + 179, + 291, + 194 + ], + "score": 1.0, + "content": "ting LLMs assess human personalities, and propose", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 193, + 292, + 208 + ], + "spans": [ + { + "bbox": [ + 67, + 193, + 292, + 208 + ], + "score": 1.0, + "content": "a general evaluation framework (illustrated Fig. 1)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 207, + 292, + 221 + ], + "spans": [ + { + "bbox": [ + 67, + 207, + 292, + 221 + ], + "score": 1.0, + "content": "to acquire quantitative human personality assess-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 220, + 292, + 234 + ], + "spans": [ + { + "bbox": [ + 68, + 220, + 292, + 234 + ], + "score": 1.0, + "content": "ments from LLMs via Myers–Briggs Type Indica-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 68, + 234, + 292, + 248 + ], + "spans": [ + { + "bbox": [ + 68, + 234, + 292, + 248 + ], + "score": 1.0, + "content": "tors (MBTI) (Myers and McCaulley, 1985). Specif-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 68, + 248, + 292, + 262 + ], + "spans": [ + { + "bbox": [ + 68, + 248, + 292, + 262 + ], + "score": 1.0, + "content": "ically, our framework consists of three key com-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 261, + 291, + 275 + ], + "spans": [ + { + "bbox": [ + 67, + 261, + 291, + 275 + ], + "score": 1.0, + "content": "ponents: (1) Unbiased prompts, which construct", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 68, + 275, + 291, + 289 + ], + "spans": [ + { + "bbox": [ + 68, + 275, + 291, + 289 + ], + "score": 1.0, + "content": "instructions of MBTI questions using randomly-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 289, + 291, + 302 + ], + "spans": [ + { + "bbox": [ + 67, + 289, + 291, + 302 + ], + "score": 1.0, + "content": "permuted options and average testing results to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 68, + 302, + 291, + 316 + ], + "spans": [ + { + "bbox": [ + 68, + 302, + 291, + 316 + ], + "score": 1.0, + "content": "achieve more consistent and impartial answers; (2)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 68, + 316, + 292, + 330 + ], + "spans": [ + { + "bbox": [ + 68, + 316, + 292, + 330 + ], + "score": 1.0, + "content": "Subject-replaced query, which converts the origi-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 68, + 329, + 291, + 344 + ], + "spans": [ + { + "bbox": [ + 68, + 329, + 291, + 344 + ], + "score": 1.0, + "content": "nal subject of the question statements into a target", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 68, + 342, + 291, + 357 + ], + "spans": [ + { + "bbox": [ + 68, + 342, + 291, + 357 + ], + "score": 1.0, + "content": "subject to enable flexible queries and assessments", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 68, + 356, + 292, + 370 + ], + "spans": [ + { + "bbox": [ + 68, + 356, + 292, + 370 + ], + "score": 1.0, + "content": "from LLMs; (3) Correctness-evaluated instruction,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 68, + 369, + 291, + 384 + ], + "spans": [ + { + "bbox": [ + 68, + 369, + 291, + 384 + ], + "score": 1.0, + "content": "which re-formulates the question instructions for", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 69, + 384, + 290, + 397 + ], + "spans": [ + { + "bbox": [ + 69, + 384, + 290, + 397 + ], + "score": 1.0, + "content": "LLMs to analyze the correctness of the question", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 397, + 291, + 411 + ], + "spans": [ + { + "bbox": [ + 68, + 397, + 291, + 411 + ], + "score": 1.0, + "content": "statements, so as to obtain clearer responses. Based", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 68, + 410, + 291, + 424 + ], + "spans": [ + { + "bbox": [ + 68, + 410, + 291, + 424 + ], + "score": 1.0, + "content": "on the above components, the proposed framework", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 68, + 424, + 291, + 436 + ], + "spans": [ + { + "bbox": [ + 68, + 424, + 291, + 436 + ], + "score": 1.0, + "content": "re-formulates the instructions and statements of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 68, + 436, + 290, + 452 + ], + "spans": [ + { + "bbox": [ + 68, + 436, + 290, + 452 + ], + "score": 1.0, + "content": "MBTI questions in a flexible and analyzable way", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 68, + 450, + 291, + 465 + ], + "spans": [ + { + "bbox": [ + 68, + 450, + 291, + 465 + ], + "score": 1.0, + "content": "for LLMs, which enables us to query them about", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 68, + 464, + 291, + 479 + ], + "spans": [ + { + "bbox": [ + 68, + 464, + 291, + 479 + ], + "score": 1.0, + "content": "human personalities. Furthermore, we propose", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 68, + 477, + 292, + 493 + ], + "spans": [ + { + "bbox": [ + 68, + 477, + 292, + 493 + ], + "score": 1.0, + "content": "three quantitative evaluation metrics to measure", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 68, + 491, + 291, + 506 + ], + "spans": [ + { + "bbox": [ + 68, + 491, + 291, + 506 + ], + "score": 1.0, + "content": "the consistency of LLMs’ assessments on the same", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 504, + 292, + 520 + ], + "spans": [ + { + "bbox": [ + 68, + 504, + 292, + 520 + ], + "score": 1.0, + "content": "subject, their assessment robustness against ran-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 68, + 518, + 290, + 532 + ], + "spans": [ + { + "bbox": [ + 68, + 518, + 290, + 532 + ], + "score": 1.0, + "content": "dom perturbations of input prompts (defined as", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 68, + 531, + 291, + 548 + ], + "spans": [ + { + "bbox": [ + 68, + 531, + 291, + 548 + ], + "score": 1.0, + "content": "“prompt biases”), and their fairness in assessing", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 69, + 546, + 291, + 560 + ], + "spans": [ + { + "bbox": [ + 69, + 546, + 291, + 560 + ], + "score": 1.0, + "content": "subjects with different genders. In our work, we", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 68, + 558, + 292, + 574 + ], + "spans": [ + { + "bbox": [ + 68, + 558, + 292, + 574 + ], + "score": 1.0, + "content": "mainly focus on evaluating ChatGPT and two repre-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 67, + 573, + 292, + 586 + ], + "spans": [ + { + "bbox": [ + 67, + 573, + 292, + 586 + ], + "score": 1.0, + "content": "sentative state-of-the-art LLMs (InstructGPT, GPT-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 69, + 587, + 290, + 600 + ], + "spans": [ + { + "bbox": [ + 69, + 587, + 290, + 600 + ], + "score": 1.0, + "content": "4) based on the proposed metrics. Experimental", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 68, + 600, + 292, + 614 + ], + "spans": [ + { + "bbox": [ + 68, + 600, + 292, + 614 + ], + "score": 1.0, + "content": "results showcase the ability of ChatGPT in ana-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 69, + 614, + 292, + 628 + ], + "spans": [ + { + "bbox": [ + 69, + 614, + 292, + 628 + ], + "score": 1.0, + "content": "lyzing personalities of different groups of people.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 69, + 627, + 291, + 641 + ], + "spans": [ + { + "bbox": [ + 69, + 627, + 291, + 641 + ], + "score": 1.0, + "content": "This can provide valuable insights for the future", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 68, + 640, + 292, + 656 + ], + "spans": [ + { + "bbox": [ + 68, + 640, + 292, + 656 + ], + "score": 1.0, + "content": "exploration of LLM psychology, sociology, and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 67, + 655, + 126, + 669 + ], + "spans": [ + { + "bbox": [ + 67, + 655, + 126, + 669 + ], + "score": 1.0, + "content": "governance.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 81, + 668, + 289, + 680 + ], + "lines": [ + { + "bbox": [ + 81, + 668, + 291, + 681 + ], + "spans": [ + { + "bbox": [ + 81, + 668, + 291, + 681 + ], + "score": 1.0, + "content": "Our contributions can be summarized as follows:", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 82, + 686, + 290, + 739 + ], + "lines": [ + { + "bbox": [ + 81, + 684, + 291, + 700 + ], + "spans": [ + { + "bbox": [ + 81, + 684, + 291, + 700 + ], + "score": 1.0, + "content": "• We for the first time explore the possibility", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 91, + 700, + 291, + 713 + ], + "spans": [ + { + "bbox": [ + 91, + 700, + 291, + 713 + ], + "score": 1.0, + "content": "of assessing human personalities by LLMs,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 90, + 713, + 291, + 726 + ], + "spans": [ + { + "bbox": [ + 90, + 713, + 291, + 726 + ], + "score": 1.0, + "content": "and propose a general framework for LLMs", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 90, + 727, + 292, + 739 + ], + "spans": [ + { + "bbox": [ + 90, + 727, + 292, + 739 + ], + "score": 1.0, + "content": "to conduct quantitative evaluations via MBTI.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + }, + { + "type": "text", + "bbox": [ + 82, + 747, + 290, + 773 + ], + "lines": [ + { + "bbox": [ + 81, + 746, + 290, + 761 + ], + "spans": [ + { + "bbox": [ + 81, + 746, + 290, + 761 + ], + "score": 1.0, + "content": "• We devise unbiased prompts, subject-replaced", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 90, + 761, + 292, + 774 + ], + "spans": [ + { + "bbox": [ + 90, + 761, + 292, + 774 + ], + "score": 1.0, + "content": "queries, and correctness-evaluated instruc-", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5 + }, + { + "type": "text", + "bbox": [ + 326, + 72, + 524, + 97 + ], + "lines": [ + { + "bbox": [ + 326, + 72, + 525, + 85 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 525, + 85 + ], + "score": 1.0, + "content": "tions to encourage LLMs to perform a reliable", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 326, + 85, + 517, + 99 + ], + "spans": [ + { + "bbox": [ + 326, + 85, + 517, + 99 + ], + "score": 1.0, + "content": "flexible assessment of human personalities.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51.5 + }, + { + "type": "text", + "bbox": [ + 318, + 106, + 525, + 146 + ], + "lines": [ + { + "bbox": [ + 317, + 106, + 527, + 120 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 527, + 120 + ], + "score": 1.0, + "content": "• We propose three evaluation metrics to mea-", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 326, + 121, + 525, + 133 + ], + "spans": [ + { + "bbox": [ + 326, + 121, + 525, + 133 + ], + "score": 1.0, + "content": "sure the consistency, robustness, and fairness", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 326, + 134, + 515, + 147 + ], + "spans": [ + { + "bbox": [ + 326, + 134, + 515, + 147 + ], + "score": 1.0, + "content": "of LLMs in assessing human personalities.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 54 + }, + { + "type": "text", + "bbox": [ + 318, + 155, + 526, + 249 + ], + "lines": [ + { + "bbox": [ + 318, + 155, + 526, + 168 + ], + "spans": [ + { + "bbox": [ + 318, + 155, + 526, + 168 + ], + "score": 1.0, + "content": "• Our experiments show that both ChatGPT and", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 326, + 169, + 527, + 182 + ], + "spans": [ + { + "bbox": [ + 326, + 169, + 527, + 182 + ], + "score": 1.0, + "content": "its counterparts can independently assess hu-", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 326, + 183, + 527, + 195 + ], + "spans": [ + { + "bbox": [ + 326, + 183, + 527, + 195 + ], + "score": 1.0, + "content": "man personalities. The average results demon-", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 326, + 196, + 525, + 208 + ], + "spans": [ + { + "bbox": [ + 326, + 196, + 525, + 208 + ], + "score": 1.0, + "content": "strate that ChatGPT and GPT-4 achieve more", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 326, + 210, + 525, + 221 + ], + "spans": [ + { + "bbox": [ + 326, + 210, + 525, + 221 + ], + "score": 1.0, + "content": "consistent and fairer assessments with less", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 326, + 223, + 527, + 236 + ], + "spans": [ + { + "bbox": [ + 326, + 223, + 527, + 236 + ], + "score": 1.0, + "content": "gender bias than InstructGPT, while their re-", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 326, + 237, + 508, + 250 + ], + "spans": [ + { + "bbox": [ + 326, + 237, + 508, + 250 + ], + "score": 1.0, + "content": "sults are more sensitive to prompt biases.", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 59 + }, + { + "type": "title", + "bbox": [ + 305, + 258, + 400, + 272 + ], + "lines": [ + { + "bbox": [ + 302, + 257, + 401, + 274 + ], + "spans": [ + { + "bbox": [ + 302, + 257, + 401, + 274 + ], + "score": 1.0, + "content": "2 Related Works", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 63 + }, + { + "type": "text", + "bbox": [ + 304, + 281, + 525, + 591 + ], + "lines": [ + { + "bbox": [ + 304, + 280, + 526, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 526, + 293 + ], + "score": 1.0, + "content": "Personality Measurement. The commonly-used", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 293, + 526, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 526, + 306 + ], + "score": 1.0, + "content": "personality modeling schemes include the three", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 305, + 308, + 525, + 320 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 525, + 320 + ], + "score": 1.0, + "content": "trait personality measure (Eysenck, 2012), the Big", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 303, + 321, + 527, + 334 + ], + "spans": [ + { + "bbox": [ + 303, + 321, + 527, + 334 + ], + "score": 1.0, + "content": "Five personality trait measure (Digman, 1990),", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 333, + 527, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 527, + 348 + ], + "score": 1.0, + "content": "the Myers–Briggs Type Indicator (MBTI) (Myers,", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 347, + 526, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 526, + 360 + ], + "score": 1.0, + "content": "1962; Myers and McCaulley, 1985), and the 16 Per-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 361, + 527, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 527, + 375 + ], + "score": 1.0, + "content": "sonality Factor questionnaire (16PF) (Schuerger,", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 374, + 526, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 526, + 388 + ], + "score": 1.0, + "content": "2000). Five dimensions are defined in the Big Five", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 389, + 527, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 527, + 401 + ], + "score": 1.0, + "content": "personality traits measure (Digman, 1990) to clas-", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 305, + 402, + 526, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 526, + 415 + ], + "score": 1.0, + "content": "sify major sources of individual differences and", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 416, + 525, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 525, + 428 + ], + "score": 1.0, + "content": "analyze a person’s characteristics. MBTI (Myers", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 429, + 525, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 525, + 442 + ], + "score": 1.0, + "content": "and McCaulley, 1985) identifies personality from", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 304, + 442, + 525, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 525, + 455 + ], + "score": 1.0, + "content": "the differences between persons on the preference", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 457, + 526, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 526, + 469 + ], + "score": 1.0, + "content": "to use perception and judgment. (Karra et al., 2022;", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 304, + 470, + 525, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 525, + 483 + ], + "score": 1.0, + "content": "Caron and Srivastava, 2022) leverage the Big Five", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 484, + 526, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 526, + 496 + ], + "score": 1.0, + "content": "trait theory to quantify the personality traits of lan-", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 303, + 497, + 526, + 510 + ], + "spans": [ + { + "bbox": [ + 303, + 497, + 526, + 510 + ], + "score": 1.0, + "content": "guage models, while (Jiang et al., 2022) further", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 511, + 526, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 526, + 524 + ], + "score": 1.0, + "content": "develops machine personality inventory to stan-", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 304, + 523, + 526, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 526, + 537 + ], + "score": 1.0, + "content": "dardize this evaluation. In (Li et al., 2022), multi-", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 304, + 538, + 525, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 525, + 551 + ], + "score": 1.0, + "content": "ple psychological tests are combined to analyze the", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 303, + 551, + 525, + 564 + ], + "spans": [ + { + "bbox": [ + 303, + 551, + 525, + 564 + ], + "score": 1.0, + "content": "LLMs’ safety. Unlike existing studies that evaluate", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 304, + 565, + 525, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 525, + 578 + ], + "score": 1.0, + "content": "personalities of LLMs, our work is the first attempt", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 304, + 579, + 521, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 521, + 591 + ], + "score": 1.0, + "content": "to explore human personality analysis via LLMs.", + "type": "text" + } + ], + "index": 86 + } + ], + "index": 75 + }, + { + "type": "text", + "bbox": [ + 305, + 598, + 526, + 773 + ], + "lines": [ + { + "bbox": [ + 304, + 598, + 527, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 527, + 610 + ], + "score": 1.0, + "content": "Biases in Language Models. Most recent lan-", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 304, + 612, + 525, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 525, + 625 + ], + "score": 1.0, + "content": "guage models are pre-trained on the large-scale", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 304, + 626, + 525, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 525, + 638 + ], + "score": 1.0, + "content": "datasets or Internet texts that usually contains", + "type": "text" + } + ], + "index": 89 + }, + { + "bbox": [ + 304, + 640, + 526, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 526, + 652 + ], + "score": 1.0, + "content": "unsafe (e.g., toxic) contents, which may cause", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 304, + 652, + 527, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 527, + 666 + ], + "score": 1.0, + "content": "the model to generate biased answers that vio-", + "type": "text" + } + ], + "index": 91 + }, + { + "bbox": [ + 304, + 667, + 526, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 667, + 526, + 679 + ], + "score": 1.0, + "content": "late prevailing societal values (Bolukbasi et al.,", + "type": "text" + } + ], + "index": 92 + }, + { + "bbox": [ + 303, + 678, + 527, + 693 + ], + "spans": [ + { + "bbox": [ + 303, + 678, + 527, + 693 + ], + "score": 1.0, + "content": "2016; Sheng et al., 2019; Bordia and Bowman,", + "type": "text" + } + ], + "index": 93 + }, + { + "bbox": [ + 304, + 693, + 527, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 693, + 527, + 706 + ], + "score": 1.0, + "content": "2019; Nadeem et al., 2021; Zong and Krishna-", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 304, + 707, + 526, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 707, + 526, + 720 + ], + "score": 1.0, + "content": "machari, 2022; Zhuo et al., 2023). (Bolukbasi et al.,", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 304, + 720, + 527, + 734 + ], + "spans": [ + { + "bbox": [ + 304, + 720, + 527, + 734 + ], + "score": 1.0, + "content": "2016) shows that biases in the geometry of word-", + "type": "text" + } + ], + "index": 96 + }, + { + "bbox": [ + 304, + 735, + 525, + 747 + ], + "spans": [ + { + "bbox": [ + 304, + 735, + 525, + 747 + ], + "score": 1.0, + "content": "embeddings can reflect gender stereotypes. The", + "type": "text" + } + ], + "index": 97 + }, + { + "bbox": [ + 304, + 748, + 527, + 761 + ], + "spans": [ + { + "bbox": [ + 304, + 748, + 527, + 761 + ], + "score": 1.0, + "content": "gender bias in word-level language models is quan-", + "type": "text" + } + ], + "index": 98 + }, + { + "bbox": [ + 304, + 761, + 527, + 774 + ], + "spans": [ + { + "bbox": [ + 304, + 761, + 527, + 774 + ], + "score": 1.0, + "content": "titatively evaluated in (Bordia and Bowman, 2019).", + "type": "text" + } + ], + "index": 99 + } + ], + "index": 93 + } + ], + "page_idx": 1, + "page_size": [ + 595, + 841 + ], + "discarded_blocks": [], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 70, + 72, + 290, + 165 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 68, + 72, + 291, + 166 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 69, + 166, + 290, + 669 + ], + "lines": [ + { + "bbox": [ + 79, + 167, + 291, + 179 + ], + "spans": [ + { + "bbox": [ + 79, + 167, + 291, + 179 + ], + "score": 1.0, + "content": "To this end, we introduce the novel idea of let-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 179, + 291, + 194 + ], + "spans": [ + { + "bbox": [ + 68, + 179, + 291, + 194 + ], + "score": 1.0, + "content": "ting LLMs assess human personalities, and propose", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 193, + 292, + 208 + ], + "spans": [ + { + "bbox": [ + 67, + 193, + 292, + 208 + ], + "score": 1.0, + "content": "a general evaluation framework (illustrated Fig. 1)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 207, + 292, + 221 + ], + "spans": [ + { + "bbox": [ + 67, + 207, + 292, + 221 + ], + "score": 1.0, + "content": "to acquire quantitative human personality assess-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 220, + 292, + 234 + ], + "spans": [ + { + "bbox": [ + 68, + 220, + 292, + 234 + ], + "score": 1.0, + "content": "ments from LLMs via Myers–Briggs Type Indica-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 68, + 234, + 292, + 248 + ], + "spans": [ + { + "bbox": [ + 68, + 234, + 292, + 248 + ], + "score": 1.0, + "content": "tors (MBTI) (Myers and McCaulley, 1985). Specif-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 68, + 248, + 292, + 262 + ], + "spans": [ + { + "bbox": [ + 68, + 248, + 292, + 262 + ], + "score": 1.0, + "content": "ically, our framework consists of three key com-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 261, + 291, + 275 + ], + "spans": [ + { + "bbox": [ + 67, + 261, + 291, + 275 + ], + "score": 1.0, + "content": "ponents: (1) Unbiased prompts, which construct", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 68, + 275, + 291, + 289 + ], + "spans": [ + { + "bbox": [ + 68, + 275, + 291, + 289 + ], + "score": 1.0, + "content": "instructions of MBTI questions using randomly-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 289, + 291, + 302 + ], + "spans": [ + { + "bbox": [ + 67, + 289, + 291, + 302 + ], + "score": 1.0, + "content": "permuted options and average testing results to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 68, + 302, + 291, + 316 + ], + "spans": [ + { + "bbox": [ + 68, + 302, + 291, + 316 + ], + "score": 1.0, + "content": "achieve more consistent and impartial answers; (2)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 68, + 316, + 292, + 330 + ], + "spans": [ + { + "bbox": [ + 68, + 316, + 292, + 330 + ], + "score": 1.0, + "content": "Subject-replaced query, which converts the origi-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 68, + 329, + 291, + 344 + ], + "spans": [ + { + "bbox": [ + 68, + 329, + 291, + 344 + ], + "score": 1.0, + "content": "nal subject of the question statements into a target", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 68, + 342, + 291, + 357 + ], + "spans": [ + { + "bbox": [ + 68, + 342, + 291, + 357 + ], + "score": 1.0, + "content": "subject to enable flexible queries and assessments", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 68, + 356, + 292, + 370 + ], + "spans": [ + { + "bbox": [ + 68, + 356, + 292, + 370 + ], + "score": 1.0, + "content": "from LLMs; (3) Correctness-evaluated instruction,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 68, + 369, + 291, + 384 + ], + "spans": [ + { + "bbox": [ + 68, + 369, + 291, + 384 + ], + "score": 1.0, + "content": "which re-formulates the question instructions for", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 69, + 384, + 290, + 397 + ], + "spans": [ + { + "bbox": [ + 69, + 384, + 290, + 397 + ], + "score": 1.0, + "content": "LLMs to analyze the correctness of the question", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 397, + 291, + 411 + ], + "spans": [ + { + "bbox": [ + 68, + 397, + 291, + 411 + ], + "score": 1.0, + "content": "statements, so as to obtain clearer responses. Based", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 68, + 410, + 291, + 424 + ], + "spans": [ + { + "bbox": [ + 68, + 410, + 291, + 424 + ], + "score": 1.0, + "content": "on the above components, the proposed framework", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 68, + 424, + 291, + 436 + ], + "spans": [ + { + "bbox": [ + 68, + 424, + 291, + 436 + ], + "score": 1.0, + "content": "re-formulates the instructions and statements of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 68, + 436, + 290, + 452 + ], + "spans": [ + { + "bbox": [ + 68, + 436, + 290, + 452 + ], + "score": 1.0, + "content": "MBTI questions in a flexible and analyzable way", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 68, + 450, + 291, + 465 + ], + "spans": [ + { + "bbox": [ + 68, + 450, + 291, + 465 + ], + "score": 1.0, + "content": "for LLMs, which enables us to query them about", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 68, + 464, + 291, + 479 + ], + "spans": [ + { + "bbox": [ + 68, + 464, + 291, + 479 + ], + "score": 1.0, + "content": "human personalities. Furthermore, we propose", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 68, + 477, + 292, + 493 + ], + "spans": [ + { + "bbox": [ + 68, + 477, + 292, + 493 + ], + "score": 1.0, + "content": "three quantitative evaluation metrics to measure", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 68, + 491, + 291, + 506 + ], + "spans": [ + { + "bbox": [ + 68, + 491, + 291, + 506 + ], + "score": 1.0, + "content": "the consistency of LLMs’ assessments on the same", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 504, + 292, + 520 + ], + "spans": [ + { + "bbox": [ + 68, + 504, + 292, + 520 + ], + "score": 1.0, + "content": "subject, their assessment robustness against ran-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 68, + 518, + 290, + 532 + ], + "spans": [ + { + "bbox": [ + 68, + 518, + 290, + 532 + ], + "score": 1.0, + "content": "dom perturbations of input prompts (defined as", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 68, + 531, + 291, + 548 + ], + "spans": [ + { + "bbox": [ + 68, + 531, + 291, + 548 + ], + "score": 1.0, + "content": "“prompt biases”), and their fairness in assessing", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 69, + 546, + 291, + 560 + ], + "spans": [ + { + "bbox": [ + 69, + 546, + 291, + 560 + ], + "score": 1.0, + "content": "subjects with different genders. In our work, we", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 68, + 558, + 292, + 574 + ], + "spans": [ + { + "bbox": [ + 68, + 558, + 292, + 574 + ], + "score": 1.0, + "content": "mainly focus on evaluating ChatGPT and two repre-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 67, + 573, + 292, + 586 + ], + "spans": [ + { + "bbox": [ + 67, + 573, + 292, + 586 + ], + "score": 1.0, + "content": "sentative state-of-the-art LLMs (InstructGPT, GPT-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 69, + 587, + 290, + 600 + ], + "spans": [ + { + "bbox": [ + 69, + 587, + 290, + 600 + ], + "score": 1.0, + "content": "4) based on the proposed metrics. Experimental", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 68, + 600, + 292, + 614 + ], + "spans": [ + { + "bbox": [ + 68, + 600, + 292, + 614 + ], + "score": 1.0, + "content": "results showcase the ability of ChatGPT in ana-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 69, + 614, + 292, + 628 + ], + "spans": [ + { + "bbox": [ + 69, + 614, + 292, + 628 + ], + "score": 1.0, + "content": "lyzing personalities of different groups of people.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 69, + 627, + 291, + 641 + ], + "spans": [ + { + "bbox": [ + 69, + 627, + 291, + 641 + ], + "score": 1.0, + "content": "This can provide valuable insights for the future", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 68, + 640, + 292, + 656 + ], + "spans": [ + { + "bbox": [ + 68, + 640, + 292, + 656 + ], + "score": 1.0, + "content": "exploration of LLM psychology, sociology, and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 67, + 655, + 126, + 669 + ], + "spans": [ + { + "bbox": [ + 67, + 655, + 126, + 669 + ], + "score": 1.0, + "content": "governance.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 25, + "bbox_fs": [ + 67, + 167, + 292, + 669 + ] + }, + { + "type": "text", + "bbox": [ + 81, + 668, + 289, + 680 + ], + "lines": [ + { + "bbox": [ + 81, + 668, + 291, + 681 + ], + "spans": [ + { + "bbox": [ + 81, + 668, + 291, + 681 + ], + "score": 1.0, + "content": "Our contributions can be summarized as follows:", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44, + "bbox_fs": [ + 81, + 668, + 291, + 681 + ] + }, + { + "type": "text", + "bbox": [ + 82, + 686, + 290, + 739 + ], + "lines": [ + { + "bbox": [ + 81, + 684, + 291, + 700 + ], + "spans": [ + { + "bbox": [ + 81, + 684, + 291, + 700 + ], + "score": 1.0, + "content": "• We for the first time explore the possibility", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 91, + 700, + 291, + 713 + ], + "spans": [ + { + "bbox": [ + 91, + 700, + 291, + 713 + ], + "score": 1.0, + "content": "of assessing human personalities by LLMs,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 90, + 713, + 291, + 726 + ], + "spans": [ + { + "bbox": [ + 90, + 713, + 291, + 726 + ], + "score": 1.0, + "content": "and propose a general framework for LLMs", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 90, + 727, + 292, + 739 + ], + "spans": [ + { + "bbox": [ + 90, + 727, + 292, + 739 + ], + "score": 1.0, + "content": "to conduct quantitative evaluations via MBTI.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5, + "bbox_fs": [ + 81, + 684, + 292, + 739 + ] + }, + { + "type": "text", + "bbox": [ + 82, + 747, + 290, + 773 + ], + "lines": [ + { + "bbox": [ + 81, + 746, + 290, + 761 + ], + "spans": [ + { + "bbox": [ + 81, + 746, + 290, + 761 + ], + "score": 1.0, + "content": "• We devise unbiased prompts, subject-replaced", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 90, + 761, + 292, + 774 + ], + "spans": [ + { + "bbox": [ + 90, + 761, + 292, + 774 + ], + "score": 1.0, + "content": "queries, and correctness-evaluated instruc-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 326, + 72, + 525, + 85 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 525, + 85 + ], + "score": 1.0, + "content": "tions to encourage LLMs to perform a reliable", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 326, + 85, + 517, + 99 + ], + "spans": [ + { + "bbox": [ + 326, + 85, + 517, + 99 + ], + "score": 1.0, + "content": "flexible assessment of human personalities.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5, + "bbox_fs": [ + 81, + 746, + 292, + 774 + ] + }, + { + "type": "text", + "bbox": [ + 326, + 72, + 524, + 97 + ], + "lines": [], + "index": 51.5, + "bbox_fs": [ + 326, + 72, + 525, + 99 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 318, + 106, + 525, + 146 + ], + "lines": [ + { + "bbox": [ + 317, + 106, + 527, + 120 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 527, + 120 + ], + "score": 1.0, + "content": "• We propose three evaluation metrics to mea-", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 326, + 121, + 525, + 133 + ], + "spans": [ + { + "bbox": [ + 326, + 121, + 525, + 133 + ], + "score": 1.0, + "content": "sure the consistency, robustness, and fairness", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 326, + 134, + 515, + 147 + ], + "spans": [ + { + "bbox": [ + 326, + 134, + 515, + 147 + ], + "score": 1.0, + "content": "of LLMs in assessing human personalities.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 54, + "bbox_fs": [ + 317, + 106, + 527, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 318, + 155, + 526, + 249 + ], + "lines": [ + { + "bbox": [ + 318, + 155, + 526, + 168 + ], + "spans": [ + { + "bbox": [ + 318, + 155, + 526, + 168 + ], + "score": 1.0, + "content": "• Our experiments show that both ChatGPT and", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 326, + 169, + 527, + 182 + ], + "spans": [ + { + "bbox": [ + 326, + 169, + 527, + 182 + ], + "score": 1.0, + "content": "its counterparts can independently assess hu-", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 326, + 183, + 527, + 195 + ], + "spans": [ + { + "bbox": [ + 326, + 183, + 527, + 195 + ], + "score": 1.0, + "content": "man personalities. The average results demon-", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 326, + 196, + 525, + 208 + ], + "spans": [ + { + "bbox": [ + 326, + 196, + 525, + 208 + ], + "score": 1.0, + "content": "strate that ChatGPT and GPT-4 achieve more", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 326, + 210, + 525, + 221 + ], + "spans": [ + { + "bbox": [ + 326, + 210, + 525, + 221 + ], + "score": 1.0, + "content": "consistent and fairer assessments with less", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 326, + 223, + 527, + 236 + ], + "spans": [ + { + "bbox": [ + 326, + 223, + 527, + 236 + ], + "score": 1.0, + "content": "gender bias than InstructGPT, while their re-", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 326, + 237, + 508, + 250 + ], + "spans": [ + { + "bbox": [ + 326, + 237, + 508, + 250 + ], + "score": 1.0, + "content": "sults are more sensitive to prompt biases.", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 59, + "bbox_fs": [ + 318, + 155, + 527, + 250 + ] + }, + { + "type": "title", + "bbox": [ + 305, + 258, + 400, + 272 + ], + "lines": [ + { + "bbox": [ + 302, + 257, + 401, + 274 + ], + "spans": [ + { + "bbox": [ + 302, + 257, + 401, + 274 + ], + "score": 1.0, + "content": "2 Related Works", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 63 + }, + { + "type": "text", + "bbox": [ + 304, + 281, + 525, + 591 + ], + "lines": [ + { + "bbox": [ + 304, + 280, + 526, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 526, + 293 + ], + "score": 1.0, + "content": "Personality Measurement. The commonly-used", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 293, + 526, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 526, + 306 + ], + "score": 1.0, + "content": "personality modeling schemes include the three", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 305, + 308, + 525, + 320 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 525, + 320 + ], + "score": 1.0, + "content": "trait personality measure (Eysenck, 2012), the Big", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 303, + 321, + 527, + 334 + ], + "spans": [ + { + "bbox": [ + 303, + 321, + 527, + 334 + ], + "score": 1.0, + "content": "Five personality trait measure (Digman, 1990),", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 333, + 527, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 527, + 348 + ], + "score": 1.0, + "content": "the Myers–Briggs Type Indicator (MBTI) (Myers,", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 347, + 526, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 526, + 360 + ], + "score": 1.0, + "content": "1962; Myers and McCaulley, 1985), and the 16 Per-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 361, + 527, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 527, + 375 + ], + "score": 1.0, + "content": "sonality Factor questionnaire (16PF) (Schuerger,", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 374, + 526, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 526, + 388 + ], + "score": 1.0, + "content": "2000). Five dimensions are defined in the Big Five", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 389, + 527, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 527, + 401 + ], + "score": 1.0, + "content": "personality traits measure (Digman, 1990) to clas-", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 305, + 402, + 526, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 526, + 415 + ], + "score": 1.0, + "content": "sify major sources of individual differences and", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 416, + 525, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 525, + 428 + ], + "score": 1.0, + "content": "analyze a person’s characteristics. MBTI (Myers", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 429, + 525, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 525, + 442 + ], + "score": 1.0, + "content": "and McCaulley, 1985) identifies personality from", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 304, + 442, + 525, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 525, + 455 + ], + "score": 1.0, + "content": "the differences between persons on the preference", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 457, + 526, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 526, + 469 + ], + "score": 1.0, + "content": "to use perception and judgment. (Karra et al., 2022;", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 304, + 470, + 525, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 525, + 483 + ], + "score": 1.0, + "content": "Caron and Srivastava, 2022) leverage the Big Five", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 484, + 526, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 526, + 496 + ], + "score": 1.0, + "content": "trait theory to quantify the personality traits of lan-", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 303, + 497, + 526, + 510 + ], + "spans": [ + { + "bbox": [ + 303, + 497, + 526, + 510 + ], + "score": 1.0, + "content": "guage models, while (Jiang et al., 2022) further", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 511, + 526, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 526, + 524 + ], + "score": 1.0, + "content": "develops machine personality inventory to stan-", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 304, + 523, + 526, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 526, + 537 + ], + "score": 1.0, + "content": "dardize this evaluation. In (Li et al., 2022), multi-", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 304, + 538, + 525, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 525, + 551 + ], + "score": 1.0, + "content": "ple psychological tests are combined to analyze the", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 303, + 551, + 525, + 564 + ], + "spans": [ + { + "bbox": [ + 303, + 551, + 525, + 564 + ], + "score": 1.0, + "content": "LLMs’ safety. Unlike existing studies that evaluate", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 304, + 565, + 525, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 525, + 578 + ], + "score": 1.0, + "content": "personalities of LLMs, our work is the first attempt", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 304, + 579, + 521, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 521, + 591 + ], + "score": 1.0, + "content": "to explore human personality analysis via LLMs.", + "type": "text" + } + ], + "index": 86 + } + ], + "index": 75, + "bbox_fs": [ + 303, + 280, + 527, + 591 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 598, + 526, + 773 + ], + "lines": [ + { + "bbox": [ + 304, + 598, + 527, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 527, + 610 + ], + "score": 1.0, + "content": "Biases in Language Models. Most recent lan-", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 304, + 612, + 525, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 525, + 625 + ], + "score": 1.0, + "content": "guage models are pre-trained on the large-scale", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 304, + 626, + 525, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 525, + 638 + ], + "score": 1.0, + "content": "datasets or Internet texts that usually contains", + "type": "text" + } + ], + "index": 89 + }, + { + "bbox": [ + 304, + 640, + 526, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 526, + 652 + ], + "score": 1.0, + "content": "unsafe (e.g., toxic) contents, which may cause", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 304, + 652, + 527, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 527, + 666 + ], + "score": 1.0, + "content": "the model to generate biased answers that vio-", + "type": "text" + } + ], + "index": 91 + }, + { + "bbox": [ + 304, + 667, + 526, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 667, + 526, + 679 + ], + "score": 1.0, + "content": "late prevailing societal values (Bolukbasi et al.,", + "type": "text" + } + ], + "index": 92 + }, + { + "bbox": [ + 303, + 678, + 527, + 693 + ], + "spans": [ + { + "bbox": [ + 303, + 678, + 527, + 693 + ], + "score": 1.0, + "content": "2016; Sheng et al., 2019; Bordia and Bowman,", + "type": "text" + } + ], + "index": 93 + }, + { + "bbox": [ + 304, + 693, + 527, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 693, + 527, + 706 + ], + "score": 1.0, + "content": "2019; Nadeem et al., 2021; Zong and Krishna-", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 304, + 707, + 526, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 707, + 526, + 720 + ], + "score": 1.0, + "content": "machari, 2022; Zhuo et al., 2023). (Bolukbasi et al.,", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 304, + 720, + 527, + 734 + ], + "spans": [ + { + "bbox": [ + 304, + 720, + 527, + 734 + ], + "score": 1.0, + "content": "2016) shows that biases in the geometry of word-", + "type": "text" + } + ], + "index": 96 + }, + { + "bbox": [ + 304, + 735, + 525, + 747 + ], + "spans": [ + { + "bbox": [ + 304, + 735, + 525, + 747 + ], + "score": 1.0, + "content": "embeddings can reflect gender stereotypes. The", + "type": "text" + } + ], + "index": 97 + }, + { + "bbox": [ + 304, + 748, + 527, + 761 + ], + "spans": [ + { + "bbox": [ + 304, + 748, + 527, + 761 + ], + "score": 1.0, + "content": "gender bias in word-level language models is quan-", + "type": "text" + } + ], + "index": 98 + }, + { + "bbox": [ + 304, + 761, + 527, + 774 + ], + "spans": [ + { + "bbox": [ + 304, + 761, + 527, + 774 + ], + "score": 1.0, + "content": "titatively evaluated in (Bordia and Bowman, 2019).", + "type": "text" + } + ], + "index": 99 + } + ], + "index": 93, + "bbox_fs": [ + 303, + 598, + 527, + 774 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 69, + 71, + 290, + 274 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 290, + 84 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 290, + 84 + ], + "score": 1.0, + "content": "In (Nadeem et al., 2021), the authors demonstrate", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 86, + 291, + 98 + ], + "spans": [ + { + "bbox": [ + 69, + 86, + 291, + 98 + ], + "score": 1.0, + "content": "that popular LLMs such as GPT-2 (Radford et al.,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 98, + 292, + 114 + ], + "spans": [ + { + "bbox": [ + 68, + 98, + 292, + 114 + ], + "score": 1.0, + "content": "2019) possess strong stereotypical biases on gender,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 113, + 291, + 125 + ], + "spans": [ + { + "bbox": [ + 68, + 113, + 291, + 125 + ], + "score": 1.0, + "content": "profession, race, and religion. To reduce such bi-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 127, + 290, + 137 + ], + "spans": [ + { + "bbox": [ + 68, + 127, + 290, + 137 + ], + "score": 1.0, + "content": "ases, many state-of-the-art LLMs such as ChatGPT", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 140, + 290, + 153 + ], + "spans": [ + { + "bbox": [ + 68, + 140, + 290, + 153 + ], + "score": 1.0, + "content": "apply instruction-finetuning with non-toxic corpora", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 154, + 290, + 166 + ], + "spans": [ + { + "bbox": [ + 69, + 154, + 290, + 166 + ], + "score": 1.0, + "content": "and instructions to improve their safety. (Zhuo", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 167, + 292, + 180 + ], + "spans": [ + { + "bbox": [ + 68, + 167, + 292, + 180 + ], + "score": 1.0, + "content": "et al., 2023) reveals that ChatGPT can generate so-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 181, + 290, + 193 + ], + "spans": [ + { + "bbox": [ + 69, + 181, + 290, + 193 + ], + "score": 1.0, + "content": "cially safe responses with fewer biases than other", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 193, + 290, + 208 + ], + "spans": [ + { + "bbox": [ + 68, + 193, + 290, + 208 + ], + "score": 1.0, + "content": "LLMs under English lanuage settings. In contrast", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 208, + 290, + 219 + ], + "spans": [ + { + "bbox": [ + 69, + 208, + 290, + 219 + ], + "score": 1.0, + "content": "to previous works, our framework enables us to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 220, + 290, + 234 + ], + "spans": [ + { + "bbox": [ + 69, + 220, + 290, + 234 + ], + "score": 1.0, + "content": "evaluate whether LLMs possess biased perceptions", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 235, + 291, + 248 + ], + "spans": [ + { + "bbox": [ + 69, + 235, + 291, + 248 + ], + "score": 1.0, + "content": "and assessments on humans (e.g., personalities),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 248, + 290, + 261 + ], + "spans": [ + { + "bbox": [ + 69, + 248, + 290, + 261 + ], + "score": 1.0, + "content": "which helps us better understand the underlying", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 261, + 256, + 275 + ], + "spans": [ + { + "bbox": [ + 69, + 261, + 256, + 275 + ], + "score": 1.0, + "content": "reasons for the LLMs’ aberrant responses.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 329, + 69, + 500, + 184 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 329, + 69, + 500, + 184 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 329, + 69, + 500, + 184 + ], + "spans": [ + { + "bbox": [ + 329, + 69, + 500, + 184 + ], + "score": 0.968, + "type": "image", + "image_path": "241a0262a34b9fd2f157ee10d274aeee1c95904d5edb217c047fb61389fdb570.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 329, + 69, + 500, + 83.375 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 329, + 83.375, + 500, + 97.75 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 329, + 97.75, + 500, + 112.125 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 329, + 112.125, + 500, + 126.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 329, + 126.5, + 500, + 140.875 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 329, + 140.875, + 500, + 155.25 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 329, + 155.25, + 500, + 169.625 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 329, + 169.625, + 500, + 184.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 192, + 525, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 525, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 525, + 205 + ], + "score": 1.0, + "content": "Figure 1: Overview of our framework: (a) The queried", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 204, + 526, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 526, + 216 + ], + "score": 1.0, + "content": "subject is replaced in the original statements of MBTI", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 217, + 526, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 526, + 228 + ], + "score": 1.0, + "content": "questions; (b) We construct correctness-evaluated in-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 229, + 525, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 525, + 240 + ], + "score": 1.0, + "content": "structions and (c) randomly permute options to build", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 241, + 525, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 525, + 252 + ], + "score": 1.0, + "content": "unbiased prompts with the subject-replaced statements", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 303, + 251, + 525, + 266 + ], + "spans": [ + { + "bbox": [ + 303, + 251, + 525, + 266 + ], + "score": 1.0, + "content": "(d), which are assessed by LLMs to infer the personality.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5 + } + ], + "index": 22.0 + }, + { + "type": "title", + "bbox": [ + 71, + 285, + 222, + 299 + ], + "lines": [ + { + "bbox": [ + 66, + 283, + 224, + 303 + ], + "spans": [ + { + "bbox": [ + 66, + 283, + 224, + 303 + ], + "score": 1.0, + "content": "3 The Proposed Framework", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 70, + 308, + 212, + 321 + ], + "lines": [ + { + "bbox": [ + 67, + 306, + 213, + 324 + ], + "spans": [ + { + "bbox": [ + 67, + 306, + 213, + 324 + ], + "score": 1.0, + "content": "3.1 Unbiased Prompt Design", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 69, + 326, + 290, + 501 + ], + "lines": [ + { + "bbox": [ + 68, + 325, + 291, + 342 + ], + "spans": [ + { + "bbox": [ + 68, + 325, + 291, + 342 + ], + "score": 1.0, + "content": "LLMs are typically sensitive to prompt biases (e.g.,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 340, + 291, + 353 + ], + "spans": [ + { + "bbox": [ + 68, + 340, + 291, + 353 + ], + "score": 1.0, + "content": "varying word orders), which can significantly influ-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 69, + 354, + 290, + 367 + ], + "spans": [ + { + "bbox": [ + 69, + 354, + 290, + 367 + ], + "score": 1.0, + "content": "ence the coherence and accuracy of the generated", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 68, + 368, + 290, + 381 + ], + "spans": [ + { + "bbox": [ + 68, + 368, + 290, + 381 + ], + "score": 1.0, + "content": "responses especially when dealing with long text", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 68, + 381, + 290, + 395 + ], + "spans": [ + { + "bbox": [ + 68, + 381, + 290, + 395 + ], + "score": 1.0, + "content": "sequences (Zhao et al., 2021). To encourage more", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 69, + 394, + 290, + 408 + ], + "spans": [ + { + "bbox": [ + 69, + 394, + 290, + 408 + ], + "score": 1.0, + "content": "consistent and impartial answers, we propose to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 68, + 408, + 292, + 422 + ], + "spans": [ + { + "bbox": [ + 68, + 408, + 292, + 422 + ], + "score": 1.0, + "content": "design unbiased prompts for the input questions.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 69, + 421, + 290, + 434 + ], + "spans": [ + { + "bbox": [ + 69, + 421, + 290, + 434 + ], + "score": 1.0, + "content": "In particular, for each question in an independent", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 69, + 435, + 290, + 448 + ], + "spans": [ + { + "bbox": [ + 69, + 435, + 290, + 448 + ], + "score": 1.0, + "content": "testing (i.e., MBTI questionnaire), we randomly", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 68, + 448, + 290, + 462 + ], + "spans": [ + { + "bbox": [ + 68, + 448, + 290, + 462 + ], + "score": 1.0, + "content": "permute all available options (e.g., agree, disagree)", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 68, + 462, + 290, + 476 + ], + "spans": [ + { + "bbox": [ + 68, + 462, + 290, + 476 + ], + "score": 1.0, + "content": "in its instruction while not changing the question", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 69, + 475, + 290, + 489 + ], + "spans": [ + { + "bbox": [ + 69, + 475, + 290, + 489 + ], + "score": 1.0, + "content": "statement, and adopt the average results of multiple", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 69, + 489, + 240, + 502 + ], + "spans": [ + { + "bbox": [ + 69, + 489, + 240, + 502 + ], + "score": 1.0, + "content": "independent testings as the final result.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 69, + 503, + 290, + 731 + ], + "lines": [ + { + "bbox": [ + 80, + 503, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 80, + 503, + 290, + 515 + ], + "score": 1.0, + "content": "Formally, the instruction and statement for the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 69, + 513, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 69, + 515, + 83, + 528 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 83, + 513, + 192, + 531 + ], + "score": 1.0, + "content": "question are defined as", + "type": "text" + }, + { + "bbox": [ + 193, + 517, + 203, + 528 + ], + "score": 0.87, + "content": "I _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 513, + 224, + 531 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 225, + 516, + 236, + 529 + ], + "score": 0.87, + "content": "S _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 513, + 271, + 531 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 272, + 516, + 290, + 528 + ], + "score": 0.86, + "content": "i \\in", + "type": "inline_equation" + } + ], + "index": 45 + }, + { + "bbox": [ + 70, + 529, + 291, + 543 + ], + "spans": [ + { + "bbox": [ + 70, + 529, + 118, + 543 + ], + "score": 0.93, + "content": "\\{ 1 , \\cdots , n \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 530, + 142, + 543 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 143, + 532, + 151, + 541 + ], + "score": 0.69, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 530, + 291, + 543 + ], + "score": 1.0, + "content": "is the total number of ques-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 69, + 544, + 290, + 556 + ], + "spans": [ + { + "bbox": [ + 69, + 544, + 199, + 556 + ], + "score": 1.0, + "content": "tions in the testing. We have", + "type": "text" + }, + { + "bbox": [ + 200, + 546, + 211, + 555 + ], + "score": 0.61, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 544, + 290, + 556 + ], + "score": 1.0, + "content": "available options", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 69, + 556, + 291, + 571 + ], + "spans": [ + { + "bbox": [ + 69, + 557, + 175, + 571 + ], + "score": 0.91, + "content": "O _ { I } = \\{ o _ { 1 } , o _ { 2 } , \\cdots , o _ { m } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 556, + 291, + 571 + ], + "score": 1.0, + "content": "in the instruction, which", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 69, + 570, + 290, + 584 + ], + "spans": [ + { + "bbox": [ + 69, + 570, + 137, + 584 + ], + "score": 1.0, + "content": "corresponds to", + "type": "text" + }, + { + "bbox": [ + 137, + 571, + 169, + 584 + ], + "score": 0.8, + "content": "\\{ A g r e e", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 570, + 290, + 584 + ], + "score": 1.0, + "content": ", Generally agree, Partially", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 68, + 585, + 292, + 597 + ], + "spans": [ + { + "bbox": [ + 68, + 585, + 292, + 597 + ], + "score": 1.0, + "content": "agree, Neither agree nor disagree, Partially dis-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 69, + 598, + 291, + 612 + ], + "spans": [ + { + "bbox": [ + 69, + 598, + 291, + 612 + ], + "score": 1.0, + "content": "agree, Generally disagree, Disagree} including", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 69, + 612, + 291, + 624 + ], + "spans": [ + { + "bbox": [ + 69, + 612, + 150, + 624 + ], + "score": 1.0, + "content": "seven levels (i.e.,", + "type": "text" + }, + { + "bbox": [ + 150, + 612, + 182, + 623 + ], + "score": 0.86, + "content": "m = 7", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 612, + 291, + 624 + ], + "score": 1.0, + "content": ") from agreement to dis-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 68, + 625, + 291, + 638 + ], + "spans": [ + { + "bbox": [ + 68, + 625, + 291, + 638 + ], + "score": 1.0, + "content": "agreement in the MBTI questionnaire. We use", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 69, + 636, + 292, + 653 + ], + "spans": [ + { + "bbox": [ + 69, + 638, + 100, + 651 + ], + "score": 0.92, + "content": "\\Omega ( O _ { I } )", + "type": "inline_equation" + }, + { + "bbox": [ + 100, + 636, + 292, + 653 + ], + "score": 1.0, + "content": "to denote all possible permutations of op-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 68, + 651, + 291, + 666 + ], + "spans": [ + { + "bbox": [ + 68, + 651, + 104, + 666 + ], + "score": 1.0, + "content": "tions in", + "type": "text" + }, + { + "bbox": [ + 105, + 652, + 119, + 664 + ], + "score": 0.88, + "content": "O _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 651, + 291, + 666 + ], + "score": 1.0, + "content": ", and a random permutation can be rep-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 67, + 663, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 67, + 663, + 122, + 680 + ], + "score": 1.0, + "content": "resented as", + "type": "text" + }, + { + "bbox": [ + 123, + 665, + 288, + 679 + ], + "score": 0.91, + "content": "{ \\cal O } _ { \\mathcal { R } } = \\{ o _ { r _ { 1 } } , o _ { r _ { 2 } } , \\cdots , o _ { r _ { m } } \\} \\in \\Omega ( { \\cal O } _ { I } )", + "type": "inline_equation" + } + ], + "index": 56 + }, + { + "bbox": [ + 68, + 678, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 68, + 678, + 99, + 693 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 99, + 679, + 182, + 692 + ], + "score": 0.92, + "content": "r _ { i } \\in \\{ 1 , 2 , \\cdots , m \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 678, + 206, + 693 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 206, + 680, + 248, + 693 + ], + "score": 0.92, + "content": "o _ { r _ { i } } \\neq o _ { r _ { j } }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 678, + 263, + 693 + ], + "score": 1.0, + "content": "iff", + "type": "text" + }, + { + "bbox": [ + 263, + 680, + 288, + 692 + ], + "score": 0.91, + "content": "i \\neq j", + "type": "inline_equation" + } + ], + "index": 57 + }, + { + "bbox": [ + 68, + 692, + 291, + 706 + ], + "spans": [ + { + "bbox": [ + 68, + 692, + 291, + 706 + ], + "score": 1.0, + "content": "Then, we utilize the randomly permuted options", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 69, + 704, + 293, + 720 + ], + "spans": [ + { + "bbox": [ + 69, + 706, + 87, + 718 + ], + "score": 0.88, + "content": "O _ { \\mathcal { R } }", + "type": "inline_equation" + }, + { + "bbox": [ + 87, + 704, + 208, + 720 + ], + "score": 1.0, + "content": "to construct the instruction", + "type": "text" + }, + { + "bbox": [ + 208, + 707, + 218, + 718 + ], + "score": 0.87, + "content": "I _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 704, + 250, + 720 + ], + "score": 1.0, + "content": "for the", + "type": "text" + }, + { + "bbox": [ + 250, + 705, + 264, + 717 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 704, + 293, + 720 + ], + "score": 1.0, + "content": "ques-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 69, + 719, + 113, + 731 + ], + "spans": [ + { + "bbox": [ + 69, + 719, + 113, + 731 + ], + "score": 1.0, + "content": "tion with:", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 52 + }, + { + "type": "text", + "bbox": [ + 69, + 734, + 289, + 760 + ], + "lines": [ + { + "bbox": [ + 67, + 731, + 291, + 750 + ], + "spans": [ + { + "bbox": [ + 67, + 731, + 166, + 750 + ], + "score": 1.0, + "content": "Instruction: Do you", + "type": "text" + }, + { + "bbox": [ + 167, + 736, + 218, + 747 + ], + "score": 0.84, + "content": "o _ { r _ { 1 } } , o _ { r _ { 2 } } , \\cdots", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 731, + 231, + 750 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 232, + 736, + 249, + 747 + ], + "score": 0.88, + "content": "o _ { r _ { m } }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 731, + 291, + 750 + ], + "score": 1.0, + "content": "with the", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 68, + 746, + 191, + 762 + ], + "spans": [ + { + "bbox": [ + 68, + 746, + 191, + 762 + ], + "score": 1.0, + "content": "following statement. Why?", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 61.5 + }, + { + "type": "text", + "bbox": [ + 81, + 761, + 289, + 774 + ], + "lines": [ + { + "bbox": [ + 80, + 759, + 291, + 775 + ], + "spans": [ + { + "bbox": [ + 80, + 759, + 291, + 775 + ], + "score": 1.0, + "content": "We combine the above instruction and the ques-", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 63 + }, + { + "type": "text", + "bbox": [ + 305, + 286, + 525, + 326 + ], + "lines": [ + { + "bbox": [ + 304, + 286, + 525, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 525, + 300 + ], + "score": 1.0, + "content": "tion statement as the prompt to query LLMs. An", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 301, + 527, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 527, + 313 + ], + "score": 1.0, + "content": "example prompt for a question in the MBTI ques-", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 313, + 447, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 447, + 326 + ], + "score": 1.0, + "content": "tionnaire is provided as follows.", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 65 + }, + { + "type": "text", + "bbox": [ + 304, + 327, + 525, + 380 + ], + "lines": [ + { + "bbox": [ + 303, + 326, + 525, + 341 + ], + "spans": [ + { + "bbox": [ + 303, + 326, + 525, + 341 + ], + "score": 1.0, + "content": "Instruction: Do you disagree, agree, generally", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 341, + 527, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 527, + 354 + ], + "score": 1.0, + "content": "disagree, generally agree, neither agree nor dis-", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 355, + 525, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 525, + 367 + ], + "score": 1.0, + "content": "agree, partially disagree or partially agree with", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 367, + 442, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 442, + 382 + ], + "score": 1.0, + "content": "the following statement. Why?", + "type": "text" + } + ], + "index": 70 + } + ], + "index": 68.5 + }, + { + "type": "text", + "bbox": [ + 305, + 382, + 525, + 407 + ], + "lines": [ + { + "bbox": [ + 303, + 380, + 527, + 395 + ], + "spans": [ + { + "bbox": [ + 303, + 380, + 527, + 395 + ], + "score": 1.0, + "content": "Statement: You usually stay calm, even under a", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 394, + 373, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 373, + 408 + ], + "score": 1.0, + "content": "lot of pressure.", + "type": "text" + } + ], + "index": 72 + } + ], + "index": 71.5 + }, + { + "type": "text", + "bbox": [ + 304, + 409, + 525, + 489 + ], + "lines": [ + { + "bbox": [ + 303, + 407, + 526, + 422 + ], + "spans": [ + { + "bbox": [ + 303, + 407, + 526, + 422 + ], + "score": 1.0, + "content": "Answer: “As an AI language model, I do not", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 423, + 526, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 526, + 435 + ], + "score": 1.0, + "content": "have personal experiences or emotions, so I cannot", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 437, + 526, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 526, + 448 + ], + "score": 1.0, + "content": "agree or disagree with the statement. However, I", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 304, + 450, + 527, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 527, + 462 + ], + "score": 1.0, + "content": "can provide some general information on the topic.", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 462, + 527, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 527, + 477 + ], + "score": 1.0, + "content": "Some individuals are naturally more composed un-", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 304, + 476, + 464, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 464, + 490 + ], + "score": 1.0, + "content": "der pressure, while others may ......”", + "type": "text" + } + ], + "index": 78 + } + ], + "index": 75.5 + }, + { + "type": "text", + "bbox": [ + 305, + 490, + 526, + 583 + ], + "lines": [ + { + "bbox": [ + 315, + 490, + 527, + 503 + ], + "spans": [ + { + "bbox": [ + 315, + 490, + 527, + 503 + ], + "score": 1.0, + "content": "However, such a query, conducted in a self-", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 504, + 526, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 526, + 516 + ], + "score": 1.0, + "content": "testing manner, can only elicit neutral answers as", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 516, + 526, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 526, + 529 + ], + "score": 1.0, + "content": "shown above, since LLMs such as ChatGPT are", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 304, + 531, + 527, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 527, + 544 + ], + "score": 1.0, + "content": "trained to not possess personal thinking (e.g., emo-", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 304, + 544, + 527, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 527, + 558 + ], + "score": 1.0, + "content": "tions). This motivates us to propose the subject-", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 304, + 558, + 528, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 528, + 570 + ], + "score": 1.0, + "content": "replaced query and correctness-evaluated instruc-", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 304, + 571, + 413, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 413, + 583 + ], + "score": 1.0, + "content": "tion as illustrated below.", + "type": "text" + } + ], + "index": 85 + } + ], + "index": 82 + }, + { + "type": "title", + "bbox": [ + 305, + 594, + 445, + 607 + ], + "lines": [ + { + "bbox": [ + 303, + 591, + 446, + 610 + ], + "spans": [ + { + "bbox": [ + 303, + 591, + 446, + 610 + ], + "score": 1.0, + "content": "3.2 Subject-Replaced Query", + "type": "text" + } + ], + "index": 86 + } + ], + "index": 86 + }, + { + "type": "text", + "bbox": [ + 304, + 612, + 525, + 773 + ], + "lines": [ + { + "bbox": [ + 304, + 612, + 527, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 527, + 625 + ], + "score": 1.0, + "content": "As our goal is to let LLMs analyze human personal-", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 303, + 625, + 527, + 639 + ], + "spans": [ + { + "bbox": [ + 303, + 625, + 527, + 639 + ], + "score": 1.0, + "content": "ities instead of querying itself (i.e., self-reporting),", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 304, + 640, + 526, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 526, + 653 + ], + "score": 1.0, + "content": "we propose the subject-replaced query (SRQ) by", + "type": "text" + } + ], + "index": 89 + }, + { + "bbox": [ + 303, + 653, + 525, + 666 + ], + "spans": [ + { + "bbox": [ + 303, + 653, + 525, + 666 + ], + "score": 1.0, + "content": "converting the original subject (i.e., “You”) of each", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 304, + 667, + 527, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 667, + 527, + 679 + ], + "score": 1.0, + "content": "question into a specific subject-of-interest. For ex-", + "type": "text" + } + ], + "index": 91 + }, + { + "bbox": [ + 304, + 680, + 527, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 680, + 527, + 693 + ], + "score": 1.0, + "content": "ample, when we hope to let LLMs assess the gen-", + "type": "text" + } + ], + "index": 92 + }, + { + "bbox": [ + 304, + 694, + 526, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 694, + 526, + 707 + ], + "score": 1.0, + "content": "eral personality of men, we can replace the subject", + "type": "text" + } + ], + "index": 93 + }, + { + "bbox": [ + 303, + 705, + 526, + 720 + ], + "spans": [ + { + "bbox": [ + 303, + 705, + 526, + 720 + ], + "score": 1.0, + "content": "“You” with “Men”, and correspondingly change the", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 304, + 721, + 527, + 734 + ], + "spans": [ + { + "bbox": [ + 304, + 721, + 527, + 734 + ], + "score": 1.0, + "content": "pronoun “your” to “their” (see the example below).", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 303, + 733, + 526, + 748 + ], + "spans": [ + { + "bbox": [ + 303, + 733, + 526, + 748 + ], + "score": 1.0, + "content": "Original Statement: You spend a lot of your free", + "type": "text" + } + ], + "index": 96 + }, + { + "bbox": [ + 304, + 747, + 525, + 761 + ], + "spans": [ + { + "bbox": [ + 304, + 747, + 525, + 761 + ], + "score": 1.0, + "content": "time exploring various random topics that pique", + "type": "text" + } + ], + "index": 97 + }, + { + "bbox": [ + 303, + 762, + 366, + 774 + ], + "spans": [ + { + "bbox": [ + 303, + 762, + 366, + 774 + ], + "score": 1.0, + "content": "your interest.", + "type": "text" + } + ], + "index": 98 + } + ], + "index": 92.5 + } + ], + "page_idx": 2, + "page_size": [ + 595, + 841 + ], + "discarded_blocks": [], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 69, + 71, + 290, + 274 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 290, + 84 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 290, + 84 + ], + "score": 1.0, + "content": "In (Nadeem et al., 2021), the authors demonstrate", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 86, + 291, + 98 + ], + "spans": [ + { + "bbox": [ + 69, + 86, + 291, + 98 + ], + "score": 1.0, + "content": "that popular LLMs such as GPT-2 (Radford et al.,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 98, + 292, + 114 + ], + "spans": [ + { + "bbox": [ + 68, + 98, + 292, + 114 + ], + "score": 1.0, + "content": "2019) possess strong stereotypical biases on gender,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 113, + 291, + 125 + ], + "spans": [ + { + "bbox": [ + 68, + 113, + 291, + 125 + ], + "score": 1.0, + "content": "profession, race, and religion. To reduce such bi-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 127, + 290, + 137 + ], + "spans": [ + { + "bbox": [ + 68, + 127, + 290, + 137 + ], + "score": 1.0, + "content": "ases, many state-of-the-art LLMs such as ChatGPT", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 140, + 290, + 153 + ], + "spans": [ + { + "bbox": [ + 68, + 140, + 290, + 153 + ], + "score": 1.0, + "content": "apply instruction-finetuning with non-toxic corpora", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 154, + 290, + 166 + ], + "spans": [ + { + "bbox": [ + 69, + 154, + 290, + 166 + ], + "score": 1.0, + "content": "and instructions to improve their safety. (Zhuo", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 167, + 292, + 180 + ], + "spans": [ + { + "bbox": [ + 68, + 167, + 292, + 180 + ], + "score": 1.0, + "content": "et al., 2023) reveals that ChatGPT can generate so-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 181, + 290, + 193 + ], + "spans": [ + { + "bbox": [ + 69, + 181, + 290, + 193 + ], + "score": 1.0, + "content": "cially safe responses with fewer biases than other", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 193, + 290, + 208 + ], + "spans": [ + { + "bbox": [ + 68, + 193, + 290, + 208 + ], + "score": 1.0, + "content": "LLMs under English lanuage settings. In contrast", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 208, + 290, + 219 + ], + "spans": [ + { + "bbox": [ + 69, + 208, + 290, + 219 + ], + "score": 1.0, + "content": "to previous works, our framework enables us to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 220, + 290, + 234 + ], + "spans": [ + { + "bbox": [ + 69, + 220, + 290, + 234 + ], + "score": 1.0, + "content": "evaluate whether LLMs possess biased perceptions", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 235, + 291, + 248 + ], + "spans": [ + { + "bbox": [ + 69, + 235, + 291, + 248 + ], + "score": 1.0, + "content": "and assessments on humans (e.g., personalities),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 248, + 290, + 261 + ], + "spans": [ + { + "bbox": [ + 69, + 248, + 290, + 261 + ], + "score": 1.0, + "content": "which helps us better understand the underlying", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 261, + 256, + 275 + ], + "spans": [ + { + "bbox": [ + 69, + 261, + 256, + 275 + ], + "score": 1.0, + "content": "reasons for the LLMs’ aberrant responses.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 7, + "bbox_fs": [ + 68, + 72, + 292, + 275 + ] + }, + { + "type": "image", + "bbox": [ + 329, + 69, + 500, + 184 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 329, + 69, + 500, + 184 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 329, + 69, + 500, + 184 + ], + "spans": [ + { + "bbox": [ + 329, + 69, + 500, + 184 + ], + "score": 0.968, + "type": "image", + "image_path": "241a0262a34b9fd2f157ee10d274aeee1c95904d5edb217c047fb61389fdb570.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 329, + 69, + 500, + 83.375 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 329, + 83.375, + 500, + 97.75 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 329, + 97.75, + 500, + 112.125 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 329, + 112.125, + 500, + 126.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 329, + 126.5, + 500, + 140.875 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 329, + 140.875, + 500, + 155.25 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 329, + 155.25, + 500, + 169.625 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 329, + 169.625, + 500, + 184.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 192, + 525, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 525, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 525, + 205 + ], + "score": 1.0, + "content": "Figure 1: Overview of our framework: (a) The queried", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 204, + 526, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 526, + 216 + ], + "score": 1.0, + "content": "subject is replaced in the original statements of MBTI", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 217, + 526, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 526, + 228 + ], + "score": 1.0, + "content": "questions; (b) We construct correctness-evaluated in-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 229, + 525, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 525, + 240 + ], + "score": 1.0, + "content": "structions and (c) randomly permute options to build", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 241, + 525, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 525, + 252 + ], + "score": 1.0, + "content": "unbiased prompts with the subject-replaced statements", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 303, + 251, + 525, + 266 + ], + "spans": [ + { + "bbox": [ + 303, + 251, + 525, + 266 + ], + "score": 1.0, + "content": "(d), which are assessed by LLMs to infer the personality.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5 + } + ], + "index": 22.0 + }, + { + "type": "title", + "bbox": [ + 71, + 285, + 222, + 299 + ], + "lines": [ + { + "bbox": [ + 66, + 283, + 224, + 303 + ], + "spans": [ + { + "bbox": [ + 66, + 283, + 224, + 303 + ], + "score": 1.0, + "content": "3 The Proposed Framework", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 70, + 308, + 212, + 321 + ], + "lines": [ + { + "bbox": [ + 67, + 306, + 213, + 324 + ], + "spans": [ + { + "bbox": [ + 67, + 306, + 213, + 324 + ], + "score": 1.0, + "content": "3.1 Unbiased Prompt Design", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 69, + 326, + 290, + 501 + ], + "lines": [ + { + "bbox": [ + 68, + 325, + 291, + 342 + ], + "spans": [ + { + "bbox": [ + 68, + 325, + 291, + 342 + ], + "score": 1.0, + "content": "LLMs are typically sensitive to prompt biases (e.g.,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 340, + 291, + 353 + ], + "spans": [ + { + "bbox": [ + 68, + 340, + 291, + 353 + ], + "score": 1.0, + "content": "varying word orders), which can significantly influ-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 69, + 354, + 290, + 367 + ], + "spans": [ + { + "bbox": [ + 69, + 354, + 290, + 367 + ], + "score": 1.0, + "content": "ence the coherence and accuracy of the generated", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 68, + 368, + 290, + 381 + ], + "spans": [ + { + "bbox": [ + 68, + 368, + 290, + 381 + ], + "score": 1.0, + "content": "responses especially when dealing with long text", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 68, + 381, + 290, + 395 + ], + "spans": [ + { + "bbox": [ + 68, + 381, + 290, + 395 + ], + "score": 1.0, + "content": "sequences (Zhao et al., 2021). To encourage more", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 69, + 394, + 290, + 408 + ], + "spans": [ + { + "bbox": [ + 69, + 394, + 290, + 408 + ], + "score": 1.0, + "content": "consistent and impartial answers, we propose to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 68, + 408, + 292, + 422 + ], + "spans": [ + { + "bbox": [ + 68, + 408, + 292, + 422 + ], + "score": 1.0, + "content": "design unbiased prompts for the input questions.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 69, + 421, + 290, + 434 + ], + "spans": [ + { + "bbox": [ + 69, + 421, + 290, + 434 + ], + "score": 1.0, + "content": "In particular, for each question in an independent", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 69, + 435, + 290, + 448 + ], + "spans": [ + { + "bbox": [ + 69, + 435, + 290, + 448 + ], + "score": 1.0, + "content": "testing (i.e., MBTI questionnaire), we randomly", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 68, + 448, + 290, + 462 + ], + "spans": [ + { + "bbox": [ + 68, + 448, + 290, + 462 + ], + "score": 1.0, + "content": "permute all available options (e.g., agree, disagree)", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 68, + 462, + 290, + 476 + ], + "spans": [ + { + "bbox": [ + 68, + 462, + 290, + 476 + ], + "score": 1.0, + "content": "in its instruction while not changing the question", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 69, + 475, + 290, + 489 + ], + "spans": [ + { + "bbox": [ + 69, + 475, + 290, + 489 + ], + "score": 1.0, + "content": "statement, and adopt the average results of multiple", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 69, + 489, + 240, + 502 + ], + "spans": [ + { + "bbox": [ + 69, + 489, + 240, + 502 + ], + "score": 1.0, + "content": "independent testings as the final result.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 37, + "bbox_fs": [ + 68, + 325, + 292, + 502 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 503, + 290, + 731 + ], + "lines": [ + { + "bbox": [ + 80, + 503, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 80, + 503, + 290, + 515 + ], + "score": 1.0, + "content": "Formally, the instruction and statement for the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 69, + 513, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 69, + 515, + 83, + 528 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 83, + 513, + 192, + 531 + ], + "score": 1.0, + "content": "question are defined as", + "type": "text" + }, + { + "bbox": [ + 193, + 517, + 203, + 528 + ], + "score": 0.87, + "content": "I _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 513, + 224, + 531 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 225, + 516, + 236, + 529 + ], + "score": 0.87, + "content": "S _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 513, + 271, + 531 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 272, + 516, + 290, + 528 + ], + "score": 0.86, + "content": "i \\in", + "type": "inline_equation" + } + ], + "index": 45 + }, + { + "bbox": [ + 70, + 529, + 291, + 543 + ], + "spans": [ + { + "bbox": [ + 70, + 529, + 118, + 543 + ], + "score": 0.93, + "content": "\\{ 1 , \\cdots , n \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 530, + 142, + 543 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 143, + 532, + 151, + 541 + ], + "score": 0.69, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 530, + 291, + 543 + ], + "score": 1.0, + "content": "is the total number of ques-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 69, + 544, + 290, + 556 + ], + "spans": [ + { + "bbox": [ + 69, + 544, + 199, + 556 + ], + "score": 1.0, + "content": "tions in the testing. We have", + "type": "text" + }, + { + "bbox": [ + 200, + 546, + 211, + 555 + ], + "score": 0.61, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 544, + 290, + 556 + ], + "score": 1.0, + "content": "available options", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 69, + 556, + 291, + 571 + ], + "spans": [ + { + "bbox": [ + 69, + 557, + 175, + 571 + ], + "score": 0.91, + "content": "O _ { I } = \\{ o _ { 1 } , o _ { 2 } , \\cdots , o _ { m } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 556, + 291, + 571 + ], + "score": 1.0, + "content": "in the instruction, which", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 69, + 570, + 290, + 584 + ], + "spans": [ + { + "bbox": [ + 69, + 570, + 137, + 584 + ], + "score": 1.0, + "content": "corresponds to", + "type": "text" + }, + { + "bbox": [ + 137, + 571, + 169, + 584 + ], + "score": 0.8, + "content": "\\{ A g r e e", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 570, + 290, + 584 + ], + "score": 1.0, + "content": ", Generally agree, Partially", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 68, + 585, + 292, + 597 + ], + "spans": [ + { + "bbox": [ + 68, + 585, + 292, + 597 + ], + "score": 1.0, + "content": "agree, Neither agree nor disagree, Partially dis-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 69, + 598, + 291, + 612 + ], + "spans": [ + { + "bbox": [ + 69, + 598, + 291, + 612 + ], + "score": 1.0, + "content": "agree, Generally disagree, Disagree} including", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 69, + 612, + 291, + 624 + ], + "spans": [ + { + "bbox": [ + 69, + 612, + 150, + 624 + ], + "score": 1.0, + "content": "seven levels (i.e.,", + "type": "text" + }, + { + "bbox": [ + 150, + 612, + 182, + 623 + ], + "score": 0.86, + "content": "m = 7", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 612, + 291, + 624 + ], + "score": 1.0, + "content": ") from agreement to dis-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 68, + 625, + 291, + 638 + ], + "spans": [ + { + "bbox": [ + 68, + 625, + 291, + 638 + ], + "score": 1.0, + "content": "agreement in the MBTI questionnaire. We use", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 69, + 636, + 292, + 653 + ], + "spans": [ + { + "bbox": [ + 69, + 638, + 100, + 651 + ], + "score": 0.92, + "content": "\\Omega ( O _ { I } )", + "type": "inline_equation" + }, + { + "bbox": [ + 100, + 636, + 292, + 653 + ], + "score": 1.0, + "content": "to denote all possible permutations of op-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 68, + 651, + 291, + 666 + ], + "spans": [ + { + "bbox": [ + 68, + 651, + 104, + 666 + ], + "score": 1.0, + "content": "tions in", + "type": "text" + }, + { + "bbox": [ + 105, + 652, + 119, + 664 + ], + "score": 0.88, + "content": "O _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 651, + 291, + 666 + ], + "score": 1.0, + "content": ", and a random permutation can be rep-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 67, + 663, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 67, + 663, + 122, + 680 + ], + "score": 1.0, + "content": "resented as", + "type": "text" + }, + { + "bbox": [ + 123, + 665, + 288, + 679 + ], + "score": 0.91, + "content": "{ \\cal O } _ { \\mathcal { R } } = \\{ o _ { r _ { 1 } } , o _ { r _ { 2 } } , \\cdots , o _ { r _ { m } } \\} \\in \\Omega ( { \\cal O } _ { I } )", + "type": "inline_equation" + } + ], + "index": 56 + }, + { + "bbox": [ + 68, + 678, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 68, + 678, + 99, + 693 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 99, + 679, + 182, + 692 + ], + "score": 0.92, + "content": "r _ { i } \\in \\{ 1 , 2 , \\cdots , m \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 678, + 206, + 693 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 206, + 680, + 248, + 693 + ], + "score": 0.92, + "content": "o _ { r _ { i } } \\neq o _ { r _ { j } }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 678, + 263, + 693 + ], + "score": 1.0, + "content": "iff", + "type": "text" + }, + { + "bbox": [ + 263, + 680, + 288, + 692 + ], + "score": 0.91, + "content": "i \\neq j", + "type": "inline_equation" + } + ], + "index": 57 + }, + { + "bbox": [ + 68, + 692, + 291, + 706 + ], + "spans": [ + { + "bbox": [ + 68, + 692, + 291, + 706 + ], + "score": 1.0, + "content": "Then, we utilize the randomly permuted options", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 69, + 704, + 293, + 720 + ], + "spans": [ + { + "bbox": [ + 69, + 706, + 87, + 718 + ], + "score": 0.88, + "content": "O _ { \\mathcal { R } }", + "type": "inline_equation" + }, + { + "bbox": [ + 87, + 704, + 208, + 720 + ], + "score": 1.0, + "content": "to construct the instruction", + "type": "text" + }, + { + "bbox": [ + 208, + 707, + 218, + 718 + ], + "score": 0.87, + "content": "I _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 704, + 250, + 720 + ], + "score": 1.0, + "content": "for the", + "type": "text" + }, + { + "bbox": [ + 250, + 705, + 264, + 717 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 704, + 293, + 720 + ], + "score": 1.0, + "content": "ques-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 69, + 719, + 113, + 731 + ], + "spans": [ + { + "bbox": [ + 69, + 719, + 113, + 731 + ], + "score": 1.0, + "content": "tion with:", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 52, + "bbox_fs": [ + 67, + 503, + 293, + 731 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 734, + 289, + 760 + ], + "lines": [ + { + "bbox": [ + 67, + 731, + 291, + 750 + ], + "spans": [ + { + "bbox": [ + 67, + 731, + 166, + 750 + ], + "score": 1.0, + "content": "Instruction: Do you", + "type": "text" + }, + { + "bbox": [ + 167, + 736, + 218, + 747 + ], + "score": 0.84, + "content": "o _ { r _ { 1 } } , o _ { r _ { 2 } } , \\cdots", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 731, + 231, + 750 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 232, + 736, + 249, + 747 + ], + "score": 0.88, + "content": "o _ { r _ { m } }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 731, + 291, + 750 + ], + "score": 1.0, + "content": "with the", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 68, + 746, + 191, + 762 + ], + "spans": [ + { + "bbox": [ + 68, + 746, + 191, + 762 + ], + "score": 1.0, + "content": "following statement. Why?", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 61.5, + "bbox_fs": [ + 67, + 731, + 291, + 762 + ] + }, + { + "type": "text", + "bbox": [ + 81, + 761, + 289, + 774 + ], + "lines": [ + { + "bbox": [ + 80, + 759, + 291, + 775 + ], + "spans": [ + { + "bbox": [ + 80, + 759, + 291, + 775 + ], + "score": 1.0, + "content": "We combine the above instruction and the ques-", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 286, + 525, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 525, + 300 + ], + "score": 1.0, + "content": "tion statement as the prompt to query LLMs. An", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 301, + 527, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 527, + 313 + ], + "score": 1.0, + "content": "example prompt for a question in the MBTI ques-", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 313, + 447, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 447, + 326 + ], + "score": 1.0, + "content": "tionnaire is provided as follows.", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 63, + "bbox_fs": [ + 80, + 759, + 291, + 775 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 286, + 525, + 326 + ], + "lines": [], + "index": 65, + "bbox_fs": [ + 304, + 286, + 527, + 326 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 304, + 327, + 525, + 380 + ], + "lines": [ + { + "bbox": [ + 303, + 326, + 525, + 341 + ], + "spans": [ + { + "bbox": [ + 303, + 326, + 525, + 341 + ], + "score": 1.0, + "content": "Instruction: Do you disagree, agree, generally", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 341, + 527, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 527, + 354 + ], + "score": 1.0, + "content": "disagree, generally agree, neither agree nor dis-", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 355, + 525, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 525, + 367 + ], + "score": 1.0, + "content": "agree, partially disagree or partially agree with", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 367, + 442, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 442, + 382 + ], + "score": 1.0, + "content": "the following statement. Why?", + "type": "text" + } + ], + "index": 70 + } + ], + "index": 68.5, + "bbox_fs": [ + 303, + 326, + 527, + 382 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 382, + 525, + 407 + ], + "lines": [ + { + "bbox": [ + 303, + 380, + 527, + 395 + ], + "spans": [ + { + "bbox": [ + 303, + 380, + 527, + 395 + ], + "score": 1.0, + "content": "Statement: You usually stay calm, even under a", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 394, + 373, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 373, + 408 + ], + "score": 1.0, + "content": "lot of pressure.", + "type": "text" + } + ], + "index": 72 + } + ], + "index": 71.5, + "bbox_fs": [ + 303, + 380, + 527, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 409, + 525, + 489 + ], + "lines": [ + { + "bbox": [ + 303, + 407, + 526, + 422 + ], + "spans": [ + { + "bbox": [ + 303, + 407, + 526, + 422 + ], + "score": 1.0, + "content": "Answer: “As an AI language model, I do not", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 423, + 526, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 526, + 435 + ], + "score": 1.0, + "content": "have personal experiences or emotions, so I cannot", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 437, + 526, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 526, + 448 + ], + "score": 1.0, + "content": "agree or disagree with the statement. However, I", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 304, + 450, + 527, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 527, + 462 + ], + "score": 1.0, + "content": "can provide some general information on the topic.", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 462, + 527, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 527, + 477 + ], + "score": 1.0, + "content": "Some individuals are naturally more composed un-", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 304, + 476, + 464, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 464, + 490 + ], + "score": 1.0, + "content": "der pressure, while others may ......”", + "type": "text" + } + ], + "index": 78 + } + ], + "index": 75.5, + "bbox_fs": [ + 303, + 407, + 527, + 490 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 490, + 526, + 583 + ], + "lines": [ + { + "bbox": [ + 315, + 490, + 527, + 503 + ], + "spans": [ + { + "bbox": [ + 315, + 490, + 527, + 503 + ], + "score": 1.0, + "content": "However, such a query, conducted in a self-", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 504, + 526, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 526, + 516 + ], + "score": 1.0, + "content": "testing manner, can only elicit neutral answers as", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 516, + 526, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 526, + 529 + ], + "score": 1.0, + "content": "shown above, since LLMs such as ChatGPT are", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 304, + 531, + 527, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 527, + 544 + ], + "score": 1.0, + "content": "trained to not possess personal thinking (e.g., emo-", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 304, + 544, + 527, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 527, + 558 + ], + "score": 1.0, + "content": "tions). This motivates us to propose the subject-", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 304, + 558, + 528, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 528, + 570 + ], + "score": 1.0, + "content": "replaced query and correctness-evaluated instruc-", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 304, + 571, + 413, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 413, + 583 + ], + "score": 1.0, + "content": "tion as illustrated below.", + "type": "text" + } + ], + "index": 85 + } + ], + "index": 82, + "bbox_fs": [ + 304, + 490, + 528, + 583 + ] + }, + { + "type": "title", + "bbox": [ + 305, + 594, + 445, + 607 + ], + "lines": [ + { + "bbox": [ + 303, + 591, + 446, + 610 + ], + "spans": [ + { + "bbox": [ + 303, + 591, + 446, + 610 + ], + "score": 1.0, + "content": "3.2 Subject-Replaced Query", + "type": "text" + } + ], + "index": 86 + } + ], + "index": 86 + }, + { + "type": "text", + "bbox": [ + 304, + 612, + 525, + 773 + ], + "lines": [ + { + "bbox": [ + 304, + 612, + 527, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 527, + 625 + ], + "score": 1.0, + "content": "As our goal is to let LLMs analyze human personal-", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 303, + 625, + 527, + 639 + ], + "spans": [ + { + "bbox": [ + 303, + 625, + 527, + 639 + ], + "score": 1.0, + "content": "ities instead of querying itself (i.e., self-reporting),", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 304, + 640, + 526, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 526, + 653 + ], + "score": 1.0, + "content": "we propose the subject-replaced query (SRQ) by", + "type": "text" + } + ], + "index": 89 + }, + { + "bbox": [ + 303, + 653, + 525, + 666 + ], + "spans": [ + { + "bbox": [ + 303, + 653, + 525, + 666 + ], + "score": 1.0, + "content": "converting the original subject (i.e., “You”) of each", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 304, + 667, + 527, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 667, + 527, + 679 + ], + "score": 1.0, + "content": "question into a specific subject-of-interest. For ex-", + "type": "text" + } + ], + "index": 91 + }, + { + "bbox": [ + 304, + 680, + 527, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 680, + 527, + 693 + ], + "score": 1.0, + "content": "ample, when we hope to let LLMs assess the gen-", + "type": "text" + } + ], + "index": 92 + }, + { + "bbox": [ + 304, + 694, + 526, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 694, + 526, + 707 + ], + "score": 1.0, + "content": "eral personality of men, we can replace the subject", + "type": "text" + } + ], + "index": 93 + }, + { + "bbox": [ + 303, + 705, + 526, + 720 + ], + "spans": [ + { + "bbox": [ + 303, + 705, + 526, + 720 + ], + "score": 1.0, + "content": "“You” with “Men”, and correspondingly change the", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 304, + 721, + 527, + 734 + ], + "spans": [ + { + "bbox": [ + 304, + 721, + 527, + 734 + ], + "score": 1.0, + "content": "pronoun “your” to “their” (see the example below).", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 303, + 733, + 526, + 748 + ], + "spans": [ + { + "bbox": [ + 303, + 733, + 526, + 748 + ], + "score": 1.0, + "content": "Original Statement: You spend a lot of your free", + "type": "text" + } + ], + "index": 96 + }, + { + "bbox": [ + 304, + 747, + 525, + 761 + ], + "spans": [ + { + "bbox": [ + 304, + 747, + 525, + 761 + ], + "score": 1.0, + "content": "time exploring various random topics that pique", + "type": "text" + } + ], + "index": 97 + }, + { + "bbox": [ + 303, + 762, + 366, + 774 + ], + "spans": [ + { + "bbox": [ + 303, + 762, + 366, + 774 + ], + "score": 1.0, + "content": "your interest.", + "type": "text" + } + ], + "index": 98 + } + ], + "index": 92.5, + "bbox_fs": [ + 303, + 612, + 527, + 774 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 69, + 71, + 290, + 111 + ], + "lines": [ + { + "bbox": [ + 68, + 70, + 290, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 290, + 84 + ], + "score": 1.0, + "content": "SRQ Statement: Men spend a lot of their free", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 85, + 290, + 99 + ], + "spans": [ + { + "bbox": [ + 69, + 85, + 290, + 99 + ], + "score": 1.0, + "content": "time exploring various random topics that pique", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 98, + 136, + 112 + ], + "spans": [ + { + "bbox": [ + 69, + 98, + 136, + 112 + ], + "score": 1.0, + "content": "their interests.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 69, + 114, + 290, + 235 + ], + "lines": [ + { + "bbox": [ + 79, + 113, + 292, + 127 + ], + "spans": [ + { + "bbox": [ + 79, + 113, + 292, + 127 + ], + "score": 1.0, + "content": "In this way, we can request the LLMs to ana-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 127, + 290, + 141 + ], + "spans": [ + { + "bbox": [ + 69, + 127, + 290, + 141 + ], + "score": 1.0, + "content": "lyze and infer the choices/answers of a specific", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 142, + 290, + 155 + ], + "spans": [ + { + "bbox": [ + 69, + 142, + 290, + 155 + ], + "score": 1.0, + "content": "subject, so as to query LLMs about the personality", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 154, + 291, + 169 + ], + "spans": [ + { + "bbox": [ + 68, + 154, + 291, + 169 + ], + "score": 1.0, + "content": "of such subject based on a certain personality mea-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 168, + 290, + 182 + ], + "spans": [ + { + "bbox": [ + 69, + 168, + 290, + 182 + ], + "score": 1.0, + "content": "sure (e.g., MBTI). The proposed SRQ is general", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 182, + 290, + 195 + ], + "spans": [ + { + "bbox": [ + 69, + 182, + 290, + 195 + ], + "score": 1.0, + "content": "and scalable. By simply replacing the subject in the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 195, + 292, + 209 + ], + "spans": [ + { + "bbox": [ + 69, + 195, + 292, + 209 + ], + "score": 1.0, + "content": "test (see Fig. 1), we can convert the original self-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 210, + 291, + 222 + ], + "spans": [ + { + "bbox": [ + 68, + 210, + 291, + 222 + ], + "score": 1.0, + "content": "report questionnaire into an analysis of expected", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 222, + 217, + 236 + ], + "spans": [ + { + "bbox": [ + 69, + 222, + 217, + 236 + ], + "score": 1.0, + "content": "subjects from the point of LLMs.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 69, + 237, + 290, + 453 + ], + "lines": [ + { + "bbox": [ + 79, + 237, + 290, + 251 + ], + "spans": [ + { + "bbox": [ + 79, + 237, + 290, + 251 + ], + "score": 1.0, + "content": "In our work, we choose large groups of people", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 68, + 250, + 291, + 265 + ], + "spans": [ + { + "bbox": [ + 68, + 250, + 291, + 265 + ], + "score": 1.0, + "content": "(e.g., “Men”, “Barbers”) instead of certain persons", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 265, + 290, + 277 + ], + "spans": [ + { + "bbox": [ + 69, + 265, + 290, + 277 + ], + "score": 1.0, + "content": "as the assessed subjects. First, as our framework", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 279, + 290, + 291 + ], + "spans": [ + { + "bbox": [ + 69, + 279, + 290, + 291 + ], + "score": 1.0, + "content": "only uses the subject name without extra personal", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 292, + 291, + 305 + ], + "spans": [ + { + "bbox": [ + 69, + 292, + 291, + 305 + ], + "score": 1.0, + "content": "information to construct MBTI queries, it is un-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 306, + 290, + 317 + ], + "spans": [ + { + "bbox": [ + 69, + 306, + 290, + 317 + ], + "score": 1.0, + "content": "realistic to let LLMs assess the MBTI answers or", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 68, + 319, + 290, + 331 + ], + "spans": [ + { + "bbox": [ + 68, + 319, + 290, + 331 + ], + "score": 1.0, + "content": "personality of a certain person who is out of their", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 69, + 333, + 290, + 345 + ], + "spans": [ + { + "bbox": [ + 69, + 333, + 290, + 345 + ], + "score": 1.0, + "content": "learned knowledge. Second, the selected subjects", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 68, + 347, + 290, + 358 + ], + "spans": [ + { + "bbox": [ + 68, + 347, + 290, + 358 + ], + "score": 1.0, + "content": "are common in the knowledge base of LLMs and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 69, + 360, + 290, + 372 + ], + "spans": [ + { + "bbox": [ + 69, + 360, + 290, + 372 + ], + "score": 1.0, + "content": "can test the basic personality assessment ability", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 69, + 373, + 291, + 385 + ], + "spans": [ + { + "bbox": [ + 69, + 373, + 291, + 385 + ], + "score": 1.0, + "content": "of LLMs, which is the main focus of our work.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 69, + 387, + 290, + 399 + ], + "spans": [ + { + "bbox": [ + 69, + 387, + 290, + 399 + ], + "score": 1.0, + "content": "Moreover, subjects with different professions such", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 399, + 290, + 413 + ], + "spans": [ + { + "bbox": [ + 68, + 399, + 290, + 413 + ], + "score": 1.0, + "content": "as “Barbers” are frequently used to measure the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 69, + 414, + 290, + 426 + ], + "spans": [ + { + "bbox": [ + 69, + 414, + 290, + 426 + ], + "score": 1.0, + "content": "bias in LLMs (Nadeem et al., 2021), thus we select", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 69, + 428, + 290, + 440 + ], + "spans": [ + { + "bbox": [ + 69, + 428, + 290, + 440 + ], + "score": 1.0, + "content": "such representative professions to better evaluate", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 70, + 441, + 291, + 453 + ], + "spans": [ + { + "bbox": [ + 70, + 441, + 291, + 453 + ], + "score": 1.0, + "content": "the consistency, robustness, and fairness of LLMs.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 70, + 468, + 257, + 480 + ], + "lines": [ + { + "bbox": [ + 67, + 466, + 258, + 482 + ], + "spans": [ + { + "bbox": [ + 67, + 466, + 258, + 482 + ], + "score": 1.0, + "content": "3.3 Correctness-Evaluated Instruction", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 69, + 488, + 290, + 745 + ], + "lines": [ + { + "bbox": [ + 69, + 488, + 292, + 502 + ], + "spans": [ + { + "bbox": [ + 69, + 488, + 292, + 502 + ], + "score": 1.0, + "content": "Directly querying LLMs about human personali-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 69, + 502, + 291, + 515 + ], + "spans": [ + { + "bbox": [ + 69, + 502, + 291, + 515 + ], + "score": 1.0, + "content": "ties with the original instruction can be intractable,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 68, + 516, + 290, + 528 + ], + "spans": [ + { + "bbox": [ + 68, + 516, + 290, + 528 + ], + "score": 1.0, + "content": "as LLMs such as ChatGPT are trained to NOT", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 530, + 290, + 541 + ], + "spans": [ + { + "bbox": [ + 68, + 530, + 290, + 541 + ], + "score": 1.0, + "content": "possess personal emotions or beliefs. As shown", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 69, + 543, + 291, + 556 + ], + "spans": [ + { + "bbox": [ + 69, + 543, + 291, + 556 + ], + "score": 1.0, + "content": "in Fig. 2, they can only generate a neutral opin-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 69, + 556, + 290, + 569 + ], + "spans": [ + { + "bbox": [ + 69, + 556, + 290, + 569 + ], + "score": 1.0, + "content": "ion when we query their agreement or disagree-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 68, + 571, + 290, + 582 + ], + "spans": [ + { + "bbox": [ + 68, + 571, + 290, + 582 + ], + "score": 1.0, + "content": "ment, regardless of different subjects. To solve", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 68, + 583, + 292, + 598 + ], + "spans": [ + { + "bbox": [ + 68, + 583, + 292, + 598 + ], + "score": 1.0, + "content": "this challenge, we propose to convert the origi-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 68, + 596, + 291, + 611 + ], + "spans": [ + { + "bbox": [ + 68, + 596, + 291, + 611 + ], + "score": 1.0, + "content": "nal agreement-measured instruction (i.e., querying", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 69, + 611, + 291, + 624 + ], + "spans": [ + { + "bbox": [ + 69, + 611, + 291, + 624 + ], + "score": 1.0, + "content": "degree of agreement) into correctness-evaluated", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 69, + 625, + 291, + 637 + ], + "spans": [ + { + "bbox": [ + 69, + 625, + 291, + 637 + ], + "score": 1.0, + "content": "instruction (CEI) by letting LLMs evaluate the cor-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 68, + 637, + 291, + 651 + ], + "spans": [ + { + "bbox": [ + 68, + 637, + 291, + 651 + ], + "score": 1.0, + "content": "rectness of the statement in questions. Specifically,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 68, + 651, + 290, + 664 + ], + "spans": [ + { + "bbox": [ + 68, + 652, + 208, + 664 + ], + "score": 1.0, + "content": "we convert the original options", + "type": "text" + }, + { + "bbox": [ + 208, + 651, + 240, + 664 + ], + "score": 0.29, + "content": "\\{ A g r e e", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 652, + 290, + 664 + ], + "score": 1.0, + "content": ", Generally", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 69, + 665, + 291, + 678 + ], + "spans": [ + { + "bbox": [ + 69, + 665, + 291, + 678 + ], + "score": 1.0, + "content": "agree, Partially agree, Neither agree nor disagree,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 68, + 678, + 290, + 692 + ], + "spans": [ + { + "bbox": [ + 68, + 678, + 290, + 692 + ], + "score": 1.0, + "content": "Partially disagree, Generally disagree, Disagree}", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 68, + 691, + 291, + 705 + ], + "spans": [ + { + "bbox": [ + 68, + 691, + 291, + 705 + ], + "score": 1.0, + "content": "into {Correct, Generally correct, Partially correct,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 68, + 705, + 292, + 718 + ], + "spans": [ + { + "bbox": [ + 68, + 705, + 292, + 718 + ], + "score": 1.0, + "content": "Neither correct nor wrong, Partially wrong, Gener-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 68, + 718, + 291, + 733 + ], + "spans": [ + { + "bbox": [ + 68, + 718, + 291, + 733 + ], + "score": 1.0, + "content": "ally wrong, Wrong}, and then construct an unbiased", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 68, + 732, + 289, + 747 + ], + "spans": [ + { + "bbox": [ + 68, + 732, + 289, + 747 + ], + "score": 1.0, + "content": "prompt (see Sec. 3.1) based on the proposed CEI.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 69, + 747, + 290, + 774 + ], + "lines": [ + { + "bbox": [ + 80, + 747, + 290, + 760 + ], + "spans": [ + { + "bbox": [ + 80, + 747, + 290, + 760 + ], + "score": 1.0, + "content": "As shown in Fig. 2, using CEI enables ChatGPT", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "spans": [ + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "score": 1.0, + "content": "to provide a clearer response to the question instead", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5 + }, + { + "type": "image", + "bbox": [ + 333, + 70, + 496, + 284 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 333, + 70, + 496, + 284 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 333, + 70, + 496, + 284 + ], + "spans": [ + { + "bbox": [ + 333, + 70, + 496, + 284 + ], + "score": 0.87, + "type": "image", + "image_path": "979a135d67d9ff852a13a7117066c31de30d6ee8b00a70ee3bd2a84454c865f9.jpg" + } + ] + } + ], + "index": 57.5, + "virtual_lines": [ + { + "bbox": [ + 333, + 70, + 496, + 83.375 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 333, + 83.375, + 496, + 96.75 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 333, + 96.75, + 496, + 110.125 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 333, + 110.125, + 496, + 123.5 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 333, + 123.5, + 496, + 136.875 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 333, + 136.875, + 496, + 150.25 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 333, + 150.25, + 496, + 163.625 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 333, + 163.625, + 496, + 177.0 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 333, + 177.0, + 496, + 190.375 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 333, + 190.375, + 496, + 203.75 + ], + "spans": [], + "index": 59 + }, + { + "bbox": [ + 333, + 203.75, + 496, + 217.125 + ], + "spans": [], + "index": 60 + }, + { + "bbox": [ + 333, + 217.125, + 496, + 230.5 + ], + "spans": [], + "index": 61 + }, + { + "bbox": [ + 333, + 230.5, + 496, + 243.875 + ], + "spans": [], + "index": 62 + }, + { + "bbox": [ + 333, + 243.875, + 496, + 257.25 + ], + "spans": [], + "index": 63 + }, + { + "bbox": [ + 333, + 257.25, + 496, + 270.625 + ], + "spans": [], + "index": 64 + }, + { + "bbox": [ + 333, + 270.625, + 496, + 284.0 + ], + "spans": [], + "index": 65 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 293, + 525, + 340 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 527, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 527, + 306 + ], + "score": 1.0, + "content": "Figure 2: Comparison of answers generated by Chat-", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 305, + 525, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 525, + 317 + ], + "score": 1.0, + "content": "GPT when adopting different types of instructions. Note", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 317, + 525, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 525, + 329 + ], + "score": 1.0, + "content": "that the agreement-measured instruction always leads", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 329, + 429, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 429, + 341 + ], + "score": 1.0, + "content": "to a neutral answer in practice.", + "type": "text" + } + ], + "index": 69 + } + ], + "index": 67.5 + } + ], + "index": 62.5 + }, + { + "type": "text", + "bbox": [ + 305, + 364, + 525, + 418 + ], + "lines": [ + { + "bbox": [ + 303, + 364, + 526, + 377 + ], + "spans": [ + { + "bbox": [ + 303, + 364, + 526, + 377 + ], + "score": 1.0, + "content": "of giving a neutral response. Note that the CEI is", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 303, + 378, + 525, + 391 + ], + "spans": [ + { + "bbox": [ + 303, + 378, + 525, + 391 + ], + "score": 1.0, + "content": "essentially equivalent to the agreement-measured", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 392, + 526, + 404 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 526, + 404 + ], + "score": 1.0, + "content": "instruction and can be flexibly extended with other", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 304, + 405, + 497, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 497, + 418 + ], + "score": 1.0, + "content": "forms (e.g., replacing “correct” by “right”).", + "type": "text" + } + ], + "index": 73 + } + ], + "index": 71.5 + }, + { + "type": "title", + "bbox": [ + 305, + 430, + 439, + 443 + ], + "lines": [ + { + "bbox": [ + 303, + 429, + 440, + 444 + ], + "spans": [ + { + "bbox": [ + 303, + 429, + 440, + 444 + ], + "score": 1.0, + "content": "3.4 The Entire Framework", + "type": "text" + } + ], + "index": 74 + } + ], + "index": 74 + }, + { + "type": "text", + "bbox": [ + 305, + 449, + 525, + 543 + ], + "lines": [ + { + "bbox": [ + 304, + 448, + 527, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 527, + 463 + ], + "score": 1.0, + "content": "The overview of our framework is shown in Fig. 1.", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 305, + 463, + 524, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 463, + 433, + 475 + ], + "score": 1.0, + "content": "Given the original statement", + "type": "text" + }, + { + "bbox": [ + 433, + 463, + 444, + 475 + ], + "score": 0.88, + "content": "S _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 463, + 515, + 475 + ], + "score": 1.0, + "content": "and instruction", + "type": "text" + }, + { + "bbox": [ + 515, + 463, + 524, + 475 + ], + "score": 0.85, + "content": "I _ { i }", + "type": "inline_equation" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 475, + 527, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 332, + 491 + ], + "score": 1.0, + "content": "of the", + "type": "text" + }, + { + "bbox": [ + 332, + 475, + 346, + 488 + ], + "score": 0.87, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 475, + 527, + 491 + ], + "score": 1.0, + "content": "question, we construct the new statement", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 305, + 488, + 527, + 505 + ], + "spans": [ + { + "bbox": [ + 305, + 490, + 316, + 503 + ], + "score": 0.89, + "content": "S _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 488, + 527, + 505 + ], + "score": 1.0, + "content": "based on SRQ (Sec. 3.2) and the new instruction", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 503, + 526, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 315, + 517 + ], + "score": 0.88, + "content": "I _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 503, + 526, + 517 + ], + "score": 1.0, + "content": "based on CEI (Sec. 3.3), which are combined to", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 518, + 525, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 443, + 531 + ], + "score": 1.0, + "content": "construct the unbiased prompt", + "type": "text" + }, + { + "bbox": [ + 443, + 518, + 455, + 529 + ], + "score": 0.86, + "content": "P _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 518, + 525, + 531 + ], + "score": 1.0, + "content": "(Sec. 3.1). We", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 303, + 531, + 493, + 545 + ], + "spans": [ + { + "bbox": [ + 303, + 531, + 465, + 545 + ], + "score": 1.0, + "content": "query the LLM to obtain the answer", + "type": "text" + }, + { + "bbox": [ + 465, + 531, + 478, + 543 + ], + "score": 0.89, + "content": "A _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 531, + 493, + 545 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 81 + } + ], + "index": 78 + }, + { + "type": "interline_equation", + "bbox": [ + 379, + 556, + 448, + 572 + ], + "lines": [ + { + "bbox": [ + 379, + 556, + 448, + 572 + ], + "spans": [ + { + "bbox": [ + 379, + 556, + 448, + 572 + ], + "score": 0.88, + "content": "A _ { i } \\sim { \\mathcal { M } } _ { \\tau } ( P _ { i } ) ,", + "type": "interline_equation", + "image_path": "bb54f955a13519ed64ec8410487f0954ffba661306ed264ecbb69d154603c621.jpg" + } + ] + } + ], + "index": 82, + "virtual_lines": [ + { + "bbox": [ + 379, + 556, + 448, + 572 + ], + "spans": [], + "index": 82 + } + ] + }, + { + "type": "text", + "bbox": [ + 304, + 585, + 525, + 774 + ], + "lines": [ + { + "bbox": [ + 304, + 584, + 527, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 334, + 598 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 334, + 585, + 353, + 597 + ], + "score": 0.91, + "content": "\\mathcal { M } _ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 584, + 527, + 598 + ], + "score": 1.0, + "content": "denotes the LLM trained with the tem-", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 304, + 598, + 528, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 345, + 611 + ], + "score": 1.0, + "content": "perature", + "type": "text" + }, + { + "bbox": [ + 345, + 600, + 353, + 609 + ], + "score": 0.6, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 599, + 358, + 611 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 359, + 598, + 397, + 612 + ], + "score": 0.92, + "content": "\\mathcal { M } _ { \\tau } ( P _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 599, + 528, + 611 + ], + "score": 1.0, + "content": "represents the answer sam-", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 304, + 612, + 526, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 526, + 625 + ], + "score": 1.0, + "content": "pling distribution of LLM conditioned on the input", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 304, + 626, + 526, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 340, + 639 + ], + "score": 1.0, + "content": "prompt", + "type": "text" + }, + { + "bbox": [ + 340, + 626, + 352, + 638 + ], + "score": 0.76, + "content": "P _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 626, + 357, + 639 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 357, + 626, + 371, + 638 + ], + "score": 0.79, + "content": "A _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 626, + 526, + 639 + ], + "score": 1.0, + "content": "represents the most likely answer", + "type": "text" + } + ], + "index": 86 + }, + { + "bbox": [ + 303, + 639, + 525, + 653 + ], + "spans": [ + { + "bbox": [ + 303, + 639, + 376, + 653 + ], + "score": 1.0, + "content": "generated from", + "type": "text" + }, + { + "bbox": [ + 376, + 639, + 414, + 652 + ], + "score": 0.66, + "content": "\\mathcal { M } _ { \\tau } ( P _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 639, + 419, + 653 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 419, + 639, + 497, + 652 + ], + "score": 0.79, + "content": "i \\in \\{ 1 , 2 , \\cdots , n \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 639, + 525, + 653 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 304, + 653, + 525, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 452, + 665 + ], + "score": 1.0, + "content": "index of different questions, and", + "type": "text" + }, + { + "bbox": [ + 453, + 655, + 461, + 663 + ], + "score": 0.78, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 653, + 525, + 665 + ], + "score": 1.0, + "content": "is the number", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 304, + 666, + 526, + 680 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 526, + 680 + ], + "score": 1.0, + "content": "of all questions in MBTI. We adopt the default", + "type": "text" + } + ], + "index": 89 + }, + { + "bbox": [ + 303, + 680, + 527, + 693 + ], + "spans": [ + { + "bbox": [ + 303, + 680, + 527, + 693 + ], + "score": 1.0, + "content": "temperature used in training standard GPT models.", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 304, + 693, + 526, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 693, + 526, + 707 + ], + "score": 1.0, + "content": "The generated answer is further parsed with several", + "type": "text" + } + ], + "index": 91 + }, + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "score": 1.0, + "content": "simple rules, which ensures that it contains or can", + "type": "text" + } + ], + "index": 92 + }, + { + "bbox": [ + 303, + 719, + 527, + 734 + ], + "spans": [ + { + "bbox": [ + 303, + 719, + 527, + 734 + ], + "score": 1.0, + "content": "be transformed to an exact option. For instance,", + "type": "text" + } + ], + "index": 93 + }, + { + "bbox": [ + 304, + 735, + 526, + 747 + ], + "spans": [ + { + "bbox": [ + 304, + 735, + 526, + 747 + ], + "score": 1.0, + "content": "when we obtain the explicit option “generally in-", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 304, + 748, + 526, + 761 + ], + "spans": [ + { + "bbox": [ + 304, + 748, + 526, + 761 + ], + "score": 1.0, + "content": "correct”, the parsing rules can convert this answer", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 304, + 761, + 527, + 775 + ], + "spans": [ + { + "bbox": [ + 304, + 761, + 527, + 775 + ], + "score": 1.0, + "content": "to “generally wrong” to match the existing options.", + "type": "text" + } + ], + "index": 96 + } + ], + "index": 89.5 + } + ], + "page_idx": 3, + "page_size": [ + 595, + 841 + ], + "discarded_blocks": [], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 69, + 71, + 290, + 111 + ], + "lines": [ + { + "bbox": [ + 68, + 70, + 290, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 290, + 84 + ], + "score": 1.0, + "content": "SRQ Statement: Men spend a lot of their free", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 85, + 290, + 99 + ], + "spans": [ + { + "bbox": [ + 69, + 85, + 290, + 99 + ], + "score": 1.0, + "content": "time exploring various random topics that pique", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 98, + 136, + 112 + ], + "spans": [ + { + "bbox": [ + 69, + 98, + 136, + 112 + ], + "score": 1.0, + "content": "their interests.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 68, + 70, + 290, + 112 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 114, + 290, + 235 + ], + "lines": [ + { + "bbox": [ + 79, + 113, + 292, + 127 + ], + "spans": [ + { + "bbox": [ + 79, + 113, + 292, + 127 + ], + "score": 1.0, + "content": "In this way, we can request the LLMs to ana-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 127, + 290, + 141 + ], + "spans": [ + { + "bbox": [ + 69, + 127, + 290, + 141 + ], + "score": 1.0, + "content": "lyze and infer the choices/answers of a specific", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 142, + 290, + 155 + ], + "spans": [ + { + "bbox": [ + 69, + 142, + 290, + 155 + ], + "score": 1.0, + "content": "subject, so as to query LLMs about the personality", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 154, + 291, + 169 + ], + "spans": [ + { + "bbox": [ + 68, + 154, + 291, + 169 + ], + "score": 1.0, + "content": "of such subject based on a certain personality mea-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 168, + 290, + 182 + ], + "spans": [ + { + "bbox": [ + 69, + 168, + 290, + 182 + ], + "score": 1.0, + "content": "sure (e.g., MBTI). The proposed SRQ is general", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 182, + 290, + 195 + ], + "spans": [ + { + "bbox": [ + 69, + 182, + 290, + 195 + ], + "score": 1.0, + "content": "and scalable. By simply replacing the subject in the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 195, + 292, + 209 + ], + "spans": [ + { + "bbox": [ + 69, + 195, + 292, + 209 + ], + "score": 1.0, + "content": "test (see Fig. 1), we can convert the original self-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 210, + 291, + 222 + ], + "spans": [ + { + "bbox": [ + 68, + 210, + 291, + 222 + ], + "score": 1.0, + "content": "report questionnaire into an analysis of expected", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 222, + 217, + 236 + ], + "spans": [ + { + "bbox": [ + 69, + 222, + 217, + 236 + ], + "score": 1.0, + "content": "subjects from the point of LLMs.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7, + "bbox_fs": [ + 68, + 113, + 292, + 236 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 237, + 290, + 453 + ], + "lines": [ + { + "bbox": [ + 79, + 237, + 290, + 251 + ], + "spans": [ + { + "bbox": [ + 79, + 237, + 290, + 251 + ], + "score": 1.0, + "content": "In our work, we choose large groups of people", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 68, + 250, + 291, + 265 + ], + "spans": [ + { + "bbox": [ + 68, + 250, + 291, + 265 + ], + "score": 1.0, + "content": "(e.g., “Men”, “Barbers”) instead of certain persons", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 265, + 290, + 277 + ], + "spans": [ + { + "bbox": [ + 69, + 265, + 290, + 277 + ], + "score": 1.0, + "content": "as the assessed subjects. First, as our framework", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 279, + 290, + 291 + ], + "spans": [ + { + "bbox": [ + 69, + 279, + 290, + 291 + ], + "score": 1.0, + "content": "only uses the subject name without extra personal", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 292, + 291, + 305 + ], + "spans": [ + { + "bbox": [ + 69, + 292, + 291, + 305 + ], + "score": 1.0, + "content": "information to construct MBTI queries, it is un-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 306, + 290, + 317 + ], + "spans": [ + { + "bbox": [ + 69, + 306, + 290, + 317 + ], + "score": 1.0, + "content": "realistic to let LLMs assess the MBTI answers or", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 68, + 319, + 290, + 331 + ], + "spans": [ + { + "bbox": [ + 68, + 319, + 290, + 331 + ], + "score": 1.0, + "content": "personality of a certain person who is out of their", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 69, + 333, + 290, + 345 + ], + "spans": [ + { + "bbox": [ + 69, + 333, + 290, + 345 + ], + "score": 1.0, + "content": "learned knowledge. Second, the selected subjects", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 68, + 347, + 290, + 358 + ], + "spans": [ + { + "bbox": [ + 68, + 347, + 290, + 358 + ], + "score": 1.0, + "content": "are common in the knowledge base of LLMs and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 69, + 360, + 290, + 372 + ], + "spans": [ + { + "bbox": [ + 69, + 360, + 290, + 372 + ], + "score": 1.0, + "content": "can test the basic personality assessment ability", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 69, + 373, + 291, + 385 + ], + "spans": [ + { + "bbox": [ + 69, + 373, + 291, + 385 + ], + "score": 1.0, + "content": "of LLMs, which is the main focus of our work.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 69, + 387, + 290, + 399 + ], + "spans": [ + { + "bbox": [ + 69, + 387, + 290, + 399 + ], + "score": 1.0, + "content": "Moreover, subjects with different professions such", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 399, + 290, + 413 + ], + "spans": [ + { + "bbox": [ + 68, + 399, + 290, + 413 + ], + "score": 1.0, + "content": "as “Barbers” are frequently used to measure the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 69, + 414, + 290, + 426 + ], + "spans": [ + { + "bbox": [ + 69, + 414, + 290, + 426 + ], + "score": 1.0, + "content": "bias in LLMs (Nadeem et al., 2021), thus we select", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 69, + 428, + 290, + 440 + ], + "spans": [ + { + "bbox": [ + 69, + 428, + 290, + 440 + ], + "score": 1.0, + "content": "such representative professions to better evaluate", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 70, + 441, + 291, + 453 + ], + "spans": [ + { + "bbox": [ + 70, + 441, + 291, + 453 + ], + "score": 1.0, + "content": "the consistency, robustness, and fairness of LLMs.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 19.5, + "bbox_fs": [ + 68, + 237, + 291, + 453 + ] + }, + { + "type": "title", + "bbox": [ + 70, + 468, + 257, + 480 + ], + "lines": [ + { + "bbox": [ + 67, + 466, + 258, + 482 + ], + "spans": [ + { + "bbox": [ + 67, + 466, + 258, + 482 + ], + "score": 1.0, + "content": "3.3 Correctness-Evaluated Instruction", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 69, + 488, + 290, + 745 + ], + "lines": [ + { + "bbox": [ + 69, + 488, + 292, + 502 + ], + "spans": [ + { + "bbox": [ + 69, + 488, + 292, + 502 + ], + "score": 1.0, + "content": "Directly querying LLMs about human personali-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 69, + 502, + 291, + 515 + ], + "spans": [ + { + "bbox": [ + 69, + 502, + 291, + 515 + ], + "score": 1.0, + "content": "ties with the original instruction can be intractable,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 68, + 516, + 290, + 528 + ], + "spans": [ + { + "bbox": [ + 68, + 516, + 290, + 528 + ], + "score": 1.0, + "content": "as LLMs such as ChatGPT are trained to NOT", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 530, + 290, + 541 + ], + "spans": [ + { + "bbox": [ + 68, + 530, + 290, + 541 + ], + "score": 1.0, + "content": "possess personal emotions or beliefs. As shown", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 69, + 543, + 291, + 556 + ], + "spans": [ + { + "bbox": [ + 69, + 543, + 291, + 556 + ], + "score": 1.0, + "content": "in Fig. 2, they can only generate a neutral opin-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 69, + 556, + 290, + 569 + ], + "spans": [ + { + "bbox": [ + 69, + 556, + 290, + 569 + ], + "score": 1.0, + "content": "ion when we query their agreement or disagree-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 68, + 571, + 290, + 582 + ], + "spans": [ + { + "bbox": [ + 68, + 571, + 290, + 582 + ], + "score": 1.0, + "content": "ment, regardless of different subjects. To solve", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 68, + 583, + 292, + 598 + ], + "spans": [ + { + "bbox": [ + 68, + 583, + 292, + 598 + ], + "score": 1.0, + "content": "this challenge, we propose to convert the origi-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 68, + 596, + 291, + 611 + ], + "spans": [ + { + "bbox": [ + 68, + 596, + 291, + 611 + ], + "score": 1.0, + "content": "nal agreement-measured instruction (i.e., querying", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 69, + 611, + 291, + 624 + ], + "spans": [ + { + "bbox": [ + 69, + 611, + 291, + 624 + ], + "score": 1.0, + "content": "degree of agreement) into correctness-evaluated", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 69, + 625, + 291, + 637 + ], + "spans": [ + { + "bbox": [ + 69, + 625, + 291, + 637 + ], + "score": 1.0, + "content": "instruction (CEI) by letting LLMs evaluate the cor-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 68, + 637, + 291, + 651 + ], + "spans": [ + { + "bbox": [ + 68, + 637, + 291, + 651 + ], + "score": 1.0, + "content": "rectness of the statement in questions. Specifically,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 68, + 651, + 290, + 664 + ], + "spans": [ + { + "bbox": [ + 68, + 652, + 208, + 664 + ], + "score": 1.0, + "content": "we convert the original options", + "type": "text" + }, + { + "bbox": [ + 208, + 651, + 240, + 664 + ], + "score": 0.29, + "content": "\\{ A g r e e", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 652, + 290, + 664 + ], + "score": 1.0, + "content": ", Generally", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 69, + 665, + 291, + 678 + ], + "spans": [ + { + "bbox": [ + 69, + 665, + 291, + 678 + ], + "score": 1.0, + "content": "agree, Partially agree, Neither agree nor disagree,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 68, + 678, + 290, + 692 + ], + "spans": [ + { + "bbox": [ + 68, + 678, + 290, + 692 + ], + "score": 1.0, + "content": "Partially disagree, Generally disagree, Disagree}", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 68, + 691, + 291, + 705 + ], + "spans": [ + { + "bbox": [ + 68, + 691, + 291, + 705 + ], + "score": 1.0, + "content": "into {Correct, Generally correct, Partially correct,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 68, + 705, + 292, + 718 + ], + "spans": [ + { + "bbox": [ + 68, + 705, + 292, + 718 + ], + "score": 1.0, + "content": "Neither correct nor wrong, Partially wrong, Gener-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 68, + 718, + 291, + 733 + ], + "spans": [ + { + "bbox": [ + 68, + 718, + 291, + 733 + ], + "score": 1.0, + "content": "ally wrong, Wrong}, and then construct an unbiased", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 68, + 732, + 289, + 747 + ], + "spans": [ + { + "bbox": [ + 68, + 732, + 289, + 747 + ], + "score": 1.0, + "content": "prompt (see Sec. 3.1) based on the proposed CEI.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 38, + "bbox_fs": [ + 68, + 488, + 292, + 747 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 747, + 290, + 774 + ], + "lines": [ + { + "bbox": [ + 80, + 747, + 290, + 760 + ], + "spans": [ + { + "bbox": [ + 80, + 747, + 290, + 760 + ], + "score": 1.0, + "content": "As shown in Fig. 2, using CEI enables ChatGPT", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "spans": [ + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "score": 1.0, + "content": "to provide a clearer response to the question instead", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 303, + 364, + 526, + 377 + ], + "spans": [ + { + "bbox": [ + 303, + 364, + 526, + 377 + ], + "score": 1.0, + "content": "of giving a neutral response. Note that the CEI is", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 303, + 378, + 525, + 391 + ], + "spans": [ + { + "bbox": [ + 303, + 378, + 525, + 391 + ], + "score": 1.0, + "content": "essentially equivalent to the agreement-measured", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 392, + 526, + 404 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 526, + 404 + ], + "score": 1.0, + "content": "instruction and can be flexibly extended with other", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 304, + 405, + 497, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 497, + 418 + ], + "score": 1.0, + "content": "forms (e.g., replacing “correct” by “right”).", + "type": "text" + } + ], + "index": 73 + } + ], + "index": 48.5, + "bbox_fs": [ + 69, + 747, + 290, + 774 + ] + }, + { + "type": "image", + "bbox": [ + 333, + 70, + 496, + 284 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 333, + 70, + 496, + 284 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 333, + 70, + 496, + 284 + ], + "spans": [ + { + "bbox": [ + 333, + 70, + 496, + 284 + ], + "score": 0.87, + "type": "image", + "image_path": "979a135d67d9ff852a13a7117066c31de30d6ee8b00a70ee3bd2a84454c865f9.jpg" + } + ] + } + ], + "index": 57.5, + "virtual_lines": [ + { + "bbox": [ + 333, + 70, + 496, + 83.375 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 333, + 83.375, + 496, + 96.75 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 333, + 96.75, + 496, + 110.125 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 333, + 110.125, + 496, + 123.5 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 333, + 123.5, + 496, + 136.875 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 333, + 136.875, + 496, + 150.25 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 333, + 150.25, + 496, + 163.625 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 333, + 163.625, + 496, + 177.0 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 333, + 177.0, + 496, + 190.375 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 333, + 190.375, + 496, + 203.75 + ], + "spans": [], + "index": 59 + }, + { + "bbox": [ + 333, + 203.75, + 496, + 217.125 + ], + "spans": [], + "index": 60 + }, + { + "bbox": [ + 333, + 217.125, + 496, + 230.5 + ], + "spans": [], + "index": 61 + }, + { + "bbox": [ + 333, + 230.5, + 496, + 243.875 + ], + "spans": [], + "index": 62 + }, + { + "bbox": [ + 333, + 243.875, + 496, + 257.25 + ], + "spans": [], + "index": 63 + }, + { + "bbox": [ + 333, + 257.25, + 496, + 270.625 + ], + "spans": [], + "index": 64 + }, + { + "bbox": [ + 333, + 270.625, + 496, + 284.0 + ], + "spans": [], + "index": 65 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 293, + 525, + 340 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 527, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 527, + 306 + ], + "score": 1.0, + "content": "Figure 2: Comparison of answers generated by Chat-", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 305, + 525, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 525, + 317 + ], + "score": 1.0, + "content": "GPT when adopting different types of instructions. Note", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 317, + 525, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 525, + 329 + ], + "score": 1.0, + "content": "that the agreement-measured instruction always leads", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 329, + 429, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 429, + 341 + ], + "score": 1.0, + "content": "to a neutral answer in practice.", + "type": "text" + } + ], + "index": 69 + } + ], + "index": 67.5 + } + ], + "index": 62.5 + }, + { + "type": "text", + "bbox": [ + 305, + 364, + 525, + 418 + ], + "lines": [], + "index": 71.5, + "bbox_fs": [ + 303, + 364, + 526, + 418 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 305, + 430, + 439, + 443 + ], + "lines": [ + { + "bbox": [ + 303, + 429, + 440, + 444 + ], + "spans": [ + { + "bbox": [ + 303, + 429, + 440, + 444 + ], + "score": 1.0, + "content": "3.4 The Entire Framework", + "type": "text" + } + ], + "index": 74 + } + ], + "index": 74 + }, + { + "type": "text", + "bbox": [ + 305, + 449, + 525, + 543 + ], + "lines": [ + { + "bbox": [ + 304, + 448, + 527, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 527, + 463 + ], + "score": 1.0, + "content": "The overview of our framework is shown in Fig. 1.", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 305, + 463, + 524, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 463, + 433, + 475 + ], + "score": 1.0, + "content": "Given the original statement", + "type": "text" + }, + { + "bbox": [ + 433, + 463, + 444, + 475 + ], + "score": 0.88, + "content": "S _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 463, + 515, + 475 + ], + "score": 1.0, + "content": "and instruction", + "type": "text" + }, + { + "bbox": [ + 515, + 463, + 524, + 475 + ], + "score": 0.85, + "content": "I _ { i }", + "type": "inline_equation" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 475, + 527, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 332, + 491 + ], + "score": 1.0, + "content": "of the", + "type": "text" + }, + { + "bbox": [ + 332, + 475, + 346, + 488 + ], + "score": 0.87, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 475, + 527, + 491 + ], + "score": 1.0, + "content": "question, we construct the new statement", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 305, + 488, + 527, + 505 + ], + "spans": [ + { + "bbox": [ + 305, + 490, + 316, + 503 + ], + "score": 0.89, + "content": "S _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 488, + 527, + 505 + ], + "score": 1.0, + "content": "based on SRQ (Sec. 3.2) and the new instruction", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 503, + 526, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 315, + 517 + ], + "score": 0.88, + "content": "I _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 503, + 526, + 517 + ], + "score": 1.0, + "content": "based on CEI (Sec. 3.3), which are combined to", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 518, + 525, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 443, + 531 + ], + "score": 1.0, + "content": "construct the unbiased prompt", + "type": "text" + }, + { + "bbox": [ + 443, + 518, + 455, + 529 + ], + "score": 0.86, + "content": "P _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 518, + 525, + 531 + ], + "score": 1.0, + "content": "(Sec. 3.1). We", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 303, + 531, + 493, + 545 + ], + "spans": [ + { + "bbox": [ + 303, + 531, + 465, + 545 + ], + "score": 1.0, + "content": "query the LLM to obtain the answer", + "type": "text" + }, + { + "bbox": [ + 465, + 531, + 478, + 543 + ], + "score": 0.89, + "content": "A _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 531, + 493, + 545 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 81 + } + ], + "index": 78, + "bbox_fs": [ + 303, + 448, + 527, + 545 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 379, + 556, + 448, + 572 + ], + "lines": [ + { + "bbox": [ + 379, + 556, + 448, + 572 + ], + "spans": [ + { + "bbox": [ + 379, + 556, + 448, + 572 + ], + "score": 0.88, + "content": "A _ { i } \\sim { \\mathcal { M } } _ { \\tau } ( P _ { i } ) ,", + "type": "interline_equation", + "image_path": "bb54f955a13519ed64ec8410487f0954ffba661306ed264ecbb69d154603c621.jpg" + } + ] + } + ], + "index": 82, + "virtual_lines": [ + { + "bbox": [ + 379, + 556, + 448, + 572 + ], + "spans": [], + "index": 82 + } + ] + }, + { + "type": "text", + "bbox": [ + 304, + 585, + 525, + 774 + ], + "lines": [ + { + "bbox": [ + 304, + 584, + 527, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 334, + 598 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 334, + 585, + 353, + 597 + ], + "score": 0.91, + "content": "\\mathcal { M } _ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 584, + 527, + 598 + ], + "score": 1.0, + "content": "denotes the LLM trained with the tem-", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 304, + 598, + 528, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 345, + 611 + ], + "score": 1.0, + "content": "perature", + "type": "text" + }, + { + "bbox": [ + 345, + 600, + 353, + 609 + ], + "score": 0.6, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 599, + 358, + 611 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 359, + 598, + 397, + 612 + ], + "score": 0.92, + "content": "\\mathcal { M } _ { \\tau } ( P _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 599, + 528, + 611 + ], + "score": 1.0, + "content": "represents the answer sam-", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 304, + 612, + 526, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 526, + 625 + ], + "score": 1.0, + "content": "pling distribution of LLM conditioned on the input", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 304, + 626, + 526, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 340, + 639 + ], + "score": 1.0, + "content": "prompt", + "type": "text" + }, + { + "bbox": [ + 340, + 626, + 352, + 638 + ], + "score": 0.76, + "content": "P _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 626, + 357, + 639 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 357, + 626, + 371, + 638 + ], + "score": 0.79, + "content": "A _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 626, + 526, + 639 + ], + "score": 1.0, + "content": "represents the most likely answer", + "type": "text" + } + ], + "index": 86 + }, + { + "bbox": [ + 303, + 639, + 525, + 653 + ], + "spans": [ + { + "bbox": [ + 303, + 639, + 376, + 653 + ], + "score": 1.0, + "content": "generated from", + "type": "text" + }, + { + "bbox": [ + 376, + 639, + 414, + 652 + ], + "score": 0.66, + "content": "\\mathcal { M } _ { \\tau } ( P _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 639, + 419, + 653 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 419, + 639, + 497, + 652 + ], + "score": 0.79, + "content": "i \\in \\{ 1 , 2 , \\cdots , n \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 639, + 525, + 653 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 304, + 653, + 525, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 452, + 665 + ], + "score": 1.0, + "content": "index of different questions, and", + "type": "text" + }, + { + "bbox": [ + 453, + 655, + 461, + 663 + ], + "score": 0.78, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 653, + 525, + 665 + ], + "score": 1.0, + "content": "is the number", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 304, + 666, + 526, + 680 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 526, + 680 + ], + "score": 1.0, + "content": "of all questions in MBTI. We adopt the default", + "type": "text" + } + ], + "index": 89 + }, + { + "bbox": [ + 303, + 680, + 527, + 693 + ], + "spans": [ + { + "bbox": [ + 303, + 680, + 527, + 693 + ], + "score": 1.0, + "content": "temperature used in training standard GPT models.", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 304, + 693, + 526, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 693, + 526, + 707 + ], + "score": 1.0, + "content": "The generated answer is further parsed with several", + "type": "text" + } + ], + "index": 91 + }, + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "score": 1.0, + "content": "simple rules, which ensures that it contains or can", + "type": "text" + } + ], + "index": 92 + }, + { + "bbox": [ + 303, + 719, + 527, + 734 + ], + "spans": [ + { + "bbox": [ + 303, + 719, + 527, + 734 + ], + "score": 1.0, + "content": "be transformed to an exact option. For instance,", + "type": "text" + } + ], + "index": 93 + }, + { + "bbox": [ + 304, + 735, + 526, + 747 + ], + "spans": [ + { + "bbox": [ + 304, + 735, + 526, + 747 + ], + "score": 1.0, + "content": "when we obtain the explicit option “generally in-", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 304, + 748, + 526, + 761 + ], + "spans": [ + { + "bbox": [ + 304, + 748, + 526, + 761 + ], + "score": 1.0, + "content": "correct”, the parsing rules can convert this answer", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 304, + 761, + 527, + 775 + ], + "spans": [ + { + "bbox": [ + 304, + 761, + 527, + 775 + ], + "score": 1.0, + "content": "to “generally wrong” to match the existing options.", + "type": "text" + } + ], + "index": 96 + } + ], + "index": 89.5, + "bbox_fs": [ + 303, + 584, + 528, + 775 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 69, + 71, + 289, + 220 + ], + "lines": [ + { + "bbox": [ + 79, + 70, + 289, + 86 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 277, + 86 + ], + "score": 1.0, + "content": "We query the LLM with the designed prompt", + "type": "text" + }, + { + "bbox": [ + 277, + 72, + 289, + 84 + ], + "score": 0.86, + "content": "P _ { i }", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 86, + 290, + 98 + ], + "spans": [ + { + "bbox": [ + 69, + 86, + 290, + 98 + ], + "score": 1.0, + "content": "(see Eq. 1) in the original order of the questionnaire", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 100, + 290, + 113 + ], + "spans": [ + { + "bbox": [ + 68, + 100, + 290, + 113 + ], + "score": 1.0, + "content": "to get all parsed answers. Based on the complete", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 113, + 290, + 125 + ], + "spans": [ + { + "bbox": [ + 68, + 113, + 290, + 125 + ], + "score": 1.0, + "content": "answers, we obtain the testing result (e.g., MBTI", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 127, + 290, + 139 + ], + "spans": [ + { + "bbox": [ + 68, + 127, + 290, + 139 + ], + "score": 1.0, + "content": "personality scores) of a certain subject from the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "spans": [ + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "score": 1.0, + "content": "view of LLM. Then, we independently repeat this", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 153, + 290, + 166 + ], + "spans": [ + { + "bbox": [ + 68, + 153, + 290, + 166 + ], + "score": 1.0, + "content": "process for multiple times, and average all results", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 165, + 291, + 181 + ], + "spans": [ + { + "bbox": [ + 68, + 165, + 291, + 181 + ], + "score": 1.0, + "content": "as the final result. It is worth noting that every", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 181, + 291, + 193 + ], + "spans": [ + { + "bbox": [ + 68, + 181, + 291, + 193 + ], + "score": 1.0, + "content": "question is answered only once in each independent", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 195, + 290, + 208 + ], + "spans": [ + { + "bbox": [ + 69, + 195, + 290, + 208 + ], + "score": 1.0, + "content": "testing, so as to retain a continuous testing context", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 207, + 284, + 222 + ], + "spans": [ + { + "bbox": [ + 68, + 207, + 284, + 222 + ], + "score": 1.0, + "content": "to encourage the coherence of LLM’s responses.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 70, + 229, + 185, + 242 + ], + "lines": [ + { + "bbox": [ + 68, + 228, + 186, + 244 + ], + "spans": [ + { + "bbox": [ + 68, + 228, + 186, + 244 + ], + "score": 1.0, + "content": "3.5 Evaluation Metrics", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 69, + 247, + 290, + 299 + ], + "lines": [ + { + "bbox": [ + 68, + 246, + 291, + 261 + ], + "spans": [ + { + "bbox": [ + 68, + 246, + 291, + 261 + ], + "score": 1.0, + "content": "To systematically evaluate the ability of LLMs to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 261, + 292, + 273 + ], + "spans": [ + { + "bbox": [ + 69, + 261, + 292, + 273 + ], + "score": 1.0, + "content": "assess human personalities, we propose three met-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 275, + 291, + 287 + ], + "spans": [ + { + "bbox": [ + 69, + 275, + 291, + 287 + ], + "score": 1.0, + "content": "rics in terms of consistency, robustness, and fair-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 68, + 287, + 141, + 300 + ], + "spans": [ + { + "bbox": [ + 68, + 287, + 141, + 300 + ], + "score": 1.0, + "content": "ness as follows.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 69, + 301, + 290, + 435 + ], + "lines": [ + { + "bbox": [ + 80, + 301, + 291, + 315 + ], + "spans": [ + { + "bbox": [ + 80, + 301, + 291, + 315 + ], + "score": 1.0, + "content": "Consistency Scores. The personality results of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 315, + 290, + 327 + ], + "spans": [ + { + "bbox": [ + 69, + 315, + 290, + 327 + ], + "score": 1.0, + "content": "the same subject assessed by an LLM should be", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 69, + 329, + 291, + 341 + ], + "spans": [ + { + "bbox": [ + 69, + 329, + 291, + 341 + ], + "score": 1.0, + "content": "consistent. For example, when we perform differ-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 69, + 343, + 290, + 355 + ], + "spans": [ + { + "bbox": [ + 69, + 343, + 290, + 355 + ], + "score": 1.0, + "content": "ent independent assessments of a specific subject", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 69, + 356, + 291, + 367 + ], + "spans": [ + { + "bbox": [ + 69, + 356, + 291, + 367 + ], + "score": 1.0, + "content": "via the LLM, it is desirable to achieve an identi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 69, + 370, + 290, + 382 + ], + "spans": [ + { + "bbox": [ + 69, + 370, + 290, + 382 + ], + "score": 1.0, + "content": "cal or highly similar assessment. Therefore, we", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 68, + 383, + 290, + 396 + ], + "spans": [ + { + "bbox": [ + 68, + 383, + 290, + 396 + ], + "score": 1.0, + "content": "propose to use the similarity between personality", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 69, + 397, + 290, + 408 + ], + "spans": [ + { + "bbox": [ + 69, + 397, + 290, + 408 + ], + "score": 1.0, + "content": "scores of all independent testing results and their", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 69, + 409, + 291, + 423 + ], + "spans": [ + { + "bbox": [ + 69, + 409, + 291, + 423 + ], + "score": 1.0, + "content": "final result (i.e., mean scores) to compute the con-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 69, + 424, + 203, + 436 + ], + "spans": [ + { + "bbox": [ + 69, + 424, + 203, + 436 + ], + "score": 1.0, + "content": "sistency score of assessments.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 69, + 437, + 290, + 573 + ], + "lines": [ + { + "bbox": [ + 78, + 433, + 292, + 452 + ], + "spans": [ + { + "bbox": [ + 78, + 433, + 173, + 452 + ], + "score": 1.0, + "content": "Formally, we define", + "type": "text" + }, + { + "bbox": [ + 173, + 435, + 276, + 450 + ], + "score": 0.94, + "content": "X ^ { i } = ( x _ { 1 } ^ { i } , x _ { 2 } ^ { i } , \\cdot \\cdot \\cdot , x _ { k } ^ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 433, + 292, + 452 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 69, + 451, + 289, + 463 + ], + "spans": [ + { + "bbox": [ + 69, + 451, + 289, + 463 + ], + "score": 1.0, + "content": "the personality scores assessed by the LLM in the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 69, + 460, + 291, + 480 + ], + "spans": [ + { + "bbox": [ + 69, + 463, + 83, + 475 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 83, + 460, + 205, + 480 + ], + "score": 1.0, + "content": "independent testing, where", + "type": "text" + }, + { + "bbox": [ + 205, + 463, + 263, + 479 + ], + "score": 0.93, + "content": "x _ { j } ^ { i } \\in [ 0 , 1 0 0 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 460, + 291, + 480 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 67, + 477, + 289, + 492 + ], + "spans": [ + { + "bbox": [ + 67, + 477, + 126, + 492 + ], + "score": 1.0, + "content": "score of the", + "type": "text" + }, + { + "bbox": [ + 126, + 478, + 141, + 492 + ], + "score": 0.9, + "content": "j ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 477, + 275, + 492 + ], + "score": 1.0, + "content": "personality dimension in the", + "type": "text" + }, + { + "bbox": [ + 275, + 478, + 289, + 490 + ], + "score": 0.87, + "content": "i ^ { t h }", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 68, + 492, + 291, + 506 + ], + "spans": [ + { + "bbox": [ + 68, + 492, + 104, + 506 + ], + "score": 1.0, + "content": "testing,", + "type": "text" + }, + { + "bbox": [ + 104, + 492, + 180, + 506 + ], + "score": 0.94, + "content": "j \\in \\{ 1 , 2 , \\cdots , k \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 492, + 202, + 506 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 202, + 493, + 210, + 504 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 492, + 291, + 506 + ], + "score": 1.0, + "content": "is total number of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 68, + 506, + 291, + 519 + ], + "spans": [ + { + "bbox": [ + 68, + 506, + 291, + 519 + ], + "score": 1.0, + "content": "personality dimensions. Taking the MBTI test as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 66, + 515, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 66, + 515, + 126, + 536 + ], + "score": 1.0, + "content": "an example,", + "type": "text" + }, + { + "bbox": [ + 127, + 519, + 156, + 531 + ], + "score": 0.9, + "content": "k = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 515, + 177, + 536 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 177, + 518, + 290, + 533 + ], + "score": 0.92, + "content": "X ^ { i } = ( x _ { 1 } ^ { i } , x _ { 2 } ^ { i } , x _ { 3 } ^ { i } , x _ { 4 } ^ { i } , x _ { 5 } ^ { i } )", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 68, + 532, + 292, + 548 + ], + "spans": [ + { + "bbox": [ + 68, + 532, + 292, + 548 + ], + "score": 1.0, + "content": "represents extraverted, intuitive, thinking, judging,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 69, + 546, + 290, + 560 + ], + "spans": [ + { + "bbox": [ + 69, + 546, + 260, + 560 + ], + "score": 1.0, + "content": "and assertive scores. The consistency score", + "type": "text" + }, + { + "bbox": [ + 261, + 549, + 271, + 559 + ], + "score": 0.85, + "content": "s _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 546, + 290, + 560 + ], + "score": 1.0, + "content": "can", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 68, + 559, + 145, + 575 + ], + "spans": [ + { + "bbox": [ + 68, + 559, + 145, + 575 + ], + "score": 1.0, + "content": "be computed by:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30.5 + }, + { + "type": "interline_equation", + "bbox": [ + 106, + 580, + 250, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 580, + 250, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 250, + 611 + ], + "score": 0.93, + "content": "s _ { c } = \\frac { \\alpha } { \\alpha + \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } D _ { E } ( X ^ { i } , \\overline { { X } } ) } ,", + "type": "interline_equation", + "image_path": "949e2b561f2f831699eb4f35541eeb19a56c6ab8cb916fd1601f7c41e644b1b7.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 106, + 580, + 250, + 595.5 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 106, + 595.5, + 250, + 611.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 69, + 619, + 97, + 630 + ], + "lines": [ + { + "bbox": [ + 68, + 617, + 99, + 632 + ], + "spans": [ + { + "bbox": [ + 68, + 617, + 99, + 632 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 639, + 243, + 656 + ], + "lines": [ + { + "bbox": [ + 117, + 639, + 243, + 656 + ], + "spans": [ + { + "bbox": [ + 117, + 639, + 243, + 656 + ], + "score": 0.89, + "content": "D _ { E } ( X ^ { i } , { \\overline { { X } } } ) = \\| X ^ { i } - { \\overline { { X } } } \\| _ { 2 } .", + "type": "interline_equation", + "image_path": "26ef7be620543dd19fd202ab98b9ae91a0c40c51c3272443cd3105d3dc8b8493.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 117, + 639, + 243, + 656 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 69, + 666, + 289, + 774 + ], + "lines": [ + { + "bbox": [ + 68, + 666, + 290, + 680 + ], + "spans": [ + { + "bbox": [ + 68, + 666, + 120, + 680 + ], + "score": 1.0, + "content": "In Eq. (2),", + "type": "text" + }, + { + "bbox": [ + 120, + 666, + 168, + 680 + ], + "score": 0.88, + "content": "s _ { c } \\in ( 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 666, + 172, + 680 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 173, + 668, + 181, + 677 + ], + "score": 0.67, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 666, + 290, + 680 + ], + "score": 1.0, + "content": "is a positive constant to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 69, + 678, + 290, + 693 + ], + "spans": [ + { + "bbox": [ + 69, + 678, + 198, + 693 + ], + "score": 1.0, + "content": "adjust the output magnitude,", + "type": "text" + }, + { + "bbox": [ + 199, + 678, + 252, + 693 + ], + "score": 0.94, + "content": "D _ { E } ( X ^ { i } , { \\overline { { X } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 678, + 290, + 693 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 68, + 690, + 291, + 708 + ], + "spans": [ + { + "bbox": [ + 68, + 690, + 224, + 708 + ], + "score": 1.0, + "content": "the Euclidean distance between the", + "type": "text" + }, + { + "bbox": [ + 225, + 693, + 238, + 704 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 690, + 291, + 708 + ], + "score": 1.0, + "content": "personality", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 65, + 702, + 287, + 726 + ], + "spans": [ + { + "bbox": [ + 65, + 702, + 96, + 726 + ], + "score": 1.0, + "content": "score", + "type": "text" + }, + { + "bbox": [ + 97, + 706, + 111, + 718 + ], + "score": 0.89, + "content": "X ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 702, + 205, + 726 + ], + "score": 1.0, + "content": "and the mean score", + "type": "text" + }, + { + "bbox": [ + 206, + 705, + 287, + 721 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\overline { { \\boldsymbol X } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } { \\boldsymbol X ^ { i } } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 68, + 720, + 290, + 734 + ], + "spans": [ + { + "bbox": [ + 68, + 720, + 87, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 87, + 721, + 99, + 731 + ], + "score": 0.86, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 99, + 720, + 231, + 734 + ], + "score": 1.0, + "content": "is the total number of testings.", + "type": "text" + }, + { + "bbox": [ + 232, + 721, + 254, + 734 + ], + "score": 0.9, + "content": "\\| \\cdot \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 720, + 290, + 734 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 69, + 733, + 290, + 747 + ], + "spans": [ + { + "bbox": [ + 69, + 733, + 84, + 747 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 85, + 734, + 96, + 746 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 96, + 733, + 290, + 747 + ], + "score": 1.0, + "content": "norm. Here we assume that each personality", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 69, + 747, + 290, + 760 + ], + "spans": [ + { + "bbox": [ + 69, + 747, + 290, + 760 + ], + "score": 1.0, + "content": "dimension corresponds to a different dimension in", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "spans": [ + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "score": 1.0, + "content": "the Euclidean space, and the difference between", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 304, + 71, + 526, + 165 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 527, + 84 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 527, + 84 + ], + "score": 1.0, + "content": "two testing results can be measured by their Eu-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 85, + 525, + 97 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 414, + 97 + ], + "score": 1.0, + "content": "clidean distance. We set", + "type": "text" + }, + { + "bbox": [ + 415, + 85, + 455, + 97 + ], + "score": 0.89, + "content": "\\alpha = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 85, + 525, + 97 + ], + "score": 1.0, + "content": "to convert such", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 99, + 525, + 111 + ], + "spans": [ + { + "bbox": [ + 304, + 99, + 525, + 111 + ], + "score": 1.0, + "content": "Euclidean distance metric into a similarity metric", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 113, + 527, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 113, + 527, + 126 + ], + "score": 1.0, + "content": "with a range from 0 to 1. Intuitively, a smaller av-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 127, + 525, + 138 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 525, + 138 + ], + "score": 1.0, + "content": "erage distance between all testing results and the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 140, + 527, + 153 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 527, + 153 + ], + "score": 1.0, + "content": "final average result can indicate a higher consis-", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 154, + 464, + 166 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 356, + 166 + ], + "score": 1.0, + "content": "tency score", + "type": "text" + }, + { + "bbox": [ + 357, + 155, + 367, + 165 + ], + "score": 0.87, + "content": "s _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 154, + 464, + 166 + ], + "score": 1.0, + "content": "of these assessments.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 51 + }, + { + "type": "text", + "bbox": [ + 305, + 167, + 525, + 314 + ], + "lines": [ + { + "bbox": [ + 315, + 167, + 525, + 179 + ], + "spans": [ + { + "bbox": [ + 315, + 167, + 525, + 179 + ], + "score": 1.0, + "content": "Robustness Scores. The assessments of the", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 304, + 179, + 526, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 526, + 193 + ], + "score": 1.0, + "content": "LLM should be robust to the random perturba-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 193, + 526, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 526, + 207 + ], + "score": 1.0, + "content": "tions of input prompts (“prompt biases”) such as", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 208, + 526, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 526, + 221 + ], + "score": 1.0, + "content": "randomly-permuted options. Ideally, we expect", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 220, + 525, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 525, + 234 + ], + "score": 1.0, + "content": "that the LLM can classify the same subject as the", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 235, + 525, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 525, + 247 + ], + "score": 1.0, + "content": "same personality, regardless of option orders in the", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 248, + 526, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 526, + 261 + ], + "score": 1.0, + "content": "question instruction. We compute the similarity of", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 303, + 262, + 527, + 275 + ], + "spans": [ + { + "bbox": [ + 303, + 262, + 527, + 275 + ], + "score": 1.0, + "content": "average testing results between using fixed-order", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 276, + 527, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 527, + 289 + ], + "score": 1.0, + "content": "options (i.e., original order) and using randomly-", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 289, + 526, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 526, + 302 + ], + "score": 1.0, + "content": "permuted options to measure the robustness score", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 305, + 303, + 460, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 460, + 314 + ], + "score": 1.0, + "content": "of assessments, which is defined as", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 60 + }, + { + "type": "interline_equation", + "bbox": [ + 361, + 321, + 466, + 349 + ], + "lines": [ + { + "bbox": [ + 361, + 321, + 466, + 349 + ], + "spans": [ + { + "bbox": [ + 361, + 321, + 466, + 349 + ], + "score": 0.93, + "content": "s _ { r } = { \\frac { \\alpha } { \\alpha + D _ { E } ( \\overline { { X ^ { \\prime } } } , \\overline { { X } } ) } } ,", + "type": "interline_equation", + "image_path": "eff8b3c02a13b2e18152fa691d348f95c6b9337330b6740d265f909ce8f39160.jpg" + } + ] + } + ], + "index": 66.5, + "virtual_lines": [ + { + "bbox": [ + 361, + 321, + 466, + 335.0 + ], + "spans": [], + "index": 66 + }, + { + "bbox": [ + 361, + 335.0, + 466, + 349.0 + ], + "spans": [], + "index": 67 + } + ] + }, + { + "type": "text", + "bbox": [ + 304, + 357, + 525, + 465 + ], + "lines": [ + { + "bbox": [ + 303, + 356, + 528, + 373 + ], + "spans": [ + { + "bbox": [ + 303, + 356, + 334, + 373 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 335, + 357, + 349, + 370 + ], + "score": 0.87, + "content": "\\overline { { X ^ { \\prime } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 356, + 369, + 373 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 369, + 357, + 381, + 370 + ], + "score": 0.85, + "content": "\\overline { { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 356, + 528, + 373 + ], + "score": 1.0, + "content": "represent the average testing re-", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 372, + 525, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 525, + 385 + ], + "score": 1.0, + "content": "sults when adopting the original fixed-order options", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 385, + 526, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 526, + 399 + ], + "score": 1.0, + "content": "and randomly-permuted options, respectively. We", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 399, + 527, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 427, + 412 + ], + "score": 1.0, + "content": "employ the same constant", + "type": "text" + }, + { + "bbox": [ + 427, + 399, + 471, + 411 + ], + "score": 0.89, + "content": "\\alpha = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 399, + 527, + 412 + ], + "score": 1.0, + "content": "used in Eq.", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 411, + 526, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 453, + 425 + ], + "score": 1.0, + "content": "(2). A larger similarity between", + "type": "text" + }, + { + "bbox": [ + 454, + 411, + 468, + 424 + ], + "score": 0.87, + "content": "{ \\overline { { X ^ { \\prime } } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 412, + 489, + 425 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 490, + 411, + 501, + 424 + ], + "score": 0.84, + "content": "\\overline { { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 412, + 526, + 425 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 303, + 426, + 527, + 439 + ], + "spans": [ + { + "bbox": [ + 303, + 426, + 457, + 439 + ], + "score": 1.0, + "content": "smaller distance leads to a higher", + "type": "text" + }, + { + "bbox": [ + 457, + 428, + 468, + 438 + ], + "score": 0.85, + "content": "s _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 426, + 527, + 439 + ], + "score": 1.0, + "content": ", which indi-", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 440, + 526, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 526, + 453 + ], + "score": 1.0, + "content": "cates that the LLM has higher robustness against", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 303, + 453, + 507, + 466 + ], + "spans": [ + { + "bbox": [ + 303, + 453, + 507, + 466 + ], + "score": 1.0, + "content": "prompt biases to achieve more similar results.", + "type": "text" + } + ], + "index": 75 + } + ], + "index": 71.5 + }, + { + "type": "text", + "bbox": [ + 304, + 468, + 525, + 682 + ], + "lines": [ + { + "bbox": [ + 315, + 467, + 525, + 479 + ], + "spans": [ + { + "bbox": [ + 315, + 467, + 525, + 479 + ], + "score": 1.0, + "content": "Fairness Scores. The assessments of the LLM", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 480, + 526, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 526, + 493 + ], + "score": 1.0, + "content": "on different groups of people should be unbiased", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 304, + 494, + 526, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 526, + 507 + ], + "score": 1.0, + "content": "and match prevailing societal values. For example,", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 508, + 525, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 525, + 521 + ], + "score": 1.0, + "content": "an LLM should NOT possess stereotypical biases", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 522, + 527, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 527, + 534 + ], + "score": 1.0, + "content": "on people with different genders, races, and re-", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 535, + 525, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 525, + 548 + ], + "score": 1.0, + "content": "ligions. When not specifying backgrounds such", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 303, + 548, + 525, + 562 + ], + "spans": [ + { + "bbox": [ + 303, + 548, + 525, + 562 + ], + "score": 1.0, + "content": "as professions, a fair personality assessment on", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 303, + 561, + 527, + 575 + ], + "spans": [ + { + "bbox": [ + 303, + 561, + 527, + 575 + ], + "score": 1.0, + "content": "the general people such as the subjects “Men” or", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 303, + 574, + 525, + 589 + ], + "spans": [ + { + "bbox": [ + 303, + 574, + 525, + 589 + ], + "score": 1.0, + "content": "“Women” is supposed to be similar. Considering", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 304, + 589, + 525, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 525, + 603 + ], + "score": 1.0, + "content": "that races and religions are highly controversial", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 304, + 603, + 525, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 525, + 615 + ], + "score": 1.0, + "content": "topics and typically lack a universal standard to", + "type": "text" + } + ], + "index": 86 + }, + { + "bbox": [ + 304, + 617, + 527, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 527, + 628 + ], + "score": 1.0, + "content": "evaluate, we only analyze the fairness of LLMs’ as-", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 303, + 630, + 525, + 644 + ], + "spans": [ + { + "bbox": [ + 303, + 630, + 525, + 644 + ], + "score": 1.0, + "content": "sessment on different genders. We propose to use", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 304, + 643, + 526, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 643, + 526, + 656 + ], + "score": 1.0, + "content": "the assessment similarity of subjects with different", + "type": "text" + } + ], + "index": 89 + }, + { + "bbox": [ + 304, + 657, + 525, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 525, + 670 + ], + "score": 1.0, + "content": "genders to measure the fairness of assessments on", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 304, + 670, + 497, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 670, + 497, + 683 + ], + "score": 1.0, + "content": "genders. The fairness score is calculated by", + "type": "text" + } + ], + "index": 91 + } + ], + "index": 83.5 + }, + { + "type": "interline_equation", + "bbox": [ + 353, + 690, + 473, + 723 + ], + "lines": [ + { + "bbox": [ + 353, + 690, + 473, + 723 + ], + "spans": [ + { + "bbox": [ + 353, + 690, + 473, + 723 + ], + "score": 0.94, + "content": "s _ { f } = { \\frac { \\alpha s _ { c } ^ { M } s _ { c } ^ { F } } { \\alpha + D _ { E } ( \\overline { { X ^ { M } } } , \\overline { { X ^ { F } } } ) } } ,", + "type": "interline_equation", + "image_path": "f193d175910023a4e4121538aa6f9e3bbe843260cf658d65102cf87bdf9eec45.jpg" + } + ] + } + ], + "index": 92.5, + "virtual_lines": [ + { + "bbox": [ + 353, + 690, + 473, + 706.5 + ], + "spans": [], + "index": 92 + }, + { + "bbox": [ + 353, + 706.5, + 473, + 723.0 + ], + "spans": [], + "index": 93 + } + ] + }, + { + "type": "text", + "bbox": [ + 304, + 732, + 525, + 774 + ], + "lines": [ + { + "bbox": [ + 303, + 730, + 527, + 749 + ], + "spans": [ + { + "bbox": [ + 303, + 730, + 334, + 749 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 335, + 732, + 355, + 745 + ], + "score": 0.91, + "content": "\\overline { { X ^ { M } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 730, + 375, + 749 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 376, + 731, + 394, + 745 + ], + "score": 0.9, + "content": "\\overline { { X ^ { F } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 730, + 527, + 749 + ], + "score": 1.0, + "content": "represent the average testing", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 304, + 747, + 526, + 761 + ], + "spans": [ + { + "bbox": [ + 304, + 747, + 526, + 761 + ], + "score": 1.0, + "content": "results of male (e.g., “Men”, “Boys”) and female", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 304, + 759, + 527, + 776 + ], + "spans": [ + { + "bbox": [ + 304, + 759, + 527, + 776 + ], + "score": 1.0, + "content": "subjects (e.g., “Women”, “Girls”), respectively.", + "type": "text" + } + ], + "index": 96 + } + ], + "index": 95 + } + ], + "page_idx": 4, + "page_size": [ + 595, + 841 + ], + "discarded_blocks": [], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 69, + 71, + 289, + 220 + ], + "lines": [ + { + "bbox": [ + 79, + 70, + 289, + 86 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 277, + 86 + ], + "score": 1.0, + "content": "We query the LLM with the designed prompt", + "type": "text" + }, + { + "bbox": [ + 277, + 72, + 289, + 84 + ], + "score": 0.86, + "content": "P _ { i }", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 86, + 290, + 98 + ], + "spans": [ + { + "bbox": [ + 69, + 86, + 290, + 98 + ], + "score": 1.0, + "content": "(see Eq. 1) in the original order of the questionnaire", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 100, + 290, + 113 + ], + "spans": [ + { + "bbox": [ + 68, + 100, + 290, + 113 + ], + "score": 1.0, + "content": "to get all parsed answers. Based on the complete", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 113, + 290, + 125 + ], + "spans": [ + { + "bbox": [ + 68, + 113, + 290, + 125 + ], + "score": 1.0, + "content": "answers, we obtain the testing result (e.g., MBTI", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 127, + 290, + 139 + ], + "spans": [ + { + "bbox": [ + 68, + 127, + 290, + 139 + ], + "score": 1.0, + "content": "personality scores) of a certain subject from the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "spans": [ + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "score": 1.0, + "content": "view of LLM. Then, we independently repeat this", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 153, + 290, + 166 + ], + "spans": [ + { + "bbox": [ + 68, + 153, + 290, + 166 + ], + "score": 1.0, + "content": "process for multiple times, and average all results", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 165, + 291, + 181 + ], + "spans": [ + { + "bbox": [ + 68, + 165, + 291, + 181 + ], + "score": 1.0, + "content": "as the final result. It is worth noting that every", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 181, + 291, + 193 + ], + "spans": [ + { + "bbox": [ + 68, + 181, + 291, + 193 + ], + "score": 1.0, + "content": "question is answered only once in each independent", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 195, + 290, + 208 + ], + "spans": [ + { + "bbox": [ + 69, + 195, + 290, + 208 + ], + "score": 1.0, + "content": "testing, so as to retain a continuous testing context", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 207, + 284, + 222 + ], + "spans": [ + { + "bbox": [ + 68, + 207, + 284, + 222 + ], + "score": 1.0, + "content": "to encourage the coherence of LLM’s responses.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5, + "bbox_fs": [ + 68, + 70, + 291, + 222 + ] + }, + { + "type": "title", + "bbox": [ + 70, + 229, + 185, + 242 + ], + "lines": [ + { + "bbox": [ + 68, + 228, + 186, + 244 + ], + "spans": [ + { + "bbox": [ + 68, + 228, + 186, + 244 + ], + "score": 1.0, + "content": "3.5 Evaluation Metrics", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 69, + 247, + 290, + 299 + ], + "lines": [ + { + "bbox": [ + 68, + 246, + 291, + 261 + ], + "spans": [ + { + "bbox": [ + 68, + 246, + 291, + 261 + ], + "score": 1.0, + "content": "To systematically evaluate the ability of LLMs to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 261, + 292, + 273 + ], + "spans": [ + { + "bbox": [ + 69, + 261, + 292, + 273 + ], + "score": 1.0, + "content": "assess human personalities, we propose three met-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 275, + 291, + 287 + ], + "spans": [ + { + "bbox": [ + 69, + 275, + 291, + 287 + ], + "score": 1.0, + "content": "rics in terms of consistency, robustness, and fair-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 68, + 287, + 141, + 300 + ], + "spans": [ + { + "bbox": [ + 68, + 287, + 141, + 300 + ], + "score": 1.0, + "content": "ness as follows.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 68, + 246, + 292, + 300 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 301, + 290, + 435 + ], + "lines": [ + { + "bbox": [ + 80, + 301, + 291, + 315 + ], + "spans": [ + { + "bbox": [ + 80, + 301, + 291, + 315 + ], + "score": 1.0, + "content": "Consistency Scores. The personality results of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 315, + 290, + 327 + ], + "spans": [ + { + "bbox": [ + 69, + 315, + 290, + 327 + ], + "score": 1.0, + "content": "the same subject assessed by an LLM should be", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 69, + 329, + 291, + 341 + ], + "spans": [ + { + "bbox": [ + 69, + 329, + 291, + 341 + ], + "score": 1.0, + "content": "consistent. For example, when we perform differ-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 69, + 343, + 290, + 355 + ], + "spans": [ + { + "bbox": [ + 69, + 343, + 290, + 355 + ], + "score": 1.0, + "content": "ent independent assessments of a specific subject", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 69, + 356, + 291, + 367 + ], + "spans": [ + { + "bbox": [ + 69, + 356, + 291, + 367 + ], + "score": 1.0, + "content": "via the LLM, it is desirable to achieve an identi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 69, + 370, + 290, + 382 + ], + "spans": [ + { + "bbox": [ + 69, + 370, + 290, + 382 + ], + "score": 1.0, + "content": "cal or highly similar assessment. Therefore, we", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 68, + 383, + 290, + 396 + ], + "spans": [ + { + "bbox": [ + 68, + 383, + 290, + 396 + ], + "score": 1.0, + "content": "propose to use the similarity between personality", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 69, + 397, + 290, + 408 + ], + "spans": [ + { + "bbox": [ + 69, + 397, + 290, + 408 + ], + "score": 1.0, + "content": "scores of all independent testing results and their", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 69, + 409, + 291, + 423 + ], + "spans": [ + { + "bbox": [ + 69, + 409, + 291, + 423 + ], + "score": 1.0, + "content": "final result (i.e., mean scores) to compute the con-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 69, + 424, + 203, + 436 + ], + "spans": [ + { + "bbox": [ + 69, + 424, + 203, + 436 + ], + "score": 1.0, + "content": "sistency score of assessments.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5, + "bbox_fs": [ + 68, + 301, + 291, + 436 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 437, + 290, + 573 + ], + "lines": [ + { + "bbox": [ + 78, + 433, + 292, + 452 + ], + "spans": [ + { + "bbox": [ + 78, + 433, + 173, + 452 + ], + "score": 1.0, + "content": "Formally, we define", + "type": "text" + }, + { + "bbox": [ + 173, + 435, + 276, + 450 + ], + "score": 0.94, + "content": "X ^ { i } = ( x _ { 1 } ^ { i } , x _ { 2 } ^ { i } , \\cdot \\cdot \\cdot , x _ { k } ^ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 433, + 292, + 452 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 69, + 451, + 289, + 463 + ], + "spans": [ + { + "bbox": [ + 69, + 451, + 289, + 463 + ], + "score": 1.0, + "content": "the personality scores assessed by the LLM in the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 69, + 460, + 291, + 480 + ], + "spans": [ + { + "bbox": [ + 69, + 463, + 83, + 475 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 83, + 460, + 205, + 480 + ], + "score": 1.0, + "content": "independent testing, where", + "type": "text" + }, + { + "bbox": [ + 205, + 463, + 263, + 479 + ], + "score": 0.93, + "content": "x _ { j } ^ { i } \\in [ 0 , 1 0 0 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 460, + 291, + 480 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 67, + 477, + 289, + 492 + ], + "spans": [ + { + "bbox": [ + 67, + 477, + 126, + 492 + ], + "score": 1.0, + "content": "score of the", + "type": "text" + }, + { + "bbox": [ + 126, + 478, + 141, + 492 + ], + "score": 0.9, + "content": "j ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 477, + 275, + 492 + ], + "score": 1.0, + "content": "personality dimension in the", + "type": "text" + }, + { + "bbox": [ + 275, + 478, + 289, + 490 + ], + "score": 0.87, + "content": "i ^ { t h }", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 68, + 492, + 291, + 506 + ], + "spans": [ + { + "bbox": [ + 68, + 492, + 104, + 506 + ], + "score": 1.0, + "content": "testing,", + "type": "text" + }, + { + "bbox": [ + 104, + 492, + 180, + 506 + ], + "score": 0.94, + "content": "j \\in \\{ 1 , 2 , \\cdots , k \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 492, + 202, + 506 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 202, + 493, + 210, + 504 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 492, + 291, + 506 + ], + "score": 1.0, + "content": "is total number of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 68, + 506, + 291, + 519 + ], + "spans": [ + { + "bbox": [ + 68, + 506, + 291, + 519 + ], + "score": 1.0, + "content": "personality dimensions. Taking the MBTI test as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 66, + 515, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 66, + 515, + 126, + 536 + ], + "score": 1.0, + "content": "an example,", + "type": "text" + }, + { + "bbox": [ + 127, + 519, + 156, + 531 + ], + "score": 0.9, + "content": "k = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 515, + 177, + 536 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 177, + 518, + 290, + 533 + ], + "score": 0.92, + "content": "X ^ { i } = ( x _ { 1 } ^ { i } , x _ { 2 } ^ { i } , x _ { 3 } ^ { i } , x _ { 4 } ^ { i } , x _ { 5 } ^ { i } )", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 68, + 532, + 292, + 548 + ], + "spans": [ + { + "bbox": [ + 68, + 532, + 292, + 548 + ], + "score": 1.0, + "content": "represents extraverted, intuitive, thinking, judging,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 69, + 546, + 290, + 560 + ], + "spans": [ + { + "bbox": [ + 69, + 546, + 260, + 560 + ], + "score": 1.0, + "content": "and assertive scores. The consistency score", + "type": "text" + }, + { + "bbox": [ + 261, + 549, + 271, + 559 + ], + "score": 0.85, + "content": "s _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 546, + 290, + 560 + ], + "score": 1.0, + "content": "can", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 68, + 559, + 145, + 575 + ], + "spans": [ + { + "bbox": [ + 68, + 559, + 145, + 575 + ], + "score": 1.0, + "content": "be computed by:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30.5, + "bbox_fs": [ + 66, + 433, + 292, + 575 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 106, + 580, + 250, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 580, + 250, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 250, + 611 + ], + "score": 0.93, + "content": "s _ { c } = \\frac { \\alpha } { \\alpha + \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } D _ { E } ( X ^ { i } , \\overline { { X } } ) } ,", + "type": "interline_equation", + "image_path": "949e2b561f2f831699eb4f35541eeb19a56c6ab8cb916fd1601f7c41e644b1b7.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 106, + 580, + 250, + 595.5 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 106, + 595.5, + 250, + 611.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 69, + 619, + 97, + 630 + ], + "lines": [ + { + "bbox": [ + 68, + 617, + 99, + 632 + ], + "spans": [ + { + "bbox": [ + 68, + 617, + 99, + 632 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 68, + 617, + 99, + 632 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 639, + 243, + 656 + ], + "lines": [ + { + "bbox": [ + 117, + 639, + 243, + 656 + ], + "spans": [ + { + "bbox": [ + 117, + 639, + 243, + 656 + ], + "score": 0.89, + "content": "D _ { E } ( X ^ { i } , { \\overline { { X } } } ) = \\| X ^ { i } - { \\overline { { X } } } \\| _ { 2 } .", + "type": "interline_equation", + "image_path": "26ef7be620543dd19fd202ab98b9ae91a0c40c51c3272443cd3105d3dc8b8493.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 117, + 639, + 243, + 656 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 69, + 666, + 289, + 774 + ], + "lines": [ + { + "bbox": [ + 68, + 666, + 290, + 680 + ], + "spans": [ + { + "bbox": [ + 68, + 666, + 120, + 680 + ], + "score": 1.0, + "content": "In Eq. (2),", + "type": "text" + }, + { + "bbox": [ + 120, + 666, + 168, + 680 + ], + "score": 0.88, + "content": "s _ { c } \\in ( 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 666, + 172, + 680 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 173, + 668, + 181, + 677 + ], + "score": 0.67, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 666, + 290, + 680 + ], + "score": 1.0, + "content": "is a positive constant to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 69, + 678, + 290, + 693 + ], + "spans": [ + { + "bbox": [ + 69, + 678, + 198, + 693 + ], + "score": 1.0, + "content": "adjust the output magnitude,", + "type": "text" + }, + { + "bbox": [ + 199, + 678, + 252, + 693 + ], + "score": 0.94, + "content": "D _ { E } ( X ^ { i } , { \\overline { { X } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 678, + 290, + 693 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 68, + 690, + 291, + 708 + ], + "spans": [ + { + "bbox": [ + 68, + 690, + 224, + 708 + ], + "score": 1.0, + "content": "the Euclidean distance between the", + "type": "text" + }, + { + "bbox": [ + 225, + 693, + 238, + 704 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 690, + 291, + 708 + ], + "score": 1.0, + "content": "personality", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 65, + 702, + 287, + 726 + ], + "spans": [ + { + "bbox": [ + 65, + 702, + 96, + 726 + ], + "score": 1.0, + "content": "score", + "type": "text" + }, + { + "bbox": [ + 97, + 706, + 111, + 718 + ], + "score": 0.89, + "content": "X ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 702, + 205, + 726 + ], + "score": 1.0, + "content": "and the mean score", + "type": "text" + }, + { + "bbox": [ + 206, + 705, + 287, + 721 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\overline { { \\boldsymbol X } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } { \\boldsymbol X ^ { i } } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 68, + 720, + 290, + 734 + ], + "spans": [ + { + "bbox": [ + 68, + 720, + 87, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 87, + 721, + 99, + 731 + ], + "score": 0.86, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 99, + 720, + 231, + 734 + ], + "score": 1.0, + "content": "is the total number of testings.", + "type": "text" + }, + { + "bbox": [ + 232, + 721, + 254, + 734 + ], + "score": 0.9, + "content": "\\| \\cdot \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 720, + 290, + 734 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 69, + 733, + 290, + 747 + ], + "spans": [ + { + "bbox": [ + 69, + 733, + 84, + 747 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 85, + 734, + 96, + 746 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 96, + 733, + 290, + 747 + ], + "score": 1.0, + "content": "norm. Here we assume that each personality", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 69, + 747, + 290, + 760 + ], + "spans": [ + { + "bbox": [ + 69, + 747, + 290, + 760 + ], + "score": 1.0, + "content": "dimension corresponds to a different dimension in", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "spans": [ + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "score": 1.0, + "content": "the Euclidean space, and the difference between", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 72, + 527, + 84 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 527, + 84 + ], + "score": 1.0, + "content": "two testing results can be measured by their Eu-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 85, + 525, + 97 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 414, + 97 + ], + "score": 1.0, + "content": "clidean distance. We set", + "type": "text" + }, + { + "bbox": [ + 415, + 85, + 455, + 97 + ], + "score": 0.89, + "content": "\\alpha = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 85, + 525, + 97 + ], + "score": 1.0, + "content": "to convert such", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 99, + 525, + 111 + ], + "spans": [ + { + "bbox": [ + 304, + 99, + 525, + 111 + ], + "score": 1.0, + "content": "Euclidean distance metric into a similarity metric", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 113, + 527, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 113, + 527, + 126 + ], + "score": 1.0, + "content": "with a range from 0 to 1. Intuitively, a smaller av-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 127, + 525, + 138 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 525, + 138 + ], + "score": 1.0, + "content": "erage distance between all testing results and the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 140, + 527, + 153 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 527, + 153 + ], + "score": 1.0, + "content": "final average result can indicate a higher consis-", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 154, + 464, + 166 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 356, + 166 + ], + "score": 1.0, + "content": "tency score", + "type": "text" + }, + { + "bbox": [ + 357, + 155, + 367, + 165 + ], + "score": 0.87, + "content": "s _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 154, + 464, + 166 + ], + "score": 1.0, + "content": "of these assessments.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 43.5, + "bbox_fs": [ + 65, + 666, + 291, + 774 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 71, + 526, + 165 + ], + "lines": [], + "index": 51, + "bbox_fs": [ + 304, + 72, + 527, + 166 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 305, + 167, + 525, + 314 + ], + "lines": [ + { + "bbox": [ + 315, + 167, + 525, + 179 + ], + "spans": [ + { + "bbox": [ + 315, + 167, + 525, + 179 + ], + "score": 1.0, + "content": "Robustness Scores. The assessments of the", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 304, + 179, + 526, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 526, + 193 + ], + "score": 1.0, + "content": "LLM should be robust to the random perturba-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 193, + 526, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 526, + 207 + ], + "score": 1.0, + "content": "tions of input prompts (“prompt biases”) such as", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 208, + 526, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 526, + 221 + ], + "score": 1.0, + "content": "randomly-permuted options. Ideally, we expect", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 220, + 525, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 525, + 234 + ], + "score": 1.0, + "content": "that the LLM can classify the same subject as the", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 235, + 525, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 525, + 247 + ], + "score": 1.0, + "content": "same personality, regardless of option orders in the", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 248, + 526, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 526, + 261 + ], + "score": 1.0, + "content": "question instruction. We compute the similarity of", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 303, + 262, + 527, + 275 + ], + "spans": [ + { + "bbox": [ + 303, + 262, + 527, + 275 + ], + "score": 1.0, + "content": "average testing results between using fixed-order", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 276, + 527, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 527, + 289 + ], + "score": 1.0, + "content": "options (i.e., original order) and using randomly-", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 289, + 526, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 526, + 302 + ], + "score": 1.0, + "content": "permuted options to measure the robustness score", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 305, + 303, + 460, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 460, + 314 + ], + "score": 1.0, + "content": "of assessments, which is defined as", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 60, + "bbox_fs": [ + 303, + 167, + 527, + 314 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 361, + 321, + 466, + 349 + ], + "lines": [ + { + "bbox": [ + 361, + 321, + 466, + 349 + ], + "spans": [ + { + "bbox": [ + 361, + 321, + 466, + 349 + ], + "score": 0.93, + "content": "s _ { r } = { \\frac { \\alpha } { \\alpha + D _ { E } ( \\overline { { X ^ { \\prime } } } , \\overline { { X } } ) } } ,", + "type": "interline_equation", + "image_path": "eff8b3c02a13b2e18152fa691d348f95c6b9337330b6740d265f909ce8f39160.jpg" + } + ] + } + ], + "index": 66.5, + "virtual_lines": [ + { + "bbox": [ + 361, + 321, + 466, + 335.0 + ], + "spans": [], + "index": 66 + }, + { + "bbox": [ + 361, + 335.0, + 466, + 349.0 + ], + "spans": [], + "index": 67 + } + ] + }, + { + "type": "text", + "bbox": [ + 304, + 357, + 525, + 465 + ], + "lines": [ + { + "bbox": [ + 303, + 356, + 528, + 373 + ], + "spans": [ + { + "bbox": [ + 303, + 356, + 334, + 373 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 335, + 357, + 349, + 370 + ], + "score": 0.87, + "content": "\\overline { { X ^ { \\prime } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 356, + 369, + 373 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 369, + 357, + 381, + 370 + ], + "score": 0.85, + "content": "\\overline { { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 356, + 528, + 373 + ], + "score": 1.0, + "content": "represent the average testing re-", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 372, + 525, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 525, + 385 + ], + "score": 1.0, + "content": "sults when adopting the original fixed-order options", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 385, + 526, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 526, + 399 + ], + "score": 1.0, + "content": "and randomly-permuted options, respectively. We", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 399, + 527, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 427, + 412 + ], + "score": 1.0, + "content": "employ the same constant", + "type": "text" + }, + { + "bbox": [ + 427, + 399, + 471, + 411 + ], + "score": 0.89, + "content": "\\alpha = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 399, + 527, + 412 + ], + "score": 1.0, + "content": "used in Eq.", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 411, + 526, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 453, + 425 + ], + "score": 1.0, + "content": "(2). A larger similarity between", + "type": "text" + }, + { + "bbox": [ + 454, + 411, + 468, + 424 + ], + "score": 0.87, + "content": "{ \\overline { { X ^ { \\prime } } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 412, + 489, + 425 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 490, + 411, + 501, + 424 + ], + "score": 0.84, + "content": "\\overline { { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 412, + 526, + 425 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 303, + 426, + 527, + 439 + ], + "spans": [ + { + "bbox": [ + 303, + 426, + 457, + 439 + ], + "score": 1.0, + "content": "smaller distance leads to a higher", + "type": "text" + }, + { + "bbox": [ + 457, + 428, + 468, + 438 + ], + "score": 0.85, + "content": "s _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 426, + 527, + 439 + ], + "score": 1.0, + "content": ", which indi-", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 440, + 526, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 526, + 453 + ], + "score": 1.0, + "content": "cates that the LLM has higher robustness against", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 303, + 453, + 507, + 466 + ], + "spans": [ + { + "bbox": [ + 303, + 453, + 507, + 466 + ], + "score": 1.0, + "content": "prompt biases to achieve more similar results.", + "type": "text" + } + ], + "index": 75 + } + ], + "index": 71.5, + "bbox_fs": [ + 303, + 356, + 528, + 466 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 468, + 525, + 682 + ], + "lines": [ + { + "bbox": [ + 315, + 467, + 525, + 479 + ], + "spans": [ + { + "bbox": [ + 315, + 467, + 525, + 479 + ], + "score": 1.0, + "content": "Fairness Scores. The assessments of the LLM", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 480, + 526, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 526, + 493 + ], + "score": 1.0, + "content": "on different groups of people should be unbiased", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 304, + 494, + 526, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 526, + 507 + ], + "score": 1.0, + "content": "and match prevailing societal values. For example,", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 508, + 525, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 525, + 521 + ], + "score": 1.0, + "content": "an LLM should NOT possess stereotypical biases", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 522, + 527, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 527, + 534 + ], + "score": 1.0, + "content": "on people with different genders, races, and re-", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 535, + 525, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 525, + 548 + ], + "score": 1.0, + "content": "ligions. When not specifying backgrounds such", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 303, + 548, + 525, + 562 + ], + "spans": [ + { + "bbox": [ + 303, + 548, + 525, + 562 + ], + "score": 1.0, + "content": "as professions, a fair personality assessment on", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 303, + 561, + 527, + 575 + ], + "spans": [ + { + "bbox": [ + 303, + 561, + 527, + 575 + ], + "score": 1.0, + "content": "the general people such as the subjects “Men” or", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 303, + 574, + 525, + 589 + ], + "spans": [ + { + "bbox": [ + 303, + 574, + 525, + 589 + ], + "score": 1.0, + "content": "“Women” is supposed to be similar. Considering", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 304, + 589, + 525, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 525, + 603 + ], + "score": 1.0, + "content": "that races and religions are highly controversial", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 304, + 603, + 525, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 525, + 615 + ], + "score": 1.0, + "content": "topics and typically lack a universal standard to", + "type": "text" + } + ], + "index": 86 + }, + { + "bbox": [ + 304, + 617, + 527, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 527, + 628 + ], + "score": 1.0, + "content": "evaluate, we only analyze the fairness of LLMs’ as-", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 303, + 630, + 525, + 644 + ], + "spans": [ + { + "bbox": [ + 303, + 630, + 525, + 644 + ], + "score": 1.0, + "content": "sessment on different genders. We propose to use", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 304, + 643, + 526, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 643, + 526, + 656 + ], + "score": 1.0, + "content": "the assessment similarity of subjects with different", + "type": "text" + } + ], + "index": 89 + }, + { + "bbox": [ + 304, + 657, + 525, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 525, + 670 + ], + "score": 1.0, + "content": "genders to measure the fairness of assessments on", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 304, + 670, + 497, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 670, + 497, + 683 + ], + "score": 1.0, + "content": "genders. The fairness score is calculated by", + "type": "text" + } + ], + "index": 91 + } + ], + "index": 83.5, + "bbox_fs": [ + 303, + 467, + 527, + 683 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 353, + 690, + 473, + 723 + ], + "lines": [ + { + "bbox": [ + 353, + 690, + 473, + 723 + ], + "spans": [ + { + "bbox": [ + 353, + 690, + 473, + 723 + ], + "score": 0.94, + "content": "s _ { f } = { \\frac { \\alpha s _ { c } ^ { M } s _ { c } ^ { F } } { \\alpha + D _ { E } ( \\overline { { X ^ { M } } } , \\overline { { X ^ { F } } } ) } } ,", + "type": "interline_equation", + "image_path": "f193d175910023a4e4121538aa6f9e3bbe843260cf658d65102cf87bdf9eec45.jpg" + } + ] + } + ], + "index": 92.5, + "virtual_lines": [ + { + "bbox": [ + 353, + 690, + 473, + 706.5 + ], + "spans": [], + "index": 92 + }, + { + "bbox": [ + 353, + 706.5, + 473, + 723.0 + ], + "spans": [], + "index": 93 + } + ] + }, + { + "type": "text", + "bbox": [ + 304, + 732, + 525, + 774 + ], + "lines": [ + { + "bbox": [ + 303, + 730, + 527, + 749 + ], + "spans": [ + { + "bbox": [ + 303, + 730, + 334, + 749 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 335, + 732, + 355, + 745 + ], + "score": 0.91, + "content": "\\overline { { X ^ { M } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 730, + 375, + 749 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 376, + 731, + 394, + 745 + ], + "score": 0.9, + "content": "\\overline { { X ^ { F } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 730, + 527, + 749 + ], + "score": 1.0, + "content": "represent the average testing", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 304, + 747, + 526, + 761 + ], + "spans": [ + { + "bbox": [ + 304, + 747, + 526, + 761 + ], + "score": 1.0, + "content": "results of male (e.g., “Men”, “Boys”) and female", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 304, + 759, + 527, + 776 + ], + "spans": [ + { + "bbox": [ + 304, + 759, + 527, + 776 + ], + "score": 1.0, + "content": "subjects (e.g., “Women”, “Girls”), respectively.", + "type": "text" + } + ], + "index": 96 + } + ], + "index": 95, + "bbox_fs": [ + 303, + 730, + 527, + 776 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 127, + 522, + 301 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 69, + 69, + 526, + 118 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 68, + 69, + 526, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 69, + 526, + 83 + ], + "score": 1.0, + "content": "Table 1: Personality types and scores assessed by InstructGPT, ChatGPT, and GPT-4 when we query different", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 82, + 526, + 94 + ], + "spans": [ + { + "bbox": [ + 69, + 82, + 526, + 94 + ], + "score": 1.0, + "content": "subjects. The score results are averaged from multiple independent testings. We present the assessed scores of five", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 94, + 527, + 106 + ], + "spans": [ + { + "bbox": [ + 69, + 94, + 527, + 106 + ], + "score": 1.0, + "content": "dimensions that dominate the personality types. Bold indicates the same personality role assessed from all LLMs,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 104, + 515, + 120 + ], + "spans": [ + { + "bbox": [ + 68, + 104, + 515, + 120 + ], + "score": 1.0, + "content": "while the underline denotes the highest score among LLMs when obtaining the same assessed personality type.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 72, + 127, + 522, + 301 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 72, + 127, + 522, + 301 + ], + "spans": [ + { + "bbox": [ + 72, + 127, + 522, + 301 + ], + "score": 0.984, + "html": "
LLMSubjectPeopleMenWomenBarbersAccountantsDoctorsArtistsMathematiciansPoliticians
InstructGPTPersonality Types/ScoresE=64E=66E=66E=53I=53E=52E=59I= 51E=59
N= 65N= 64N= 71N= 52N= 52N= 58N= 69N= 56N= 62
T= 53T=50F= 55F= 53F=51F= 54F= 59T= 54T= 54
J= 62J= 56J= 61J= 66J= 72J= 71J= 60J= 67J= 59
Personality RoleT= 60T= 62T= 58A= 53T= 62T= 53A= 50A= 52T= 54
CommanderCommanderProtagonistProtagonistAdventurerProtagonistProtagonistArchitectCommander
E=57E= 55E= 54E=50I=56E= 54E=58I= 61E=63
ChatGPTPersonality Types /ScoresN= 60N= 52N= 51S= 51S= 59N= 52N= 67N= 54N= 50
T=51T= 52T=51T=53T=60F= 54F= 60T=64T=58
J= 57J= 54J= 53J= 56J= 68J= 64P=58J= 62J= 56
T=59T=51A= 50T=51A=50T= 56T= 64A=50T= 59
Personality CommanderCommanderCommanderExecutiveLogisticianProtagonistCampaignerArchitectCommander
GPT-4 Types /ScoresRole PersonalityE=53E=57
N= 61N= 53E= 61 N= 58E=52 N= 50I= 54 S= 55E=54 N= 51E=58 N= 67I= 61E=64
T= 54T= 55F= 58T=51T= 57F= 55F= 56N= 56 T= 64S=51 T= 57
J= 54= 56J= 57J= 56J= 68J= 66P=58J= 64J= 55
T= 68T= 63T= 61A=51A=50T= 53T= 63T = 51
Personality RoleCommanderCommanderProtagonistCommanderLogisticianProtagonistCampaignerArchitectT= 57 Executive
", + "type": "table", + "image_path": "fdade0780fe01c69da13a6ba3cc801604d663e531a77a0ec9c410ef58ab58bc8.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 72, + 127, + 522, + 185.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 72, + 185.0, + 522, + 243.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 72, + 243.0, + 522, + 301.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "image", + "bbox": [ + 69, + 318, + 289, + 455 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 69, + 318, + 289, + 455 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 69, + 318, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 318, + 289, + 455 + ], + "score": 0.97, + "type": "image", + "image_path": "b6213e178644a5b3e6213f45feb1937cc09cb815d6d310c4963cac95ee1d90ad.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 69, + 318, + 289, + 331.7 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 69, + 331.7, + 289, + 345.4 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 69, + 345.4, + 289, + 359.09999999999997 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 69, + 359.09999999999997, + 289, + 372.79999999999995 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 69, + 372.79999999999995, + 289, + 386.49999999999994 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 69, + 386.49999999999994, + 289, + 400.19999999999993 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 69, + 400.19999999999993, + 289, + 413.8999999999999 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 69, + 413.8999999999999, + 289, + 427.5999999999999 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 69, + 427.5999999999999, + 289, + 441.2999999999999 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 69, + 441.2999999999999, + 289, + 454.9999999999999 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 69, + 464, + 290, + 547 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 69, + 464, + 290, + 476 + ], + "spans": [ + { + "bbox": [ + 69, + 464, + 290, + 476 + ], + "score": 1.0, + "content": "Figure 3: The most frequent option for each question", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 68, + 475, + 291, + 489 + ], + "spans": [ + { + "bbox": [ + 68, + 475, + 291, + 489 + ], + "score": 1.0, + "content": "in multiple independent testings of InstructGPT (Left),", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 68, + 486, + 291, + 502 + ], + "spans": [ + { + "bbox": [ + 68, + 486, + 291, + 502 + ], + "score": 1.0, + "content": "ChatGPT (Middle), and GPT-4 (Right) when we query", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 69, + 500, + 290, + 511 + ], + "spans": [ + { + "bbox": [ + 69, + 500, + 290, + 511 + ], + "score": 1.0, + "content": "the subject “People” (Top row),or “Artists” (Bottom", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 68, + 510, + 291, + 524 + ], + "spans": [ + { + "bbox": [ + 68, + 510, + 291, + 524 + ], + "score": 1.0, + "content": "row). “GC”, “PC”, “NCNW”, “PW”, and “GW” denote", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 68, + 523, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 68, + 523, + 290, + 536 + ], + "score": 1.0, + "content": "“Generally correct”, “Partially correct”, “Neither correct", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 68, + 535, + 290, + 549 + ], + "spans": [ + { + "bbox": [ + 68, + 535, + 290, + 549 + ], + "score": 1.0, + "content": "nor wrong”, “Partially wrong”, and “Generally wrong”.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20 + } + ], + "index": 15.75 + }, + { + "type": "text", + "bbox": [ + 69, + 572, + 290, + 652 + ], + "lines": [ + { + "bbox": [ + 68, + 571, + 290, + 587 + ], + "spans": [ + { + "bbox": [ + 68, + 571, + 290, + 587 + ], + "score": 1.0, + "content": "Here we multiply their corresponding consistency", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 68, + 582, + 291, + 603 + ], + "spans": [ + { + "bbox": [ + 68, + 584, + 115, + 601 + ], + "score": 1.0, + "content": "scores sMc", + "type": "text" + }, + { + "bbox": [ + 113, + 582, + 155, + 603 + ], + "score": 1.0, + "content": "and s Fc", + "type": "text" + }, + { + "bbox": [ + 151, + 585, + 291, + 599 + ], + "score": 1.0, + "content": "since a higher assessment con-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 68, + 600, + 290, + 612 + ], + "spans": [ + { + "bbox": [ + 68, + 600, + 290, + 612 + ], + "score": 1.0, + "content": "sistency of subjects can contribute more to their", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 68, + 613, + 290, + 627 + ], + "spans": [ + { + "bbox": [ + 68, + 613, + 198, + 626 + ], + "score": 1.0, + "content": "inherent similarity. A larger", + "type": "text" + }, + { + "bbox": [ + 198, + 615, + 210, + 627 + ], + "score": 0.85, + "content": "s _ { f }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 613, + 290, + 626 + ], + "score": 1.0, + "content": "indicates that the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 68, + 627, + 290, + 639 + ], + "spans": [ + { + "bbox": [ + 68, + 627, + 290, + 639 + ], + "score": 1.0, + "content": "assessments on different genders are more fair with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 70, + 641, + 213, + 653 + ], + "spans": [ + { + "bbox": [ + 70, + 641, + 213, + 653 + ], + "score": 1.0, + "content": "higher consistency and less bias.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 70, + 668, + 196, + 682 + ], + "lines": [ + { + "bbox": [ + 66, + 665, + 197, + 686 + ], + "spans": [ + { + "bbox": [ + 66, + 665, + 197, + 686 + ], + "score": 1.0, + "content": "4 Experimental Setups", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 69, + 693, + 290, + 773 + ], + "lines": [ + { + "bbox": [ + 80, + 693, + 290, + 706 + ], + "spans": [ + { + "bbox": [ + 80, + 693, + 290, + 706 + ], + "score": 1.0, + "content": "GPT Models. InstructGPT (text-davinci-003", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 707, + 290, + 720 + ], + "spans": [ + { + "bbox": [ + 68, + 707, + 290, + 720 + ], + "score": 1.0, + "content": "model) (Ouyang et al., 2022) is a fine-tuned series", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 69, + 720, + 290, + 732 + ], + "spans": [ + { + "bbox": [ + 69, + 720, + 290, + 732 + ], + "score": 1.0, + "content": "of GPT-3 (Brown et al., 2020) using reinforcement", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 69, + 735, + 290, + 747 + ], + "spans": [ + { + "bbox": [ + 69, + 735, + 290, + 747 + ], + "score": 1.0, + "content": "learning from human feedback (RLHF). Compared", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 69, + 748, + 290, + 760 + ], + "spans": [ + { + "bbox": [ + 69, + 748, + 290, + 760 + ], + "score": 1.0, + "content": "with InstructGPT, ChatGPT (gpt-3.5-turbo model)", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 68, + 761, + 290, + 774 + ], + "spans": [ + { + "bbox": [ + 68, + 761, + 290, + 774 + ], + "score": 1.0, + "content": "is trained on a more diverse range of internet text", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 304, + 321, + 525, + 416 + ], + "lines": [ + { + "bbox": [ + 303, + 321, + 526, + 335 + ], + "spans": [ + { + "bbox": [ + 303, + 321, + 526, + 335 + ], + "score": 1.0, + "content": "(e.g., social media, news) and can better and faster", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 336, + 527, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 527, + 348 + ], + "score": 1.0, + "content": "respond to prompts in a conversational manner.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 349, + 525, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 525, + 361 + ], + "score": 1.0, + "content": "GPT-4 (gpt-4 model) (Bubeck et al., 2023) can", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 363, + 526, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 526, + 375 + ], + "score": 1.0, + "content": "be viewed as an enhanced version of ChatGPT, and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 376, + 525, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 525, + 390 + ], + "score": 1.0, + "content": "it can solve more complex problems and support", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 303, + 388, + 525, + 404 + ], + "spans": [ + { + "bbox": [ + 303, + 388, + 525, + 404 + ], + "score": 1.0, + "content": "multi-modal chat with broader general knowledge", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 404, + 461, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 461, + 417 + ], + "score": 1.0, + "content": "and stronger reasoning capabilities.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 305, + 418, + 525, + 565 + ], + "lines": [ + { + "bbox": [ + 315, + 417, + 528, + 431 + ], + "spans": [ + { + "bbox": [ + 315, + 417, + 528, + 431 + ], + "score": 1.0, + "content": "Myers–Briggs Type Indicator. The My-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 431, + 527, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 527, + 444 + ], + "score": 1.0, + "content": "ers–Briggs Type Indicator (MBTI) (Myers and Mc-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 443, + 527, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 527, + 459 + ], + "score": 1.0, + "content": "Caulley, 1985) assesses the psychological prefer-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 459, + 526, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 526, + 471 + ], + "score": 1.0, + "content": "ences of individuals in how they perceive the world", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 471, + 526, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 526, + 486 + ], + "score": 1.0, + "content": "and make decisions via an introspective question-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 303, + 485, + 525, + 499 + ], + "spans": [ + { + "bbox": [ + 303, + 485, + 525, + 499 + ], + "score": 1.0, + "content": "naire, so as to identify different personality types", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 303, + 498, + 526, + 512 + ], + "spans": [ + { + "bbox": [ + 303, + 498, + 526, + 512 + ], + "score": 1.0, + "content": "based on five dichotomies1: (1) Extraverted versus", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 512, + 525, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 525, + 524 + ], + "score": 1.0, + "content": "Introverted (E vs. I); (2) Intuitive versus Observant", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 527, + 526, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 526, + 539 + ], + "score": 1.0, + "content": "(N vs. S); (3) Thinking versus Feeling (T vs. F); (4)", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 540, + 525, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 525, + 552 + ], + "score": 1.0, + "content": "Judging versus Prospecting (J vs. P); (5) Assertive", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "versus Turbulent (A vs. T) (see Appendix C).", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 49 + }, + { + "type": "text", + "bbox": [ + 305, + 568, + 526, + 688 + ], + "lines": [ + { + "bbox": [ + 315, + 568, + 527, + 579 + ], + "spans": [ + { + "bbox": [ + 315, + 568, + 527, + 579 + ], + "score": 1.0, + "content": "Implementation Details. The number of inde-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 304, + 581, + 524, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 487, + 594 + ], + "score": 1.0, + "content": "pendent testings for each subject is set to", + "type": "text" + }, + { + "bbox": [ + 487, + 581, + 524, + 592 + ], + "score": 0.89, + "content": "N = 1 5", + "type": "inline_equation" + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 594, + 525, + 607 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 525, + 607 + ], + "score": 1.0, + "content": "We evaluate the consistency and robustness scores", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 607, + 525, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 607, + 525, + 622 + ], + "score": 1.0, + "content": "of LLMs’ assessments on the general population", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 621, + 527, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 527, + 635 + ], + "score": 1.0, + "content": "(“People”, “Men”, “Women”) and specific profes-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 635, + 527, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 527, + 647 + ], + "score": 1.0, + "content": "sions following (Nadeem et al., 2021). The fair-", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 649, + 527, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 649, + 527, + 662 + ], + "score": 1.0, + "content": "ness score is measured based on two gender pairs,", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 304, + 663, + 527, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 663, + 527, + 676 + ], + "score": 1.0, + "content": "namely (“Men”, “Women”) and (“Boys”, “Girls”).", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 675, + 501, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 675, + 501, + 690 + ], + "score": 1.0, + "content": "More details are provided in the appendices.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 59 + }, + { + "type": "title", + "bbox": [ + 305, + 701, + 432, + 715 + ], + "lines": [ + { + "bbox": [ + 302, + 700, + 433, + 717 + ], + "spans": [ + { + "bbox": [ + 302, + 700, + 433, + 717 + ], + "score": 1.0, + "content": "5 Results and Analyses", + "type": "text" + } + ], + "index": 64 + } + ], + "index": 64 + }, + { + "type": "text", + "bbox": [ + 305, + 725, + 525, + 751 + ], + "lines": [ + { + "bbox": [ + 303, + 725, + 525, + 738 + ], + "spans": [ + { + "bbox": [ + 303, + 725, + 525, + 738 + ], + "score": 1.0, + "content": "We query ChatGPT, InstructGPT, and GPT-4 to", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 303, + 738, + 526, + 753 + ], + "spans": [ + { + "bbox": [ + 303, + 738, + 526, + 753 + ], + "score": 1.0, + "content": "assess the personalities of different subjects, and", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 65.5 + } + ], + "page_idx": 5, + "page_size": [ + 595, + 841 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 318, + 763, + 440, + 773 + ], + "lines": [ + { + "bbox": [ + 319, + 762, + 441, + 774 + ], + "spans": [ + { + "bbox": [ + 319, + 762, + 441, + 774 + ], + "score": 1.0, + "content": "1https://www.16personalities.com", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 127, + 522, + 301 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 69, + 69, + 526, + 118 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 68, + 69, + 526, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 69, + 526, + 83 + ], + "score": 1.0, + "content": "Table 1: Personality types and scores assessed by InstructGPT, ChatGPT, and GPT-4 when we query different", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 82, + 526, + 94 + ], + "spans": [ + { + "bbox": [ + 69, + 82, + 526, + 94 + ], + "score": 1.0, + "content": "subjects. The score results are averaged from multiple independent testings. We present the assessed scores of five", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 94, + 527, + 106 + ], + "spans": [ + { + "bbox": [ + 69, + 94, + 527, + 106 + ], + "score": 1.0, + "content": "dimensions that dominate the personality types. Bold indicates the same personality role assessed from all LLMs,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 104, + 515, + 120 + ], + "spans": [ + { + "bbox": [ + 68, + 104, + 515, + 120 + ], + "score": 1.0, + "content": "while the underline denotes the highest score among LLMs when obtaining the same assessed personality type.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 72, + 127, + 522, + 301 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 72, + 127, + 522, + 301 + ], + "spans": [ + { + "bbox": [ + 72, + 127, + 522, + 301 + ], + "score": 0.984, + "html": "
LLMSubjectPeopleMenWomenBarbersAccountantsDoctorsArtistsMathematiciansPoliticians
InstructGPTPersonality Types/ScoresE=64E=66E=66E=53I=53E=52E=59I= 51E=59
N= 65N= 64N= 71N= 52N= 52N= 58N= 69N= 56N= 62
T= 53T=50F= 55F= 53F=51F= 54F= 59T= 54T= 54
J= 62J= 56J= 61J= 66J= 72J= 71J= 60J= 67J= 59
Personality RoleT= 60T= 62T= 58A= 53T= 62T= 53A= 50A= 52T= 54
CommanderCommanderProtagonistProtagonistAdventurerProtagonistProtagonistArchitectCommander
E=57E= 55E= 54E=50I=56E= 54E=58I= 61E=63
ChatGPTPersonality Types /ScoresN= 60N= 52N= 51S= 51S= 59N= 52N= 67N= 54N= 50
T=51T= 52T=51T=53T=60F= 54F= 60T=64T=58
J= 57J= 54J= 53J= 56J= 68J= 64P=58J= 62J= 56
T=59T=51A= 50T=51A=50T= 56T= 64A=50T= 59
Personality CommanderCommanderCommanderExecutiveLogisticianProtagonistCampaignerArchitectCommander
GPT-4 Types /ScoresRole PersonalityE=53E=57
N= 61N= 53E= 61 N= 58E=52 N= 50I= 54 S= 55E=54 N= 51E=58 N= 67I= 61E=64
T= 54T= 55F= 58T=51T= 57F= 55F= 56N= 56 T= 64S=51 T= 57
J= 54= 56J= 57J= 56J= 68J= 66P=58J= 64J= 55
T= 68T= 63T= 61A=51A=50T= 53T= 63T = 51
Personality RoleCommanderCommanderProtagonistCommanderLogisticianProtagonistCampaignerArchitectT= 57 Executive
", + "type": "table", + "image_path": "fdade0780fe01c69da13a6ba3cc801604d663e531a77a0ec9c410ef58ab58bc8.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 72, + 127, + 522, + 185.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 72, + 185.0, + 522, + 243.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 72, + 243.0, + 522, + 301.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "image", + "bbox": [ + 69, + 318, + 289, + 455 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 69, + 318, + 289, + 455 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 69, + 318, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 318, + 289, + 455 + ], + "score": 0.97, + "type": "image", + "image_path": "b6213e178644a5b3e6213f45feb1937cc09cb815d6d310c4963cac95ee1d90ad.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 69, + 318, + 289, + 331.7 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 69, + 331.7, + 289, + 345.4 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 69, + 345.4, + 289, + 359.09999999999997 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 69, + 359.09999999999997, + 289, + 372.79999999999995 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 69, + 372.79999999999995, + 289, + 386.49999999999994 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 69, + 386.49999999999994, + 289, + 400.19999999999993 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 69, + 400.19999999999993, + 289, + 413.8999999999999 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 69, + 413.8999999999999, + 289, + 427.5999999999999 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 69, + 427.5999999999999, + 289, + 441.2999999999999 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 69, + 441.2999999999999, + 289, + 454.9999999999999 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 69, + 464, + 290, + 547 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 69, + 464, + 290, + 476 + ], + "spans": [ + { + "bbox": [ + 69, + 464, + 290, + 476 + ], + "score": 1.0, + "content": "Figure 3: The most frequent option for each question", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 68, + 475, + 291, + 489 + ], + "spans": [ + { + "bbox": [ + 68, + 475, + 291, + 489 + ], + "score": 1.0, + "content": "in multiple independent testings of InstructGPT (Left),", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 68, + 486, + 291, + 502 + ], + "spans": [ + { + "bbox": [ + 68, + 486, + 291, + 502 + ], + "score": 1.0, + "content": "ChatGPT (Middle), and GPT-4 (Right) when we query", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 69, + 500, + 290, + 511 + ], + "spans": [ + { + "bbox": [ + 69, + 500, + 290, + 511 + ], + "score": 1.0, + "content": "the subject “People” (Top row),or “Artists” (Bottom", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 68, + 510, + 291, + 524 + ], + "spans": [ + { + "bbox": [ + 68, + 510, + 291, + 524 + ], + "score": 1.0, + "content": "row). “GC”, “PC”, “NCNW”, “PW”, and “GW” denote", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 68, + 523, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 68, + 523, + 290, + 536 + ], + "score": 1.0, + "content": "“Generally correct”, “Partially correct”, “Neither correct", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 68, + 535, + 290, + 549 + ], + "spans": [ + { + "bbox": [ + 68, + 535, + 290, + 549 + ], + "score": 1.0, + "content": "nor wrong”, “Partially wrong”, and “Generally wrong”.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20 + } + ], + "index": 15.75 + }, + { + "type": "text", + "bbox": [ + 69, + 572, + 290, + 652 + ], + "lines": [ + { + "bbox": [ + 68, + 571, + 290, + 587 + ], + "spans": [ + { + "bbox": [ + 68, + 571, + 290, + 587 + ], + "score": 1.0, + "content": "Here we multiply their corresponding consistency", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 68, + 582, + 291, + 603 + ], + "spans": [ + { + "bbox": [ + 68, + 584, + 115, + 601 + ], + "score": 1.0, + "content": "scores sMc", + "type": "text" + }, + { + "bbox": [ + 113, + 582, + 155, + 603 + ], + "score": 1.0, + "content": "and s Fc", + "type": "text" + }, + { + "bbox": [ + 151, + 585, + 291, + 599 + ], + "score": 1.0, + "content": "since a higher assessment con-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 68, + 600, + 290, + 612 + ], + "spans": [ + { + "bbox": [ + 68, + 600, + 290, + 612 + ], + "score": 1.0, + "content": "sistency of subjects can contribute more to their", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 68, + 613, + 290, + 627 + ], + "spans": [ + { + "bbox": [ + 68, + 613, + 198, + 626 + ], + "score": 1.0, + "content": "inherent similarity. A larger", + "type": "text" + }, + { + "bbox": [ + 198, + 615, + 210, + 627 + ], + "score": 0.85, + "content": "s _ { f }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 613, + 290, + 626 + ], + "score": 1.0, + "content": "indicates that the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 68, + 627, + 290, + 639 + ], + "spans": [ + { + "bbox": [ + 68, + 627, + 290, + 639 + ], + "score": 1.0, + "content": "assessments on different genders are more fair with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 70, + 641, + 213, + 653 + ], + "spans": [ + { + "bbox": [ + 70, + 641, + 213, + 653 + ], + "score": 1.0, + "content": "higher consistency and less bias.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5, + "bbox_fs": [ + 68, + 571, + 291, + 653 + ] + }, + { + "type": "title", + "bbox": [ + 70, + 668, + 196, + 682 + ], + "lines": [ + { + "bbox": [ + 66, + 665, + 197, + 686 + ], + "spans": [ + { + "bbox": [ + 66, + 665, + 197, + 686 + ], + "score": 1.0, + "content": "4 Experimental Setups", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 69, + 693, + 290, + 773 + ], + "lines": [ + { + "bbox": [ + 80, + 693, + 290, + 706 + ], + "spans": [ + { + "bbox": [ + 80, + 693, + 290, + 706 + ], + "score": 1.0, + "content": "GPT Models. InstructGPT (text-davinci-003", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 707, + 290, + 720 + ], + "spans": [ + { + "bbox": [ + 68, + 707, + 290, + 720 + ], + "score": 1.0, + "content": "model) (Ouyang et al., 2022) is a fine-tuned series", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 69, + 720, + 290, + 732 + ], + "spans": [ + { + "bbox": [ + 69, + 720, + 290, + 732 + ], + "score": 1.0, + "content": "of GPT-3 (Brown et al., 2020) using reinforcement", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 69, + 735, + 290, + 747 + ], + "spans": [ + { + "bbox": [ + 69, + 735, + 290, + 747 + ], + "score": 1.0, + "content": "learning from human feedback (RLHF). Compared", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 69, + 748, + 290, + 760 + ], + "spans": [ + { + "bbox": [ + 69, + 748, + 290, + 760 + ], + "score": 1.0, + "content": "with InstructGPT, ChatGPT (gpt-3.5-turbo model)", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 68, + 761, + 290, + 774 + ], + "spans": [ + { + "bbox": [ + 68, + 761, + 290, + 774 + ], + "score": 1.0, + "content": "is trained on a more diverse range of internet text", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 303, + 321, + 526, + 335 + ], + "spans": [ + { + "bbox": [ + 303, + 321, + 526, + 335 + ], + "score": 1.0, + "content": "(e.g., social media, news) and can better and faster", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 336, + 527, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 527, + 348 + ], + "score": 1.0, + "content": "respond to prompts in a conversational manner.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 349, + 525, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 525, + 361 + ], + "score": 1.0, + "content": "GPT-4 (gpt-4 model) (Bubeck et al., 2023) can", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 363, + 526, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 526, + 375 + ], + "score": 1.0, + "content": "be viewed as an enhanced version of ChatGPT, and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 376, + 525, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 525, + 390 + ], + "score": 1.0, + "content": "it can solve more complex problems and support", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 303, + 388, + 525, + 404 + ], + "spans": [ + { + "bbox": [ + 303, + 388, + 525, + 404 + ], + "score": 1.0, + "content": "multi-modal chat with broader general knowledge", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 404, + 461, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 461, + 417 + ], + "score": 1.0, + "content": "and stronger reasoning capabilities.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 33.5, + "bbox_fs": [ + 68, + 693, + 290, + 774 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 321, + 525, + 416 + ], + "lines": [], + "index": 40, + "bbox_fs": [ + 303, + 321, + 527, + 417 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 305, + 418, + 525, + 565 + ], + "lines": [ + { + "bbox": [ + 315, + 417, + 528, + 431 + ], + "spans": [ + { + "bbox": [ + 315, + 417, + 528, + 431 + ], + "score": 1.0, + "content": "Myers–Briggs Type Indicator. The My-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 431, + 527, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 527, + 444 + ], + "score": 1.0, + "content": "ers–Briggs Type Indicator (MBTI) (Myers and Mc-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 443, + 527, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 527, + 459 + ], + "score": 1.0, + "content": "Caulley, 1985) assesses the psychological prefer-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 459, + 526, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 526, + 471 + ], + "score": 1.0, + "content": "ences of individuals in how they perceive the world", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 471, + 526, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 526, + 486 + ], + "score": 1.0, + "content": "and make decisions via an introspective question-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 303, + 485, + 525, + 499 + ], + "spans": [ + { + "bbox": [ + 303, + 485, + 525, + 499 + ], + "score": 1.0, + "content": "naire, so as to identify different personality types", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 303, + 498, + 526, + 512 + ], + "spans": [ + { + "bbox": [ + 303, + 498, + 526, + 512 + ], + "score": 1.0, + "content": "based on five dichotomies1: (1) Extraverted versus", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 512, + 525, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 525, + 524 + ], + "score": 1.0, + "content": "Introverted (E vs. I); (2) Intuitive versus Observant", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 527, + 526, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 526, + 539 + ], + "score": 1.0, + "content": "(N vs. S); (3) Thinking versus Feeling (T vs. F); (4)", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 540, + 525, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 525, + 552 + ], + "score": 1.0, + "content": "Judging versus Prospecting (J vs. P); (5) Assertive", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "versus Turbulent (A vs. T) (see Appendix C).", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 49, + "bbox_fs": [ + 303, + 417, + 528, + 566 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 568, + 526, + 688 + ], + "lines": [ + { + "bbox": [ + 315, + 568, + 527, + 579 + ], + "spans": [ + { + "bbox": [ + 315, + 568, + 527, + 579 + ], + "score": 1.0, + "content": "Implementation Details. The number of inde-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 304, + 581, + 524, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 487, + 594 + ], + "score": 1.0, + "content": "pendent testings for each subject is set to", + "type": "text" + }, + { + "bbox": [ + 487, + 581, + 524, + 592 + ], + "score": 0.89, + "content": "N = 1 5", + "type": "inline_equation" + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 594, + 525, + 607 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 525, + 607 + ], + "score": 1.0, + "content": "We evaluate the consistency and robustness scores", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 607, + 525, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 607, + 525, + 622 + ], + "score": 1.0, + "content": "of LLMs’ assessments on the general population", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 621, + 527, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 527, + 635 + ], + "score": 1.0, + "content": "(“People”, “Men”, “Women”) and specific profes-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 635, + 527, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 527, + 647 + ], + "score": 1.0, + "content": "sions following (Nadeem et al., 2021). The fair-", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 649, + 527, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 649, + 527, + 662 + ], + "score": 1.0, + "content": "ness score is measured based on two gender pairs,", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 304, + 663, + 527, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 663, + 527, + 676 + ], + "score": 1.0, + "content": "namely (“Men”, “Women”) and (“Boys”, “Girls”).", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 675, + 501, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 675, + 501, + 690 + ], + "score": 1.0, + "content": "More details are provided in the appendices.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 59, + "bbox_fs": [ + 304, + 568, + 527, + 690 + ] + }, + { + "type": "title", + "bbox": [ + 305, + 701, + 432, + 715 + ], + "lines": [ + { + "bbox": [ + 302, + 700, + 433, + 717 + ], + "spans": [ + { + "bbox": [ + 302, + 700, + 433, + 717 + ], + "score": 1.0, + "content": "5 Results and Analyses", + "type": "text" + } + ], + "index": 64 + } + ], + "index": 64 + }, + { + "type": "text", + "bbox": [ + 305, + 725, + 525, + 751 + ], + "lines": [ + { + "bbox": [ + 303, + 725, + 525, + 738 + ], + "spans": [ + { + "bbox": [ + 303, + 725, + 525, + 738 + ], + "score": 1.0, + "content": "We query ChatGPT, InstructGPT, and GPT-4 to", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 303, + 738, + 526, + 753 + ], + "spans": [ + { + "bbox": [ + 303, + 738, + 526, + 753 + ], + "score": 1.0, + "content": "assess the personalities of different subjects, and", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 65.5, + "bbox_fs": [ + 303, + 725, + 526, + 753 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 102, + 525, + 161 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 65, + 70, + 521, + 93 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 68, + 69, + 524, + 82 + ], + "spans": [ + { + "bbox": [ + 68, + 69, + 181, + 82 + ], + "score": 1.0, + "content": "Table 2: Consistency scores", + "type": "text" + }, + { + "bbox": [ + 182, + 71, + 196, + 81 + ], + "score": 0.74, + "content": "( s _ { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 69, + 285, + 82 + ], + "score": 1.0, + "content": "and robustness scores", + "type": "text" + }, + { + "bbox": [ + 285, + 71, + 304, + 81 + ], + "score": 0.46, + "content": "\\left( s _ { r } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 69, + 524, + 82 + ], + "score": 1.0, + "content": "comparison between InstructGPT, ChatGPT, and GPT-4", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 82, + 404, + 95 + ], + "spans": [ + { + "bbox": [ + 68, + 82, + 404, + 95 + ], + "score": 1.0, + "content": "in assessing different subjects. Bold shows the highest average scores among them.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 72, + 102, + 525, + 161 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 525, + 161 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 525, + 161 + ], + "score": 0.979, + "html": "
MetricLLMPeopleMenWomenBarbersAccountantsDoctorsArtistsMathematiciansPoliticiansAverage
Consistency ScoreInstructGPT ChatGPT0.9160.8880.9050.8980.9250.9010.9000.8970.9140.905
0.9070.8950.9130.9220.9320.9220.9180.9320.9190.918
GPT-40.9360.9270.9110.9090.9280.9160.9270.9220.9110.921
RobustnessInstructGPT0.9360.9240.9440.9250.9650.9360.9360.9560.9520.942
ChatGPT0.8880.9170.9600.9270.9580.9670.9400.9200.9350.935
ScoreGPT-40.9700.8930.8850.9650.9610.9800.9280.9340.9050.936
", + "type": "table", + "image_path": "65c67c771d469310a808cb95898ffcae3fd83517dea746841dab3cf8a0d423ad.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 72, + 102, + 525, + 121.66666666666667 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 72, + 121.66666666666667, + 525, + 141.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 72, + 141.33333333333334, + 525, + 161.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 73, + 223, + 288, + 258 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 69, + 179, + 290, + 215 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 69, + 179, + 291, + 191 + ], + "spans": [ + { + "bbox": [ + 69, + 179, + 171, + 191 + ], + "score": 1.0, + "content": "Table 3: Fairness scores", + "type": "text" + }, + { + "bbox": [ + 171, + 180, + 189, + 191 + ], + "score": 0.84, + "content": "( s _ { f } )", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 179, + 291, + 191 + ], + "score": 1.0, + "content": "comparison between In-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 190, + 290, + 204 + ], + "spans": [ + { + "bbox": [ + 69, + 190, + 290, + 204 + ], + "score": 1.0, + "content": "structGPT, ChatGPT, and GPT-4 in assessing different", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 203, + 290, + 216 + ], + "spans": [ + { + "bbox": [ + 69, + 203, + 290, + 216 + ], + "score": 1.0, + "content": "gender pairs. Bold indicates the highest average score.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 73, + 223, + 288, + 258 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 73, + 223, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 73, + 223, + 288, + 258 + ], + "score": 0.966, + "html": "
LLMMenvs.WomenBoys vs. GirlsAverage
InstructGPT0.7230.7830.753
ChatGPT0.7960.7560.776
GPT40.7860.7700.778
", + "type": "table", + "image_path": "d518b9ae7d6c29a6948c80a9da36ccb44e15dafc179c1739a99e567ff35cb492.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 73, + 223, + 288, + 240.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 73, + 240.5, + 288, + 258.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 7.25 + }, + { + "type": "text", + "bbox": [ + 69, + 280, + 289, + 320 + ], + "lines": [ + { + "bbox": [ + 68, + 281, + 290, + 293 + ], + "spans": [ + { + "bbox": [ + 68, + 281, + 290, + 293 + ], + "score": 1.0, + "content": "compare their assessment results in Table 1. The", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 295, + 290, + 306 + ], + "spans": [ + { + "bbox": [ + 69, + 295, + 290, + 306 + ], + "score": 1.0, + "content": "consistency, robustness, and fairness scores of their", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 68, + 309, + 255, + 319 + ], + "spans": [ + { + "bbox": [ + 68, + 309, + 255, + 319 + ], + "score": 1.0, + "content": "assessments are reported in Table 2 and 3.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 69, + 333, + 234, + 360 + ], + "lines": [ + { + "bbox": [ + 68, + 333, + 234, + 347 + ], + "spans": [ + { + "bbox": [ + 68, + 333, + 234, + 347 + ], + "score": 1.0, + "content": "5.1 Can ChatGPT Assess Human", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 93, + 348, + 161, + 360 + ], + "spans": [ + { + "bbox": [ + 93, + 348, + 161, + 360 + ], + "score": 1.0, + "content": "Personalities?", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 69, + 367, + 290, + 515 + ], + "lines": [ + { + "bbox": [ + 69, + 367, + 291, + 379 + ], + "spans": [ + { + "bbox": [ + 69, + 367, + 291, + 379 + ], + "score": 1.0, + "content": "As shown in Fig. 3, most answers and their distribu-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 68, + 381, + 291, + 393 + ], + "spans": [ + { + "bbox": [ + 68, + 381, + 291, + 393 + ], + "score": 1.0, + "content": "tions generated by three LLMs are evidently differ-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 394, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 69, + 394, + 290, + 407 + ], + "score": 1.0, + "content": "ent, which suggests that each model can be viewed", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 68, + 408, + 290, + 421 + ], + "spans": [ + { + "bbox": [ + 68, + 408, + 290, + 421 + ], + "score": 1.0, + "content": "as an individual to provide independent opinions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 68, + 421, + 290, + 434 + ], + "spans": [ + { + "bbox": [ + 68, + 421, + 290, + 434 + ], + "score": 1.0, + "content": "in assessing personalities. Notably, ChatGPT and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 69, + 435, + 290, + 448 + ], + "spans": [ + { + "bbox": [ + 69, + 435, + 290, + 448 + ], + "score": 1.0, + "content": "GPT-4 can respond to questions more flexibly (i.e.,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 69, + 449, + 290, + 461 + ], + "spans": [ + { + "bbox": [ + 69, + 449, + 290, + 461 + ], + "score": 1.0, + "content": "more diverse options and distributions) compared", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 68, + 461, + 291, + 475 + ], + "spans": [ + { + "bbox": [ + 68, + 461, + 291, + 475 + ], + "score": 1.0, + "content": "with InstructGPT. This is consistent with their prop-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 69, + 477, + 291, + 489 + ], + "spans": [ + { + "bbox": [ + 69, + 477, + 291, + 489 + ], + "score": 1.0, + "content": "erty of being trained on a a wider range of topics,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 489, + 290, + 503 + ], + "spans": [ + { + "bbox": [ + 68, + 489, + 290, + 503 + ], + "score": 1.0, + "content": "enabling them to possess stronger model capacity", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 68, + 503, + 271, + 515 + ], + "spans": [ + { + "bbox": [ + 68, + 503, + 271, + 515 + ], + "score": 1.0, + "content": "(e.g., reasoning ability) for better assessment.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 69, + 517, + 290, + 773 + ], + "lines": [ + { + "bbox": [ + 80, + 518, + 291, + 530 + ], + "spans": [ + { + "bbox": [ + 80, + 518, + 291, + 530 + ], + "score": 1.0, + "content": "Interestingly, in spite of possibly different an-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 69, + 531, + 290, + 543 + ], + "spans": [ + { + "bbox": [ + 69, + 531, + 290, + 543 + ], + "score": 1.0, + "content": "swer distributions, the average results in Table 1", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 69, + 545, + 291, + 557 + ], + "spans": [ + { + "bbox": [ + 69, + 545, + 291, + 557 + ], + "score": 1.0, + "content": "show that four subjects are assessed as the same per-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 69, + 558, + 290, + 570 + ], + "spans": [ + { + "bbox": [ + 69, + 558, + 290, + 570 + ], + "score": 1.0, + "content": "sonality types by all LLMs. This could suggest the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 69, + 571, + 290, + 585 + ], + "spans": [ + { + "bbox": [ + 69, + 571, + 290, + 585 + ], + "score": 1.0, + "content": "inherent similarity of their personality assessment", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 69, + 585, + 290, + 597 + ], + "spans": [ + { + "bbox": [ + 69, + 585, + 290, + 597 + ], + "score": 1.0, + "content": "abilities. In most of these cases, ChatGPT tends", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 597, + 290, + 613 + ], + "spans": [ + { + "bbox": [ + 68, + 597, + 290, + 613 + ], + "score": 1.0, + "content": "to achieve medium personality scores, implying", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 69, + 612, + 290, + 624 + ], + "spans": [ + { + "bbox": [ + 69, + 612, + 290, + 624 + ], + "score": 1.0, + "content": "its more neutral assessment compared with other", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 69, + 626, + 291, + 639 + ], + "spans": [ + { + "bbox": [ + 69, + 626, + 291, + 639 + ], + "score": 1.0, + "content": "two LLMs. It is worth noting that some assess-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 69, + 640, + 290, + 651 + ], + "spans": [ + { + "bbox": [ + 69, + 640, + 290, + 651 + ], + "score": 1.0, + "content": "ment results from ChatGPT and GPT-4 are close", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 69, + 653, + 290, + 665 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 290, + 665 + ], + "score": 1.0, + "content": "to our intuition: (1) Accountants are assessed as", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 68, + 666, + 290, + 680 + ], + "spans": [ + { + "bbox": [ + 68, + 666, + 290, + 680 + ], + "score": 1.0, + "content": "“Logistician” that is usually a reliable, practical", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 69, + 680, + 291, + 692 + ], + "spans": [ + { + "bbox": [ + 69, + 680, + 291, + 692 + ], + "score": 1.0, + "content": "and fact-minded individual. (2) Artists are classi-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 68, + 693, + 292, + 707 + ], + "spans": [ + { + "bbox": [ + 68, + 693, + 292, + 707 + ], + "score": 1.0, + "content": "fied as the type “ENFP-T” that often possesses cre-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 69, + 707, + 290, + 720 + ], + "spans": [ + { + "bbox": [ + 69, + 707, + 290, + 720 + ], + "score": 1.0, + "content": "ative and enthusiastic spirits. (3) Mathematicians", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 68, + 721, + 289, + 732 + ], + "spans": [ + { + "bbox": [ + 68, + 721, + 289, + 732 + ], + "score": 1.0, + "content": "are assessed to be the personality role \"Architect\"", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 69, + 734, + 290, + 747 + ], + "spans": [ + { + "bbox": [ + 69, + 734, + 290, + 747 + ], + "score": 1.0, + "content": "that are thinkers with profound ideas and strategic", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 69, + 748, + 291, + 760 + ], + "spans": [ + { + "bbox": [ + 69, + 748, + 291, + 760 + ], + "score": 1.0, + "content": "plans. To a certain extent, these results demon-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 69, + 761, + 291, + 774 + ], + "spans": [ + { + "bbox": [ + 69, + 761, + 291, + 774 + ], + "score": 1.0, + "content": "strate their effectiveness on human personality as-", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 305, + 181, + 525, + 342 + ], + "lines": [ + { + "bbox": [ + 304, + 181, + 527, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 527, + 194 + ], + "score": 1.0, + "content": "sessment. Moreover, it is observed that “People”", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 193, + 527, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 527, + 208 + ], + "score": 1.0, + "content": "and “Men” are classified as leader roles (“Com-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 209, + 526, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 526, + 221 + ], + "score": 1.0, + "content": "mander”) by all LLMs. We speculate that it is a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 222, + 527, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 527, + 234 + ], + "score": 1.0, + "content": "result of the human-centered fine-tuning (e.g., rein-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 235, + 527, + 249 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 527, + 249 + ], + "score": 1.0, + "content": "forcement learning from human feedback (RLHF)),", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 249, + 526, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 526, + 262 + ], + "score": 1.0, + "content": "which encourages LLMs to follow the prevailing", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 303, + 262, + 527, + 275 + ], + "spans": [ + { + "bbox": [ + 303, + 262, + 527, + 275 + ], + "score": 1.0, + "content": "positive societal conceptions and values such as", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 276, + 527, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 527, + 288 + ], + "score": 1.0, + "content": "the expected relations between human and LLMs.", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 289, + 525, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 525, + 302 + ], + "score": 1.0, + "content": "In this context, the assessed personality scores in", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 303, + 302, + 526, + 316 + ], + "spans": [ + { + "bbox": [ + 303, + 302, + 526, + 316 + ], + "score": 1.0, + "content": "Table 1 can shed more insights on “how LLMs view", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 304, + 316, + 525, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 525, + 330 + ], + "score": 1.0, + "content": "humans” and serve as an indicator to better develop", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 304, + 331, + 511, + 343 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 511, + 343 + ], + "score": 1.0, + "content": "human-centered and socially-beneficial LLMs.", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 50.5 + }, + { + "type": "title", + "bbox": [ + 304, + 386, + 523, + 411 + ], + "lines": [ + { + "bbox": [ + 303, + 384, + 524, + 399 + ], + "spans": [ + { + "bbox": [ + 303, + 384, + 524, + 399 + ], + "score": 1.0, + "content": "5.2 Is the Assessment Consistent, Robust and", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 327, + 399, + 357, + 412 + ], + "spans": [ + { + "bbox": [ + 327, + 399, + 357, + 412 + ], + "score": 1.0, + "content": "Fair?", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 57.5 + }, + { + "type": "text", + "bbox": [ + 305, + 436, + 525, + 772 + ], + "lines": [ + { + "bbox": [ + 304, + 435, + 525, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 525, + 448 + ], + "score": 1.0, + "content": "As shown in Table 2, ChatGPT and GPT-4 achieve", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 450, + 525, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 525, + 462 + ], + "score": 1.0, + "content": "higher consistency scores than InstructGPT in most", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 303, + 462, + 527, + 478 + ], + "spans": [ + { + "bbox": [ + 303, + 462, + 527, + 478 + ], + "score": 1.0, + "content": "cases when assessing different subjects. This sug-", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 304, + 477, + 526, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 526, + 490 + ], + "score": 1.0, + "content": "gests that ChatGPT and GPT-4 can provide more", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 490, + 527, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 527, + 503 + ], + "score": 1.0, + "content": "similar and consistent personality assessment re-", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 504, + 527, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 527, + 516 + ], + "score": 1.0, + "content": "sults under multiple independent testings. How-", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 518, + 526, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 526, + 530 + ], + "score": 1.0, + "content": "ever, their average robustness scores are slightly", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 531, + 525, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 525, + 543 + ], + "score": 1.0, + "content": "lower than that of InstructGPT, which indicates", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 544, + 525, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 525, + 557 + ], + "score": 1.0, + "content": "that their assessments could be more sensitive to", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 558, + 527, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 527, + 571 + ], + "score": 1.0, + "content": "the prompt biases (e.g., changes of option orders).", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 571, + 527, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 527, + 583 + ], + "score": 1.0, + "content": "This might lead to their more diverse answer distri-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 585, + 525, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 525, + 598 + ], + "score": 1.0, + "content": "butions in different testings as shown in Fig. 3. It", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 599, + 527, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 527, + 611 + ], + "score": 1.0, + "content": "actually verifies the necessity of the proposed unbi-", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 303, + 612, + 525, + 625 + ], + "spans": [ + { + "bbox": [ + 303, + 612, + 525, + 625 + ], + "score": 1.0, + "content": "ased prompts and the averaging of testing results", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 304, + 626, + 527, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 527, + 639 + ], + "score": 1.0, + "content": "to encourage more impartial assessments. As pre-", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 639, + 526, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 526, + 652 + ], + "score": 1.0, + "content": "sented in Table 3, ChatGPT and GPT-4 show higher", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 303, + 653, + 527, + 666 + ], + "spans": [ + { + "bbox": [ + 303, + 653, + 527, + 666 + ], + "score": 1.0, + "content": "average fairness scores than InstructGPT when as-", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 303, + 666, + 526, + 680 + ], + "spans": [ + { + "bbox": [ + 303, + 666, + 526, + 680 + ], + "score": 1.0, + "content": "sessing different genders. This indicates that they", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 680, + 525, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 680, + 525, + 692 + ], + "score": 1.0, + "content": "are more likely to equally assess subjects with less", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 304, + 694, + 525, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 694, + 525, + 707 + ], + "score": 1.0, + "content": "gender bias, which is consistent with the finding", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "score": 1.0, + "content": "of (Zhuo et al., 2023). In summary, although the", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 720, + 527, + 732 + ], + "spans": [ + { + "bbox": [ + 304, + 720, + 527, + 732 + ], + "score": 1.0, + "content": "assessments of ChatGPT and GPT-4 can be influ-", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 735, + 525, + 746 + ], + "spans": [ + { + "bbox": [ + 304, + 735, + 525, + 746 + ], + "score": 1.0, + "content": "enced by random input perturbations, their overall", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 304, + 748, + 526, + 760 + ], + "spans": [ + { + "bbox": [ + 304, + 748, + 526, + 760 + ], + "score": 1.0, + "content": "assessment results are more consistent and fairer", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 304, + 762, + 430, + 773 + ], + "spans": [ + { + "bbox": [ + 304, + 762, + 430, + 773 + ], + "score": 1.0, + "content": "compared with InstructGPT.", + "type": "text" + } + ], + "index": 83 + } + ], + "index": 71 + } + ], + "page_idx": 6, + "page_size": [ + 595, + 841 + ], + "discarded_blocks": [], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 102, + 525, + 161 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 65, + 70, + 521, + 93 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 68, + 69, + 524, + 82 + ], + "spans": [ + { + "bbox": [ + 68, + 69, + 181, + 82 + ], + "score": 1.0, + "content": "Table 2: Consistency scores", + "type": "text" + }, + { + "bbox": [ + 182, + 71, + 196, + 81 + ], + "score": 0.74, + "content": "( s _ { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 69, + 285, + 82 + ], + "score": 1.0, + "content": "and robustness scores", + "type": "text" + }, + { + "bbox": [ + 285, + 71, + 304, + 81 + ], + "score": 0.46, + "content": "\\left( s _ { r } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 69, + 524, + 82 + ], + "score": 1.0, + "content": "comparison between InstructGPT, ChatGPT, and GPT-4", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 82, + 404, + 95 + ], + "spans": [ + { + "bbox": [ + 68, + 82, + 404, + 95 + ], + "score": 1.0, + "content": "in assessing different subjects. Bold shows the highest average scores among them.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 72, + 102, + 525, + 161 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 525, + 161 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 525, + 161 + ], + "score": 0.979, + "html": "
MetricLLMPeopleMenWomenBarbersAccountantsDoctorsArtistsMathematiciansPoliticiansAverage
Consistency ScoreInstructGPT ChatGPT0.9160.8880.9050.8980.9250.9010.9000.8970.9140.905
0.9070.8950.9130.9220.9320.9220.9180.9320.9190.918
GPT-40.9360.9270.9110.9090.9280.9160.9270.9220.9110.921
RobustnessInstructGPT0.9360.9240.9440.9250.9650.9360.9360.9560.9520.942
ChatGPT0.8880.9170.9600.9270.9580.9670.9400.9200.9350.935
ScoreGPT-40.9700.8930.8850.9650.9610.9800.9280.9340.9050.936
", + "type": "table", + "image_path": "65c67c771d469310a808cb95898ffcae3fd83517dea746841dab3cf8a0d423ad.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 72, + 102, + 525, + 121.66666666666667 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 72, + 121.66666666666667, + 525, + 141.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 72, + 141.33333333333334, + 525, + 161.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 73, + 223, + 288, + 258 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 69, + 179, + 290, + 215 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 69, + 179, + 291, + 191 + ], + "spans": [ + { + "bbox": [ + 69, + 179, + 171, + 191 + ], + "score": 1.0, + "content": "Table 3: Fairness scores", + "type": "text" + }, + { + "bbox": [ + 171, + 180, + 189, + 191 + ], + "score": 0.84, + "content": "( s _ { f } )", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 179, + 291, + 191 + ], + "score": 1.0, + "content": "comparison between In-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 190, + 290, + 204 + ], + "spans": [ + { + "bbox": [ + 69, + 190, + 290, + 204 + ], + "score": 1.0, + "content": "structGPT, ChatGPT, and GPT-4 in assessing different", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 203, + 290, + 216 + ], + "spans": [ + { + "bbox": [ + 69, + 203, + 290, + 216 + ], + "score": 1.0, + "content": "gender pairs. Bold indicates the highest average score.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 73, + 223, + 288, + 258 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 73, + 223, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 73, + 223, + 288, + 258 + ], + "score": 0.966, + "html": "
LLMMenvs.WomenBoys vs. GirlsAverage
InstructGPT0.7230.7830.753
ChatGPT0.7960.7560.776
GPT40.7860.7700.778
", + "type": "table", + "image_path": "d518b9ae7d6c29a6948c80a9da36ccb44e15dafc179c1739a99e567ff35cb492.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 73, + 223, + 288, + 240.5 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 73, + 240.5, + 288, + 258.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 7.25 + }, + { + "type": "text", + "bbox": [ + 69, + 280, + 289, + 320 + ], + "lines": [ + { + "bbox": [ + 68, + 281, + 290, + 293 + ], + "spans": [ + { + "bbox": [ + 68, + 281, + 290, + 293 + ], + "score": 1.0, + "content": "compare their assessment results in Table 1. The", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 295, + 290, + 306 + ], + "spans": [ + { + "bbox": [ + 69, + 295, + 290, + 306 + ], + "score": 1.0, + "content": "consistency, robustness, and fairness scores of their", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 68, + 309, + 255, + 319 + ], + "spans": [ + { + "bbox": [ + 68, + 309, + 255, + 319 + ], + "score": 1.0, + "content": "assessments are reported in Table 2 and 3.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 68, + 281, + 290, + 319 + ] + }, + { + "type": "title", + "bbox": [ + 69, + 333, + 234, + 360 + ], + "lines": [ + { + "bbox": [ + 68, + 333, + 234, + 347 + ], + "spans": [ + { + "bbox": [ + 68, + 333, + 234, + 347 + ], + "score": 1.0, + "content": "5.1 Can ChatGPT Assess Human", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 93, + 348, + 161, + 360 + ], + "spans": [ + { + "bbox": [ + 93, + 348, + 161, + 360 + ], + "score": 1.0, + "content": "Personalities?", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 69, + 367, + 290, + 515 + ], + "lines": [ + { + "bbox": [ + 69, + 367, + 291, + 379 + ], + "spans": [ + { + "bbox": [ + 69, + 367, + 291, + 379 + ], + "score": 1.0, + "content": "As shown in Fig. 3, most answers and their distribu-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 68, + 381, + 291, + 393 + ], + "spans": [ + { + "bbox": [ + 68, + 381, + 291, + 393 + ], + "score": 1.0, + "content": "tions generated by three LLMs are evidently differ-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 394, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 69, + 394, + 290, + 407 + ], + "score": 1.0, + "content": "ent, which suggests that each model can be viewed", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 68, + 408, + 290, + 421 + ], + "spans": [ + { + "bbox": [ + 68, + 408, + 290, + 421 + ], + "score": 1.0, + "content": "as an individual to provide independent opinions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 68, + 421, + 290, + 434 + ], + "spans": [ + { + "bbox": [ + 68, + 421, + 290, + 434 + ], + "score": 1.0, + "content": "in assessing personalities. Notably, ChatGPT and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 69, + 435, + 290, + 448 + ], + "spans": [ + { + "bbox": [ + 69, + 435, + 290, + 448 + ], + "score": 1.0, + "content": "GPT-4 can respond to questions more flexibly (i.e.,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 69, + 449, + 290, + 461 + ], + "spans": [ + { + "bbox": [ + 69, + 449, + 290, + 461 + ], + "score": 1.0, + "content": "more diverse options and distributions) compared", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 68, + 461, + 291, + 475 + ], + "spans": [ + { + "bbox": [ + 68, + 461, + 291, + 475 + ], + "score": 1.0, + "content": "with InstructGPT. This is consistent with their prop-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 69, + 477, + 291, + 489 + ], + "spans": [ + { + "bbox": [ + 69, + 477, + 291, + 489 + ], + "score": 1.0, + "content": "erty of being trained on a a wider range of topics,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 489, + 290, + 503 + ], + "spans": [ + { + "bbox": [ + 68, + 489, + 290, + 503 + ], + "score": 1.0, + "content": "enabling them to possess stronger model capacity", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 68, + 503, + 271, + 515 + ], + "spans": [ + { + "bbox": [ + 68, + 503, + 271, + 515 + ], + "score": 1.0, + "content": "(e.g., reasoning ability) for better assessment.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20, + "bbox_fs": [ + 68, + 367, + 291, + 515 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 517, + 290, + 773 + ], + "lines": [ + { + "bbox": [ + 80, + 518, + 291, + 530 + ], + "spans": [ + { + "bbox": [ + 80, + 518, + 291, + 530 + ], + "score": 1.0, + "content": "Interestingly, in spite of possibly different an-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 69, + 531, + 290, + 543 + ], + "spans": [ + { + "bbox": [ + 69, + 531, + 290, + 543 + ], + "score": 1.0, + "content": "swer distributions, the average results in Table 1", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 69, + 545, + 291, + 557 + ], + "spans": [ + { + "bbox": [ + 69, + 545, + 291, + 557 + ], + "score": 1.0, + "content": "show that four subjects are assessed as the same per-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 69, + 558, + 290, + 570 + ], + "spans": [ + { + "bbox": [ + 69, + 558, + 290, + 570 + ], + "score": 1.0, + "content": "sonality types by all LLMs. This could suggest the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 69, + 571, + 290, + 585 + ], + "spans": [ + { + "bbox": [ + 69, + 571, + 290, + 585 + ], + "score": 1.0, + "content": "inherent similarity of their personality assessment", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 69, + 585, + 290, + 597 + ], + "spans": [ + { + "bbox": [ + 69, + 585, + 290, + 597 + ], + "score": 1.0, + "content": "abilities. In most of these cases, ChatGPT tends", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 597, + 290, + 613 + ], + "spans": [ + { + "bbox": [ + 68, + 597, + 290, + 613 + ], + "score": 1.0, + "content": "to achieve medium personality scores, implying", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 69, + 612, + 290, + 624 + ], + "spans": [ + { + "bbox": [ + 69, + 612, + 290, + 624 + ], + "score": 1.0, + "content": "its more neutral assessment compared with other", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 69, + 626, + 291, + 639 + ], + "spans": [ + { + "bbox": [ + 69, + 626, + 291, + 639 + ], + "score": 1.0, + "content": "two LLMs. It is worth noting that some assess-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 69, + 640, + 290, + 651 + ], + "spans": [ + { + "bbox": [ + 69, + 640, + 290, + 651 + ], + "score": 1.0, + "content": "ment results from ChatGPT and GPT-4 are close", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 69, + 653, + 290, + 665 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 290, + 665 + ], + "score": 1.0, + "content": "to our intuition: (1) Accountants are assessed as", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 68, + 666, + 290, + 680 + ], + "spans": [ + { + "bbox": [ + 68, + 666, + 290, + 680 + ], + "score": 1.0, + "content": "“Logistician” that is usually a reliable, practical", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 69, + 680, + 291, + 692 + ], + "spans": [ + { + "bbox": [ + 69, + 680, + 291, + 692 + ], + "score": 1.0, + "content": "and fact-minded individual. (2) Artists are classi-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 68, + 693, + 292, + 707 + ], + "spans": [ + { + "bbox": [ + 68, + 693, + 292, + 707 + ], + "score": 1.0, + "content": "fied as the type “ENFP-T” that often possesses cre-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 69, + 707, + 290, + 720 + ], + "spans": [ + { + "bbox": [ + 69, + 707, + 290, + 720 + ], + "score": 1.0, + "content": "ative and enthusiastic spirits. (3) Mathematicians", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 68, + 721, + 289, + 732 + ], + "spans": [ + { + "bbox": [ + 68, + 721, + 289, + 732 + ], + "score": 1.0, + "content": "are assessed to be the personality role \"Architect\"", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 69, + 734, + 290, + 747 + ], + "spans": [ + { + "bbox": [ + 69, + 734, + 290, + 747 + ], + "score": 1.0, + "content": "that are thinkers with profound ideas and strategic", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 69, + 748, + 291, + 760 + ], + "spans": [ + { + "bbox": [ + 69, + 748, + 291, + 760 + ], + "score": 1.0, + "content": "plans. To a certain extent, these results demon-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 69, + 761, + 291, + 774 + ], + "spans": [ + { + "bbox": [ + 69, + 761, + 291, + 774 + ], + "score": 1.0, + "content": "strate their effectiveness on human personality as-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 181, + 527, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 527, + 194 + ], + "score": 1.0, + "content": "sessment. Moreover, it is observed that “People”", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 193, + 527, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 527, + 208 + ], + "score": 1.0, + "content": "and “Men” are classified as leader roles (“Com-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 209, + 526, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 526, + 221 + ], + "score": 1.0, + "content": "mander”) by all LLMs. We speculate that it is a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 222, + 527, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 527, + 234 + ], + "score": 1.0, + "content": "result of the human-centered fine-tuning (e.g., rein-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 235, + 527, + 249 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 527, + 249 + ], + "score": 1.0, + "content": "forcement learning from human feedback (RLHF)),", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 249, + 526, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 526, + 262 + ], + "score": 1.0, + "content": "which encourages LLMs to follow the prevailing", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 303, + 262, + 527, + 275 + ], + "spans": [ + { + "bbox": [ + 303, + 262, + 527, + 275 + ], + "score": 1.0, + "content": "positive societal conceptions and values such as", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 276, + 527, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 527, + 288 + ], + "score": 1.0, + "content": "the expected relations between human and LLMs.", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 289, + 525, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 525, + 302 + ], + "score": 1.0, + "content": "In this context, the assessed personality scores in", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 303, + 302, + 526, + 316 + ], + "spans": [ + { + "bbox": [ + 303, + 302, + 526, + 316 + ], + "score": 1.0, + "content": "Table 1 can shed more insights on “how LLMs view", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 304, + 316, + 525, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 525, + 330 + ], + "score": 1.0, + "content": "humans” and serve as an indicator to better develop", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 304, + 331, + 511, + 343 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 511, + 343 + ], + "score": 1.0, + "content": "human-centered and socially-beneficial LLMs.", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 35, + "bbox_fs": [ + 68, + 518, + 292, + 774 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 181, + 525, + 342 + ], + "lines": [], + "index": 50.5, + "bbox_fs": [ + 303, + 181, + 527, + 343 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 304, + 386, + 523, + 411 + ], + "lines": [ + { + "bbox": [ + 303, + 384, + 524, + 399 + ], + "spans": [ + { + "bbox": [ + 303, + 384, + 524, + 399 + ], + "score": 1.0, + "content": "5.2 Is the Assessment Consistent, Robust and", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 327, + 399, + 357, + 412 + ], + "spans": [ + { + "bbox": [ + 327, + 399, + 357, + 412 + ], + "score": 1.0, + "content": "Fair?", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 57.5 + }, + { + "type": "text", + "bbox": [ + 305, + 436, + 525, + 772 + ], + "lines": [ + { + "bbox": [ + 304, + 435, + 525, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 525, + 448 + ], + "score": 1.0, + "content": "As shown in Table 2, ChatGPT and GPT-4 achieve", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 450, + 525, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 525, + 462 + ], + "score": 1.0, + "content": "higher consistency scores than InstructGPT in most", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 303, + 462, + 527, + 478 + ], + "spans": [ + { + "bbox": [ + 303, + 462, + 527, + 478 + ], + "score": 1.0, + "content": "cases when assessing different subjects. This sug-", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 304, + 477, + 526, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 526, + 490 + ], + "score": 1.0, + "content": "gests that ChatGPT and GPT-4 can provide more", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 490, + 527, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 527, + 503 + ], + "score": 1.0, + "content": "similar and consistent personality assessment re-", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 504, + 527, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 527, + 516 + ], + "score": 1.0, + "content": "sults under multiple independent testings. How-", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 518, + 526, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 526, + 530 + ], + "score": 1.0, + "content": "ever, their average robustness scores are slightly", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 531, + 525, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 525, + 543 + ], + "score": 1.0, + "content": "lower than that of InstructGPT, which indicates", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 544, + 525, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 525, + 557 + ], + "score": 1.0, + "content": "that their assessments could be more sensitive to", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 558, + 527, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 527, + 571 + ], + "score": 1.0, + "content": "the prompt biases (e.g., changes of option orders).", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 571, + 527, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 527, + 583 + ], + "score": 1.0, + "content": "This might lead to their more diverse answer distri-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 585, + 525, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 525, + 598 + ], + "score": 1.0, + "content": "butions in different testings as shown in Fig. 3. It", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 599, + 527, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 527, + 611 + ], + "score": 1.0, + "content": "actually verifies the necessity of the proposed unbi-", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 303, + 612, + 525, + 625 + ], + "spans": [ + { + "bbox": [ + 303, + 612, + 525, + 625 + ], + "score": 1.0, + "content": "ased prompts and the averaging of testing results", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 304, + 626, + 527, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 527, + 639 + ], + "score": 1.0, + "content": "to encourage more impartial assessments. As pre-", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 639, + 526, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 526, + 652 + ], + "score": 1.0, + "content": "sented in Table 3, ChatGPT and GPT-4 show higher", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 303, + 653, + 527, + 666 + ], + "spans": [ + { + "bbox": [ + 303, + 653, + 527, + 666 + ], + "score": 1.0, + "content": "average fairness scores than InstructGPT when as-", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 303, + 666, + 526, + 680 + ], + "spans": [ + { + "bbox": [ + 303, + 666, + 526, + 680 + ], + "score": 1.0, + "content": "sessing different genders. This indicates that they", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 680, + 525, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 680, + 525, + 692 + ], + "score": 1.0, + "content": "are more likely to equally assess subjects with less", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 304, + 694, + 525, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 694, + 525, + 707 + ], + "score": 1.0, + "content": "gender bias, which is consistent with the finding", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 707, + 525, + 720 + ], + "score": 1.0, + "content": "of (Zhuo et al., 2023). In summary, although the", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 720, + 527, + 732 + ], + "spans": [ + { + "bbox": [ + 304, + 720, + 527, + 732 + ], + "score": 1.0, + "content": "assessments of ChatGPT and GPT-4 can be influ-", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 735, + 525, + 746 + ], + "spans": [ + { + "bbox": [ + 304, + 735, + 525, + 746 + ], + "score": 1.0, + "content": "enced by random input perturbations, their overall", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 304, + 748, + 526, + 760 + ], + "spans": [ + { + "bbox": [ + 304, + 748, + 526, + 760 + ], + "score": 1.0, + "content": "assessment results are more consistent and fairer", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 304, + 762, + 430, + 773 + ], + "spans": [ + { + "bbox": [ + 304, + 762, + 430, + 773 + ], + "score": 1.0, + "content": "compared with InstructGPT.", + "type": "text" + } + ], + "index": 83 + } + ], + "index": 71, + "bbox_fs": [ + 303, + 435, + 527, + 773 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 127, + 524, + 205 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 69, + 69, + 525, + 118 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 68, + 69, + 526, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 69, + 526, + 83 + ], + "score": 1.0, + "content": "Table 4: Personality types and roles assessed by ChatGPT and GPT-4 when we query subjects with different", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 82, + 525, + 94 + ], + "spans": [ + { + "bbox": [ + 69, + 82, + 525, + 94 + ], + "score": 1.0, + "content": "income levels (low, middle, high), age levels (children, adolescents, adults, old adults) or different education levels", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 93, + 525, + 106 + ], + "spans": [ + { + "bbox": [ + 68, + 93, + 525, + 106 + ], + "score": 1.0, + "content": "(junior/middle/high school students, undergraduate/master/PhD students). The results are averaged from multiple", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 105, + 443, + 119 + ], + "spans": [ + { + "bbox": [ + 68, + 105, + 443, + 119 + ], + "score": 1.0, + "content": "independent testings. Bold indicates the same personality types/role assessed from all LLMs.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 73, + 127, + 524, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 73, + 127, + 524, + 205 + ], + "spans": [ + { + "bbox": [ + 73, + 127, + 524, + 205 + ], + "score": 0.955, + "html": "
LLMBackgroundIncome LevelAgeLevelevelEduEducation Level
LowMiddleHighChildrenAdolescentsAdultsOld AdultsJuniorMiddleHighUndergraduateMasterPhD
ChatGPTPersonalityTypesINFJ-TENFJ-TENTJ-TENFP-TENFP-TENTJ-TINFJ-TESFP-TENFP-TENFJ-TENFJ-TINTJ-TINTJ-T
PersonalityRoleAdvocateProtagonistCommanderCampaignerCampaignerCommanderAdvocateEntertainerCampaignerProtagonistProtagonistArchitectArchitect
GPT-4PersonalityTypesENFJ-TENFJ-TENTJ-TENFP-TENFP-TENTJ-TENFJ-TENTP-TENTP-TENTP-TENTJ-TENTJ-TENTJ-T
PersonalityRoleProtagonistProtagonistCommanderCampaignerCampaignerCommanderProtagonistDebaterDebaterDebaterCommanderCommanderCommander
", + "type": "table", + "image_path": "e78235f99159ec64fcfd0c1cc222d028fbb3b2a312e320608fbdb73ebc269c05.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 73, + 127, + 524, + 153.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 73, + 153.0, + 524, + 179.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 73, + 179.0, + 524, + 205.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "image", + "bbox": [ + 69, + 221, + 289, + 291 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 69, + 221, + 289, + 291 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 69, + 221, + 289, + 291 + ], + "spans": [ + { + "bbox": [ + 69, + 221, + 289, + 291 + ], + "score": 0.966, + "type": "image", + "image_path": "0698c6cc3c08f83b1e348e1389720d8e0c9bf0861ee79cec3eafd1ef6e0fade0.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 69, + 221, + 289, + 235.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 69, + 235.0, + 289, + 249.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 69, + 249.0, + 289, + 263.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 69, + 263.0, + 289, + 277.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 69, + 277.0, + 289, + 291.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 69, + 299, + 290, + 359 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 69, + 299, + 290, + 312 + ], + "spans": [ + { + "bbox": [ + 69, + 299, + 290, + 312 + ], + "score": 1.0, + "content": "Figure 4: The most frequent option for each question", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 311, + 291, + 324 + ], + "spans": [ + { + "bbox": [ + 69, + 311, + 291, + 324 + ], + "score": 1.0, + "content": "in multiple independent testings of InstructGPT (Left),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 68, + 323, + 291, + 336 + ], + "spans": [ + { + "bbox": [ + 68, + 323, + 291, + 336 + ], + "score": 1.0, + "content": "ChatGPT (Middle), GPT-4 (Right) when we query the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 335, + 292, + 348 + ], + "spans": [ + { + "bbox": [ + 69, + 335, + 292, + 348 + ], + "score": 1.0, + "content": "subject “Artists” without using unbiased prompts. “W”", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 347, + 290, + 360 + ], + "spans": [ + { + "bbox": [ + 69, + 347, + 290, + 360 + ], + "score": 1.0, + "content": "denotes “Wrong”, and other legends are same as Fig. 3.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "image_caption", + "bbox": [ + 304, + 313, + 525, + 337 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 526, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 526, + 326 + ], + "score": 1.0, + "content": "Figure 6: An example of uncertain answers generated", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 324, + 514, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 514, + 338 + ], + "score": 1.0, + "content": "from ChatGPT when querying a specific individual.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51.5 + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 71, + 371, + 288, + 453 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 71, + 371, + 288, + 453 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 71, + 371, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 71, + 371, + 288, + 453 + ], + "score": 0.967, + "type": "image", + "image_path": "9c53b84ba9700f4fae31b7a9847d5a97174bed4015856baf914842bd4cf356fb.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 71, + 371, + 288, + 384.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 71, + 384.6666666666667, + 288, + 398.33333333333337 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 71, + 398.33333333333337, + 288, + 412.00000000000006 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 71, + 412.00000000000006, + 288, + 425.66666666666674 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 71, + 425.66666666666674, + 288, + 439.3333333333334 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 71, + 439.3333333333334, + 288, + 453.0000000000001 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 69, + 460, + 290, + 496 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 68, + 460, + 290, + 473 + ], + "spans": [ + { + "bbox": [ + 68, + 460, + 290, + 473 + ], + "score": 1.0, + "content": "Figure 5: Personality scores of different subjects in five", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 69, + 472, + 290, + 484 + ], + "spans": [ + { + "bbox": [ + 69, + 472, + 290, + 484 + ], + "score": 1.0, + "content": "dimensions of MBTI results assessed from InstructGPT", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 69, + 484, + 262, + 497 + ], + "spans": [ + { + "bbox": [ + 69, + 484, + 262, + 497 + ], + "score": 1.0, + "content": "(Blue), ChatGPT (Orange), and GPT-4 (Green).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + } + ], + "index": 21.75 + }, + { + "type": "title", + "bbox": [ + 69, + 519, + 147, + 532 + ], + "lines": [ + { + "bbox": [ + 68, + 516, + 149, + 535 + ], + "spans": [ + { + "bbox": [ + 68, + 516, + 149, + 535 + ], + "score": 1.0, + "content": "6 Discussions", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 69, + 543, + 290, + 664 + ], + "lines": [ + { + "bbox": [ + 79, + 541, + 290, + 557 + ], + "spans": [ + { + "bbox": [ + 79, + 541, + 290, + 557 + ], + "score": 1.0, + "content": "Effects of Unbiased Prompts. Fig. 4 shows", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 69, + 557, + 290, + 570 + ], + "spans": [ + { + "bbox": [ + 69, + 557, + 290, + 570 + ], + "score": 1.0, + "content": "that using the same-order options leads to a higher", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 69, + 570, + 291, + 583 + ], + "spans": [ + { + "bbox": [ + 69, + 570, + 291, + 583 + ], + "score": 1.0, + "content": "frequency of the same option (i.e., more fixed an-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 69, + 584, + 291, + 597 + ], + "spans": [ + { + "bbox": [ + 69, + 584, + 291, + 597 + ], + "score": 1.0, + "content": "swers) for many questions compared with employ-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 69, + 597, + 291, + 611 + ], + "spans": [ + { + "bbox": [ + 69, + 597, + 291, + 611 + ], + "score": 1.0, + "content": "ing unbiased prompts (see Fig. 3). This suggests", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 610, + 291, + 624 + ], + "spans": [ + { + "bbox": [ + 68, + 610, + 291, + 624 + ], + "score": 1.0, + "content": "the effectiveness and necessity of the proposed un-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 68, + 624, + 292, + 639 + ], + "spans": [ + { + "bbox": [ + 68, + 624, + 292, + 639 + ], + "score": 1.0, + "content": "biased prompts, which introduce random perturba-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 68, + 638, + 290, + 653 + ], + "spans": [ + { + "bbox": [ + 68, + 638, + 290, + 653 + ], + "score": 1.0, + "content": "tions into question inputs and average all testing", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 69, + 653, + 280, + 665 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 280, + 665 + ], + "score": 1.0, + "content": "results to encourage more impartial assessment.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 69, + 666, + 290, + 773 + ], + "lines": [ + { + "bbox": [ + 80, + 666, + 289, + 679 + ], + "spans": [ + { + "bbox": [ + 80, + 666, + 289, + 679 + ], + "score": 1.0, + "content": "Effects of Background Prompts. We show the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 69, + 680, + 291, + 693 + ], + "spans": [ + { + "bbox": [ + 69, + 680, + 291, + 693 + ], + "score": 1.0, + "content": "effects of background prompts on LLM’s assess-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 69, + 694, + 290, + 706 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 290, + 706 + ], + "score": 1.0, + "content": "ments by adding different income, age or education", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 69, + 707, + 291, + 720 + ], + "spans": [ + { + "bbox": [ + 69, + 707, + 291, + 720 + ], + "score": 1.0, + "content": "information of the subject. As shown in Table 4,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 67, + 719, + 291, + 735 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 291, + 735 + ], + "score": 1.0, + "content": "“Middle-income people” is assessed as the type", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 67, + 732, + 291, + 748 + ], + "spans": [ + { + "bbox": [ + 67, + 732, + 291, + 748 + ], + "score": 1.0, + "content": "“ENFJ-T” that is slightly different from the type", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 67, + 746, + 291, + 761 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 291, + 761 + ], + "score": 1.0, + "content": "“ENTJ-T” of “People”. Interestingly, high educa-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "spans": [ + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "score": 1.0, + "content": "tion level subjects such as “Master” and “PhD” are", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 342, + 226, + 489, + 258 + ], + "lines": [ + { + "bbox": [ + 342, + 227, + 488, + 237 + ], + "spans": [ + { + "bbox": [ + 342, + 227, + 488, + 237 + ], + "score": 0.98, + "content": "Statement: Taylor Swift spends a lot of her", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 342, + 237, + 487, + 247 + ], + "spans": [ + { + "bbox": [ + 342, + 237, + 487, + 247 + ], + "score": 0.99, + "content": "free time exploring various random topics", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 342, + 248, + 422, + 258 + ], + "spans": [ + { + "bbox": [ + 342, + 248, + 422, + 258 + ], + "score": 0.996, + "content": "that pique her interests.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 342, + 258, + 488, + 300 + ], + "lines": [ + { + "bbox": [ + 342, + 257, + 487, + 268 + ], + "spans": [ + { + "bbox": [ + 342, + 257, + 487, + 268 + ], + "score": 0.978, + "content": "Answer: It is neither correct nor wrong.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 342, + 268, + 488, + 279 + ], + "spans": [ + { + "bbox": [ + 342, + 268, + 488, + 279 + ], + "score": 0.992, + "content": "Without more information on her personal", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 342, + 279, + 487, + 289 + ], + "spans": [ + { + "bbox": [ + 342, + 279, + 487, + 289 + ], + "score": 0.964, + "content": "lifeand interests,it is difficult to determine", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 342, + 288, + 468, + 300 + ], + "spans": [ + { + "bbox": [ + 342, + 288, + 468, + 300 + ], + "score": 0.95, + "content": "the full accuracy of the statement...", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5 + }, + { + "type": "text", + "bbox": [ + 305, + 359, + 525, + 467 + ], + "lines": [ + { + "bbox": [ + 303, + 358, + 527, + 374 + ], + "spans": [ + { + "bbox": [ + 303, + 358, + 527, + 374 + ], + "score": 1.0, + "content": "assessed as the “INTJ-T” or “ENTJ-T” type that", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 373, + 526, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 526, + 386 + ], + "score": 1.0, + "content": "often possesses strategic plans, profound ideas or", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 304, + 387, + 526, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 526, + 399 + ], + "score": 1.0, + "content": "rational minds, while junior/middle school students", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 304, + 401, + 525, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 525, + 414 + ], + "score": 1.0, + "content": "are classified to the types that are usually energetic", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 303, + 414, + 525, + 426 + ], + "spans": [ + { + "bbox": [ + 303, + 414, + 525, + 426 + ], + "score": 1.0, + "content": "or curious. This implies that ChatGPT and GPT-4", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 428, + 525, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 525, + 440 + ], + "score": 1.0, + "content": "may be able to to understand different backgrounds", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 440, + 526, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 526, + 455 + ], + "score": 1.0, + "content": "of subjects, and an appropriate background prompt", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 454, + 516, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 516, + 468 + ], + "score": 1.0, + "content": "could facilitate reliable personality assessments.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 56.5 + }, + { + "type": "text", + "bbox": [ + 305, + 469, + 525, + 549 + ], + "lines": [ + { + "bbox": [ + 315, + 467, + 527, + 483 + ], + "spans": [ + { + "bbox": [ + 315, + 467, + 527, + 483 + ], + "score": 1.0, + "content": "Visualization of Different Assessments. Fig.", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 304, + 482, + 526, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 526, + 495 + ], + "score": 1.0, + "content": "5 visualizes three subjects with different assessed", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 496, + 526, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 526, + 510 + ], + "score": 1.0, + "content": "types or scores. ChatGPT and GPT-4 achieve very", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 510, + 526, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 526, + 522 + ], + "score": 1.0, + "content": "close scores in each dimension despite different", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 523, + 525, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 525, + 536 + ], + "score": 1.0, + "content": "assessed types, which demonstrates their higher", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 305, + 537, + 502, + 550 + ], + "spans": [ + { + "bbox": [ + 305, + 537, + 502, + 550 + ], + "score": 1.0, + "content": "similarity in personality assessment abilities.", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 63.5 + }, + { + "type": "text", + "bbox": [ + 305, + 551, + 526, + 685 + ], + "lines": [ + { + "bbox": [ + 315, + 549, + 526, + 565 + ], + "spans": [ + { + "bbox": [ + 315, + 549, + 526, + 565 + ], + "score": 1.0, + "content": "Assessment of Specific Individuals. Querying", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 564, + 525, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 525, + 578 + ], + "score": 1.0, + "content": "LLMs about the personality of a certain person", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 577, + 527, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 527, + 591 + ], + "score": 1.0, + "content": "might generate uncertain answers due to the in-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 591, + 528, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 528, + 605 + ], + "score": 1.0, + "content": "sufficiency of personal backgrounds (e.g., behav-", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 605, + 527, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 527, + 618 + ], + "score": 1.0, + "content": "ior patterns) in its knowledge base (see Fig. 6).", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 617, + 527, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 527, + 632 + ], + "score": 1.0, + "content": "Considering the effects of background prompts,", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 303, + 632, + 526, + 645 + ], + "spans": [ + { + "bbox": [ + 303, + 632, + 526, + 645 + ], + "score": 1.0, + "content": "providing richer background information through", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 303, + 644, + 526, + 659 + ], + "spans": [ + { + "bbox": [ + 303, + 644, + 526, + 659 + ], + "score": 1.0, + "content": "subject-specific prompts or fine-tuning can help", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 659, + 525, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 659, + 525, + 671 + ], + "score": 1.0, + "content": "achieve a more reliable assessment. More results", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 305, + 673, + 489, + 686 + ], + "spans": [ + { + "bbox": [ + 305, + 673, + 489, + 686 + ], + "score": 1.0, + "content": "and analyses are provided in Appendix B.", + "type": "text" + } + ], + "index": 76 + } + ], + "index": 71.5 + }, + { + "type": "title", + "bbox": [ + 305, + 697, + 381, + 711 + ], + "lines": [ + { + "bbox": [ + 302, + 694, + 383, + 713 + ], + "spans": [ + { + "bbox": [ + 302, + 694, + 383, + 713 + ], + "score": 1.0, + "content": "7 Conclusion", + "type": "text" + } + ], + "index": 77 + } + ], + "index": 77 + }, + { + "type": "text", + "bbox": [ + 304, + 721, + 525, + 774 + ], + "lines": [ + { + "bbox": [ + 304, + 720, + 527, + 733 + ], + "spans": [ + { + "bbox": [ + 304, + 720, + 527, + 733 + ], + "score": 1.0, + "content": "This paper proposes a general evaluation frame-", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 303, + 734, + 526, + 748 + ], + "spans": [ + { + "bbox": [ + 303, + 734, + 526, + 748 + ], + "score": 1.0, + "content": "work for LLMs to assess human personalities via", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 303, + 748, + 525, + 760 + ], + "spans": [ + { + "bbox": [ + 303, + 748, + 525, + 760 + ], + "score": 1.0, + "content": "MBTI. We devise unbiased prompts to encourage", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 761, + 525, + 775 + ], + "spans": [ + { + "bbox": [ + 304, + 761, + 525, + 775 + ], + "score": 1.0, + "content": "LLMs to generate more impartial answers. The", + "type": "text" + } + ], + "index": 81 + } + ], + "index": 79.5 + } + ], + "page_idx": 7, + "page_size": [ + 595, + 841 + ], + "discarded_blocks": [], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 127, + 524, + 205 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 69, + 69, + 525, + 118 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 68, + 69, + 526, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 69, + 526, + 83 + ], + "score": 1.0, + "content": "Table 4: Personality types and roles assessed by ChatGPT and GPT-4 when we query subjects with different", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 82, + 525, + 94 + ], + "spans": [ + { + "bbox": [ + 69, + 82, + 525, + 94 + ], + "score": 1.0, + "content": "income levels (low, middle, high), age levels (children, adolescents, adults, old adults) or different education levels", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 93, + 525, + 106 + ], + "spans": [ + { + "bbox": [ + 68, + 93, + 525, + 106 + ], + "score": 1.0, + "content": "(junior/middle/high school students, undergraduate/master/PhD students). The results are averaged from multiple", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 105, + 443, + 119 + ], + "spans": [ + { + "bbox": [ + 68, + 105, + 443, + 119 + ], + "score": 1.0, + "content": "independent testings. Bold indicates the same personality types/role assessed from all LLMs.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 73, + 127, + 524, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 73, + 127, + 524, + 205 + ], + "spans": [ + { + "bbox": [ + 73, + 127, + 524, + 205 + ], + "score": 0.955, + "html": "
LLMBackgroundIncome LevelAgeLevelevelEduEducation Level
LowMiddleHighChildrenAdolescentsAdultsOld AdultsJuniorMiddleHighUndergraduateMasterPhD
ChatGPTPersonalityTypesINFJ-TENFJ-TENTJ-TENFP-TENFP-TENTJ-TINFJ-TESFP-TENFP-TENFJ-TENFJ-TINTJ-TINTJ-T
PersonalityRoleAdvocateProtagonistCommanderCampaignerCampaignerCommanderAdvocateEntertainerCampaignerProtagonistProtagonistArchitectArchitect
GPT-4PersonalityTypesENFJ-TENFJ-TENTJ-TENFP-TENFP-TENTJ-TENFJ-TENTP-TENTP-TENTP-TENTJ-TENTJ-TENTJ-T
PersonalityRoleProtagonistProtagonistCommanderCampaignerCampaignerCommanderProtagonistDebaterDebaterDebaterCommanderCommanderCommander
", + "type": "table", + "image_path": "e78235f99159ec64fcfd0c1cc222d028fbb3b2a312e320608fbdb73ebc269c05.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 73, + 127, + 524, + 153.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 73, + 153.0, + 524, + 179.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 73, + 179.0, + 524, + 205.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "image", + "bbox": [ + 69, + 221, + 289, + 291 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 69, + 221, + 289, + 291 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 69, + 221, + 289, + 291 + ], + "spans": [ + { + "bbox": [ + 69, + 221, + 289, + 291 + ], + "score": 0.966, + "type": "image", + "image_path": "0698c6cc3c08f83b1e348e1389720d8e0c9bf0861ee79cec3eafd1ef6e0fade0.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 69, + 221, + 289, + 235.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 69, + 235.0, + 289, + 249.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 69, + 249.0, + 289, + 263.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 69, + 263.0, + 289, + 277.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 69, + 277.0, + 289, + 291.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 69, + 299, + 290, + 359 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 69, + 299, + 290, + 312 + ], + "spans": [ + { + "bbox": [ + 69, + 299, + 290, + 312 + ], + "score": 1.0, + "content": "Figure 4: The most frequent option for each question", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 311, + 291, + 324 + ], + "spans": [ + { + "bbox": [ + 69, + 311, + 291, + 324 + ], + "score": 1.0, + "content": "in multiple independent testings of InstructGPT (Left),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 68, + 323, + 291, + 336 + ], + "spans": [ + { + "bbox": [ + 68, + 323, + 291, + 336 + ], + "score": 1.0, + "content": "ChatGPT (Middle), GPT-4 (Right) when we query the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 335, + 292, + 348 + ], + "spans": [ + { + "bbox": [ + 69, + 335, + 292, + 348 + ], + "score": 1.0, + "content": "subject “Artists” without using unbiased prompts. “W”", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 347, + 290, + 360 + ], + "spans": [ + { + "bbox": [ + 69, + 347, + 290, + 360 + ], + "score": 1.0, + "content": "denotes “Wrong”, and other legends are same as Fig. 3.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "image_caption", + "bbox": [ + 304, + 313, + 525, + 337 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 526, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 526, + 326 + ], + "score": 1.0, + "content": "Figure 6: An example of uncertain answers generated", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 324, + 514, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 514, + 338 + ], + "score": 1.0, + "content": "from ChatGPT when querying a specific individual.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51.5 + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 71, + 371, + 288, + 453 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 71, + 371, + 288, + 453 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 71, + 371, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 71, + 371, + 288, + 453 + ], + "score": 0.967, + "type": "image", + "image_path": "9c53b84ba9700f4fae31b7a9847d5a97174bed4015856baf914842bd4cf356fb.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 71, + 371, + 288, + 384.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 71, + 384.6666666666667, + 288, + 398.33333333333337 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 71, + 398.33333333333337, + 288, + 412.00000000000006 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 71, + 412.00000000000006, + 288, + 425.66666666666674 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 71, + 425.66666666666674, + 288, + 439.3333333333334 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 71, + 439.3333333333334, + 288, + 453.0000000000001 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 69, + 460, + 290, + 496 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 68, + 460, + 290, + 473 + ], + "spans": [ + { + "bbox": [ + 68, + 460, + 290, + 473 + ], + "score": 1.0, + "content": "Figure 5: Personality scores of different subjects in five", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 69, + 472, + 290, + 484 + ], + "spans": [ + { + "bbox": [ + 69, + 472, + 290, + 484 + ], + "score": 1.0, + "content": "dimensions of MBTI results assessed from InstructGPT", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 69, + 484, + 262, + 497 + ], + "spans": [ + { + "bbox": [ + 69, + 484, + 262, + 497 + ], + "score": 1.0, + "content": "(Blue), ChatGPT (Orange), and GPT-4 (Green).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + } + ], + "index": 21.75 + }, + { + "type": "title", + "bbox": [ + 69, + 519, + 147, + 532 + ], + "lines": [ + { + "bbox": [ + 68, + 516, + 149, + 535 + ], + "spans": [ + { + "bbox": [ + 68, + 516, + 149, + 535 + ], + "score": 1.0, + "content": "6 Discussions", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 69, + 543, + 290, + 664 + ], + "lines": [ + { + "bbox": [ + 79, + 541, + 290, + 557 + ], + "spans": [ + { + "bbox": [ + 79, + 541, + 290, + 557 + ], + "score": 1.0, + "content": "Effects of Unbiased Prompts. Fig. 4 shows", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 69, + 557, + 290, + 570 + ], + "spans": [ + { + "bbox": [ + 69, + 557, + 290, + 570 + ], + "score": 1.0, + "content": "that using the same-order options leads to a higher", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 69, + 570, + 291, + 583 + ], + "spans": [ + { + "bbox": [ + 69, + 570, + 291, + 583 + ], + "score": 1.0, + "content": "frequency of the same option (i.e., more fixed an-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 69, + 584, + 291, + 597 + ], + "spans": [ + { + "bbox": [ + 69, + 584, + 291, + 597 + ], + "score": 1.0, + "content": "swers) for many questions compared with employ-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 69, + 597, + 291, + 611 + ], + "spans": [ + { + "bbox": [ + 69, + 597, + 291, + 611 + ], + "score": 1.0, + "content": "ing unbiased prompts (see Fig. 3). This suggests", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 610, + 291, + 624 + ], + "spans": [ + { + "bbox": [ + 68, + 610, + 291, + 624 + ], + "score": 1.0, + "content": "the effectiveness and necessity of the proposed un-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 68, + 624, + 292, + 639 + ], + "spans": [ + { + "bbox": [ + 68, + 624, + 292, + 639 + ], + "score": 1.0, + "content": "biased prompts, which introduce random perturba-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 68, + 638, + 290, + 653 + ], + "spans": [ + { + "bbox": [ + 68, + 638, + 290, + 653 + ], + "score": 1.0, + "content": "tions into question inputs and average all testing", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 69, + 653, + 280, + 665 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 280, + 665 + ], + "score": 1.0, + "content": "results to encourage more impartial assessment.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31, + "bbox_fs": [ + 68, + 541, + 292, + 665 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 666, + 290, + 773 + ], + "lines": [ + { + "bbox": [ + 80, + 666, + 289, + 679 + ], + "spans": [ + { + "bbox": [ + 80, + 666, + 289, + 679 + ], + "score": 1.0, + "content": "Effects of Background Prompts. We show the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 69, + 680, + 291, + 693 + ], + "spans": [ + { + "bbox": [ + 69, + 680, + 291, + 693 + ], + "score": 1.0, + "content": "effects of background prompts on LLM’s assess-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 69, + 694, + 290, + 706 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 290, + 706 + ], + "score": 1.0, + "content": "ments by adding different income, age or education", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 69, + 707, + 291, + 720 + ], + "spans": [ + { + "bbox": [ + 69, + 707, + 291, + 720 + ], + "score": 1.0, + "content": "information of the subject. As shown in Table 4,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 67, + 719, + 291, + 735 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 291, + 735 + ], + "score": 1.0, + "content": "“Middle-income people” is assessed as the type", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 67, + 732, + 291, + 748 + ], + "spans": [ + { + "bbox": [ + 67, + 732, + 291, + 748 + ], + "score": 1.0, + "content": "“ENFJ-T” that is slightly different from the type", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 67, + 746, + 291, + 761 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 291, + 761 + ], + "score": 1.0, + "content": "“ENTJ-T” of “People”. Interestingly, high educa-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "spans": [ + { + "bbox": [ + 69, + 761, + 290, + 774 + ], + "score": 1.0, + "content": "tion level subjects such as “Master” and “PhD” are", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39.5, + "bbox_fs": [ + 67, + 666, + 291, + 774 + ] + }, + { + "type": "text", + "bbox": [ + 342, + 226, + 489, + 258 + ], + "lines": [ + { + "bbox": [ + 342, + 227, + 488, + 237 + ], + "spans": [ + { + "bbox": [ + 342, + 227, + 488, + 237 + ], + "score": 0.98, + "content": "Statement: Taylor Swift spends a lot of her", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 342, + 237, + 487, + 247 + ], + "spans": [ + { + "bbox": [ + 342, + 237, + 487, + 247 + ], + "score": 0.99, + "content": "free time exploring various random topics", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 342, + 248, + 422, + 258 + ], + "spans": [ + { + "bbox": [ + 342, + 248, + 422, + 258 + ], + "score": 0.996, + "content": "that pique her interests.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 342, + 227, + 488, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 342, + 258, + 488, + 300 + ], + "lines": [ + { + "bbox": [ + 342, + 257, + 487, + 268 + ], + "spans": [ + { + "bbox": [ + 342, + 257, + 487, + 268 + ], + "score": 0.978, + "content": "Answer: It is neither correct nor wrong.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 342, + 268, + 488, + 279 + ], + "spans": [ + { + "bbox": [ + 342, + 268, + 488, + 279 + ], + "score": 0.992, + "content": "Without more information on her personal", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 342, + 279, + 487, + 289 + ], + "spans": [ + { + "bbox": [ + 342, + 279, + 487, + 289 + ], + "score": 0.964, + "content": "lifeand interests,it is difficult to determine", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 342, + 288, + 468, + 300 + ], + "spans": [ + { + "bbox": [ + 342, + 288, + 468, + 300 + ], + "score": 0.95, + "content": "the full accuracy of the statement...", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5, + "bbox_fs": [ + 342, + 257, + 488, + 300 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 359, + 525, + 467 + ], + "lines": [ + { + "bbox": [ + 303, + 358, + 527, + 374 + ], + "spans": [ + { + "bbox": [ + 303, + 358, + 527, + 374 + ], + "score": 1.0, + "content": "assessed as the “INTJ-T” or “ENTJ-T” type that", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 373, + 526, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 526, + 386 + ], + "score": 1.0, + "content": "often possesses strategic plans, profound ideas or", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 304, + 387, + 526, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 526, + 399 + ], + "score": 1.0, + "content": "rational minds, while junior/middle school students", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 304, + 401, + 525, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 525, + 414 + ], + "score": 1.0, + "content": "are classified to the types that are usually energetic", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 303, + 414, + 525, + 426 + ], + "spans": [ + { + "bbox": [ + 303, + 414, + 525, + 426 + ], + "score": 1.0, + "content": "or curious. This implies that ChatGPT and GPT-4", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 428, + 525, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 525, + 440 + ], + "score": 1.0, + "content": "may be able to to understand different backgrounds", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 440, + 526, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 526, + 455 + ], + "score": 1.0, + "content": "of subjects, and an appropriate background prompt", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 454, + 516, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 516, + 468 + ], + "score": 1.0, + "content": "could facilitate reliable personality assessments.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 56.5, + "bbox_fs": [ + 303, + 358, + 527, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 469, + 525, + 549 + ], + "lines": [ + { + "bbox": [ + 315, + 467, + 527, + 483 + ], + "spans": [ + { + "bbox": [ + 315, + 467, + 527, + 483 + ], + "score": 1.0, + "content": "Visualization of Different Assessments. Fig.", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 304, + 482, + 526, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 526, + 495 + ], + "score": 1.0, + "content": "5 visualizes three subjects with different assessed", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 496, + 526, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 526, + 510 + ], + "score": 1.0, + "content": "types or scores. ChatGPT and GPT-4 achieve very", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 510, + 526, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 526, + 522 + ], + "score": 1.0, + "content": "close scores in each dimension despite different", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 523, + 525, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 525, + 536 + ], + "score": 1.0, + "content": "assessed types, which demonstrates their higher", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 305, + 537, + 502, + 550 + ], + "spans": [ + { + "bbox": [ + 305, + 537, + 502, + 550 + ], + "score": 1.0, + "content": "similarity in personality assessment abilities.", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 63.5, + "bbox_fs": [ + 304, + 467, + 527, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 551, + 526, + 685 + ], + "lines": [ + { + "bbox": [ + 315, + 549, + 526, + 565 + ], + "spans": [ + { + "bbox": [ + 315, + 549, + 526, + 565 + ], + "score": 1.0, + "content": "Assessment of Specific Individuals. Querying", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 564, + 525, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 525, + 578 + ], + "score": 1.0, + "content": "LLMs about the personality of a certain person", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 577, + 527, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 527, + 591 + ], + "score": 1.0, + "content": "might generate uncertain answers due to the in-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 591, + 528, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 528, + 605 + ], + "score": 1.0, + "content": "sufficiency of personal backgrounds (e.g., behav-", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 605, + 527, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 527, + 618 + ], + "score": 1.0, + "content": "ior patterns) in its knowledge base (see Fig. 6).", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 617, + 527, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 527, + 632 + ], + "score": 1.0, + "content": "Considering the effects of background prompts,", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 303, + 632, + 526, + 645 + ], + "spans": [ + { + "bbox": [ + 303, + 632, + 526, + 645 + ], + "score": 1.0, + "content": "providing richer background information through", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 303, + 644, + 526, + 659 + ], + "spans": [ + { + "bbox": [ + 303, + 644, + 526, + 659 + ], + "score": 1.0, + "content": "subject-specific prompts or fine-tuning can help", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 659, + 525, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 659, + 525, + 671 + ], + "score": 1.0, + "content": "achieve a more reliable assessment. More results", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 305, + 673, + 489, + 686 + ], + "spans": [ + { + "bbox": [ + 305, + 673, + 489, + 686 + ], + "score": 1.0, + "content": "and analyses are provided in Appendix B.", + "type": "text" + } + ], + "index": 76 + } + ], + "index": 71.5, + "bbox_fs": [ + 303, + 549, + 528, + 686 + ] + }, + { + "type": "title", + "bbox": [ + 305, + 697, + 381, + 711 + ], + "lines": [ + { + "bbox": [ + 302, + 694, + 383, + 713 + ], + "spans": [ + { + "bbox": [ + 302, + 694, + 383, + 713 + ], + "score": 1.0, + "content": "7 Conclusion", + "type": "text" + } + ], + "index": 77 + } + ], + "index": 77 + }, + { + "type": "text", + "bbox": [ + 304, + 721, + 525, + 774 + ], + "lines": [ + { + "bbox": [ + 304, + 720, + 527, + 733 + ], + "spans": [ + { + "bbox": [ + 304, + 720, + 527, + 733 + ], + "score": 1.0, + "content": "This paper proposes a general evaluation frame-", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 303, + 734, + 526, + 748 + ], + "spans": [ + { + "bbox": [ + 303, + 734, + 526, + 748 + ], + "score": 1.0, + "content": "work for LLMs to assess human personalities via", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 303, + 748, + 525, + 760 + ], + "spans": [ + { + "bbox": [ + 303, + 748, + 525, + 760 + ], + "score": 1.0, + "content": "MBTI. We devise unbiased prompts to encourage", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 761, + 525, + 775 + ], + "spans": [ + { + "bbox": [ + 304, + 761, + 525, + 775 + ], + "score": 1.0, + "content": "LLMs to generate more impartial answers. The", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 69, + 71, + 290, + 86 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 290, + 86 + ], + "score": 1.0, + "content": "subject-replaced query is proposed to flexibly query", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 85, + 291, + 99 + ], + "spans": [ + { + "bbox": [ + 68, + 85, + 291, + 99 + ], + "score": 1.0, + "content": "personalities of different people. We further con-", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 100, + 290, + 110 + ], + "spans": [ + { + "bbox": [ + 69, + 100, + 290, + 110 + ], + "score": 1.0, + "content": "struct correctness-evaluated instructions to enable", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 113, + 291, + 125 + ], + "spans": [ + { + "bbox": [ + 69, + 113, + 291, + 125 + ], + "score": 1.0, + "content": "clearer LLM responses. We evaluate LLMs’ consis-", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 127, + 291, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 127, + 291, + 139 + ], + "score": 1.0, + "content": "tency, robustness, and fairness in personality assess-", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "spans": [ + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "score": 1.0, + "content": "ments, and demonstrate the higher consistency and", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 154, + 290, + 165 + ], + "spans": [ + { + "bbox": [ + 70, + 154, + 290, + 165 + ], + "score": 1.0, + "content": "fairness of ChatGPT and GPT-4 than InstructGPT.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 79.5, + "bbox_fs": [ + 303, + 720, + 527, + 775 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 70, + 72, + 290, + 165 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 290, + 86 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 290, + 86 + ], + "score": 1.0, + "content": "subject-replaced query is proposed to flexibly query", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 85, + 291, + 99 + ], + "spans": [ + { + "bbox": [ + 68, + 85, + 291, + 99 + ], + "score": 1.0, + "content": "personalities of different people. We further con-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 100, + 290, + 110 + ], + "spans": [ + { + "bbox": [ + 69, + 100, + 290, + 110 + ], + "score": 1.0, + "content": "struct correctness-evaluated instructions to enable", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 113, + 291, + 125 + ], + "spans": [ + { + "bbox": [ + 69, + 113, + 291, + 125 + ], + "score": 1.0, + "content": "clearer LLM responses. We evaluate LLMs’ consis-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 127, + 291, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 127, + 291, + 139 + ], + "score": 1.0, + "content": "tency, robustness, and fairness in personality assess-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "spans": [ + { + "bbox": [ + 69, + 140, + 290, + 153 + ], + "score": 1.0, + "content": "ments, and demonstrate the higher consistency and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 154, + 290, + 165 + ], + "spans": [ + { + "bbox": [ + 70, + 154, + 290, + 165 + ], + "score": 1.0, + "content": "fairness of ChatGPT and GPT-4 than InstructGPT.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 70, + 176, + 187, + 189 + ], + "lines": [ + { + "bbox": [ + 67, + 174, + 189, + 192 + ], + "spans": [ + { + "bbox": [ + 67, + 174, + 189, + 192 + ], + "score": 1.0, + "content": "8 Acknowledgements", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 69, + 198, + 290, + 250 + ], + "lines": [ + { + "bbox": [ + 69, + 198, + 290, + 210 + ], + "spans": [ + { + "bbox": [ + 69, + 198, + 290, + 210 + ], + "score": 1.0, + "content": "This research is supported by the National Research", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 211, + 291, + 225 + ], + "spans": [ + { + "bbox": [ + 68, + 211, + 291, + 225 + ], + "score": 1.0, + "content": "Foundation, Singapore under its AI Singapore Pro-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 225, + 292, + 236 + ], + "spans": [ + { + "bbox": [ + 68, + 225, + 292, + 236 + ], + "score": 1.0, + "content": "gramme (AISG Award No: AISG2-PhD/2022-01-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 237, + 109, + 252 + ], + "spans": [ + { + "bbox": [ + 68, + 237, + 109, + 252 + ], + "score": 1.0, + "content": "034[T]).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 70, + 261, + 129, + 275 + ], + "lines": [ + { + "bbox": [ + 69, + 261, + 131, + 276 + ], + "spans": [ + { + "bbox": [ + 69, + 261, + 131, + 276 + ], + "score": 1.0, + "content": "Limitations", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 69, + 280, + 290, + 703 + ], + "lines": [ + { + "bbox": [ + 68, + 282, + 290, + 297 + ], + "spans": [ + { + "bbox": [ + 68, + 282, + 290, + 297 + ], + "score": 1.0, + "content": "While our study is a step toward the promising", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 68, + 296, + 291, + 311 + ], + "spans": [ + { + "bbox": [ + 68, + 296, + 291, + 311 + ], + "score": 1.0, + "content": "open direction of LLM-based human personality", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 310, + 291, + 324 + ], + "spans": [ + { + "bbox": [ + 69, + 310, + 291, + 324 + ], + "score": 1.0, + "content": "and psychology assessment, it possesses limitations", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 324, + 291, + 337 + ], + "spans": [ + { + "bbox": [ + 69, + 324, + 291, + 337 + ], + "score": 1.0, + "content": "and opportunities when applied to the real world.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 338, + 290, + 349 + ], + "spans": [ + { + "bbox": [ + 69, + 338, + 290, + 349 + ], + "score": 1.0, + "content": "First, our work focuses on ChatGPT model series", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 69, + 351, + 290, + 363 + ], + "spans": [ + { + "bbox": [ + 69, + 351, + 290, + 363 + ], + "score": 1.0, + "content": "and the experiments are conducted on a limited", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 68, + 364, + 290, + 377 + ], + "spans": [ + { + "bbox": [ + 68, + 364, + 290, + 377 + ], + "score": 1.0, + "content": "number of LLMs. Our framework is also scalable", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 68, + 378, + 290, + 390 + ], + "spans": [ + { + "bbox": [ + 68, + 378, + 290, + 390 + ], + "score": 1.0, + "content": "to be applied to other LLMs such as LLaMA, while", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 69, + 392, + 291, + 405 + ], + "spans": [ + { + "bbox": [ + 69, + 392, + 291, + 405 + ], + "score": 1.0, + "content": "its performance remains to be further explored. Sec-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 69, + 406, + 290, + 418 + ], + "spans": [ + { + "bbox": [ + 69, + 406, + 290, + 418 + ], + "score": 1.0, + "content": "ond, although most independent testings of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 68, + 418, + 291, + 432 + ], + "spans": [ + { + "bbox": [ + 68, + 418, + 291, + 432 + ], + "score": 1.0, + "content": "LLM under the same standard setting yield sim-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 431, + 291, + 446 + ], + "spans": [ + { + "bbox": [ + 68, + 431, + 291, + 446 + ], + "score": 1.0, + "content": "ilar assessments, the experimental setting (e.g.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 68, + 446, + 290, + 458 + ], + "spans": [ + { + "bbox": [ + 68, + 446, + 290, + 458 + ], + "score": 1.0, + "content": "hyper-parameters) or testing number can be further", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 68, + 459, + 290, + 472 + ], + "spans": [ + { + "bbox": [ + 68, + 459, + 290, + 472 + ], + "score": 1.0, + "content": "customized to test the reliability of LLMs under", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 68, + 471, + 291, + 488 + ], + "spans": [ + { + "bbox": [ + 68, + 471, + 291, + 488 + ], + "score": 1.0, + "content": "extreme cases. We will leverage the upcoming", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 68, + 486, + 291, + 500 + ], + "spans": [ + { + "bbox": [ + 68, + 486, + 291, + 500 + ], + "score": 1.0, + "content": "API that supports controllable hyper-parameters to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 68, + 498, + 292, + 514 + ], + "spans": [ + { + "bbox": [ + 68, + 498, + 292, + 514 + ], + "score": 1.0, + "content": "better evaluate GPT models. Third, the represen-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 69, + 513, + 291, + 527 + ], + "spans": [ + { + "bbox": [ + 69, + 513, + 291, + 527 + ], + "score": 1.0, + "content": "tations of different genders might be insufficient.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 68, + 527, + 292, + 540 + ], + "spans": [ + { + "bbox": [ + 68, + 527, + 292, + 540 + ], + "score": 1.0, + "content": "For example, the subjects “Ladies” and “Gentle-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 69, + 541, + 291, + 554 + ], + "spans": [ + { + "bbox": [ + 69, + 541, + 291, + 554 + ], + "score": 1.0, + "content": "men” also have different genders, while they can", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 68, + 554, + 291, + 567 + ], + "spans": [ + { + "bbox": [ + 68, + 554, + 291, + 567 + ], + "score": 1.0, + "content": "be viewed as groups that differ from “Men” and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 67, + 566, + 291, + 581 + ], + "spans": [ + { + "bbox": [ + 67, + 566, + 291, + 581 + ], + "score": 1.0, + "content": "“Women”. As the focus of this work is to devise", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 68, + 581, + 290, + 594 + ], + "spans": [ + { + "bbox": [ + 68, + 581, + 290, + 594 + ], + "score": 1.0, + "content": "a general evaluation framework, we will further", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 69, + 595, + 291, + 608 + ], + "spans": [ + { + "bbox": [ + 69, + 595, + 291, + 608 + ], + "score": 1.0, + "content": "explore the assessment of more diverse subjects in", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 68, + 607, + 291, + 622 + ], + "spans": [ + { + "bbox": [ + 68, + 607, + 291, + 622 + ], + "score": 1.0, + "content": "future works. Last, despite the popularity of MBTI", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 68, + 621, + 291, + 635 + ], + "spans": [ + { + "bbox": [ + 68, + 621, + 291, + 635 + ], + "score": 1.0, + "content": "in different areas, its scientific validity is still under", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 68, + 635, + 291, + 649 + ], + "spans": [ + { + "bbox": [ + 68, + 635, + 291, + 649 + ], + "score": 1.0, + "content": "exploration. In our work, MBTI is adopted as a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 68, + 649, + 290, + 662 + ], + "spans": [ + { + "bbox": [ + 68, + 649, + 290, + 662 + ], + "score": 1.0, + "content": "representative personality measure to help LLMs", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 69, + 662, + 291, + 676 + ], + "spans": [ + { + "bbox": [ + 69, + 662, + 291, + 676 + ], + "score": 1.0, + "content": "conduct quantitative evaluations. We will explore", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 68, + 676, + 290, + 689 + ], + "spans": [ + { + "bbox": [ + 68, + 676, + 290, + 689 + ], + "score": 1.0, + "content": "other tests such as Big Five Inventory (BFI) (John", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 69, + 690, + 259, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 690, + 259, + 703 + ], + "score": 1.0, + "content": "et al., 1999) under our scalable framework.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 28 + }, + { + "type": "title", + "bbox": [ + 70, + 712, + 183, + 726 + ], + "lines": [ + { + "bbox": [ + 68, + 711, + 184, + 727 + ], + "spans": [ + { + "bbox": [ + 68, + 711, + 184, + 727 + ], + "score": 1.0, + "content": "Ethics Considerations", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 69, + 734, + 289, + 774 + ], + "lines": [ + { + "bbox": [ + 68, + 732, + 291, + 748 + ], + "spans": [ + { + "bbox": [ + 68, + 732, + 291, + 748 + ], + "score": 1.0, + "content": "Misuse Potential. Due to the exploratory nature", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 68, + 747, + 291, + 761 + ], + "spans": [ + { + "bbox": [ + 68, + 747, + 291, + 761 + ], + "score": 1.0, + "content": "of our study, one should not directly use, generalize", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 68, + 761, + 290, + 774 + ], + "spans": [ + { + "bbox": [ + 68, + 761, + 290, + 774 + ], + "score": 1.0, + "content": "or match the assessment results (e.g., personality", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 305, + 72, + 525, + 166 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 527, + 84 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 527, + 84 + ], + "score": 1.0, + "content": "types of different professions) with certain real-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 85, + 525, + 98 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 525, + 98 + ], + "score": 1.0, + "content": "world populations. Otherwise, the misuse of the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 303, + 99, + 526, + 112 + ], + "spans": [ + { + "bbox": [ + 303, + 99, + 526, + 112 + ], + "score": 1.0, + "content": "proposed framework and LLM’s assessments might", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 112, + 526, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 112, + 526, + 126 + ], + "score": 1.0, + "content": "lead to unrealistic conclusions and even negative", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 127, + 525, + 138 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 525, + 138 + ], + "score": 1.0, + "content": "societal impacts (e.g., discrimination) on certain", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 140, + 525, + 152 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 525, + 152 + ], + "score": 1.0, + "content": "groups of people. Our framework must not be used", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 153, + 495, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 495, + 167 + ], + "score": 1.0, + "content": "for any ethically questionable applications.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 51 + }, + { + "type": "text", + "bbox": [ + 305, + 174, + 525, + 295 + ], + "lines": [ + { + "bbox": [ + 304, + 172, + 528, + 188 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 528, + 188 + ], + "score": 1.0, + "content": "Biases. The LLMs used in our study are pre-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 304, + 189, + 525, + 200 + ], + "spans": [ + { + "bbox": [ + 304, + 189, + 525, + 200 + ], + "score": 1.0, + "content": "trained on the large-scale datasets or Internet texts", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 201, + 525, + 214 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 525, + 214 + ], + "score": 1.0, + "content": "that may contain different biases or unsafe (e.g.,", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 214, + 527, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 527, + 228 + ], + "score": 1.0, + "content": "toxic) contents. Despite with human fine-tuning,", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 229, + 527, + 241 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 527, + 241 + ], + "score": 1.0, + "content": "the model could still generate some biased personal-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 242, + 525, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 525, + 255 + ], + "score": 1.0, + "content": "ity assessments that might not match the prevailing", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 256, + 527, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 527, + 268 + ], + "score": 1.0, + "content": "societal conceptions or values. Thus, the assess-", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 304, + 271, + 525, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 525, + 281 + ], + "score": 1.0, + "content": "ment results of LLMs via our framework must be", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 282, + 477, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 477, + 296 + ], + "score": 1.0, + "content": "further reviewed before generalization.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 59 + }, + { + "type": "text", + "bbox": [ + 304, + 303, + 525, + 546 + ], + "lines": [ + { + "bbox": [ + 304, + 303, + 527, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 527, + 317 + ], + "score": 1.0, + "content": "Broader Impact. Our study reveals the possi-", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 317, + 525, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 525, + 331 + ], + "score": 1.0, + "content": "bility of applying LLMs to automatically analyze", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 332, + 525, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 525, + 345 + ], + "score": 1.0, + "content": "human psychology such as personalities, and opens", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 303, + 345, + 525, + 357 + ], + "spans": [ + { + "bbox": [ + 303, + 345, + 525, + 357 + ], + "score": 1.0, + "content": "a new avenue to learn about their perceptions and", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 359, + 525, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 525, + 370 + ], + "score": 1.0, + "content": "assessments on humans, so as to better understand", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 371, + 526, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 526, + 385 + ], + "score": 1.0, + "content": "LLMs’ potential thinking modes, response moti-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 386, + 525, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 525, + 398 + ], + "score": 1.0, + "content": "vations, and communication principles. This can", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 399, + 526, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 526, + 412 + ], + "score": 1.0, + "content": "help speed up the development of more reliable,", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 413, + 525, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 525, + 425 + ], + "score": 1.0, + "content": "human-friendly, and trustworthy LLMs, as well", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 304, + 426, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 525, + 439 + ], + "score": 1.0, + "content": "as facilitate the future research of AI psychology", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 439, + 526, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 526, + 453 + ], + "score": 1.0, + "content": "and sociology. Our work suggests that LLMs such", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 452, + 527, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 527, + 466 + ], + "score": 1.0, + "content": "as InstructGPT may have biases on different gen-", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 304, + 466, + 525, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 525, + 479 + ], + "score": 1.0, + "content": "ders, which could incur societal and ethical risks", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 481, + 527, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 527, + 493 + ], + "score": 1.0, + "content": "in their applications. Based on our study, we ad-", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 303, + 493, + 526, + 507 + ], + "spans": [ + { + "bbox": [ + 303, + 493, + 526, + 507 + ], + "score": 1.0, + "content": "vocate introducing more human-like psychology", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 507, + 525, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 525, + 521 + ], + "score": 1.0, + "content": "and personality testings into the design and training", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 520, + 526, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 526, + 534 + ], + "score": 1.0, + "content": "of LLMs, so as to improve model safety and user", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 535, + 357, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 357, + 548 + ], + "score": 1.0, + "content": "experience.", + "type": "text" + } + ], + "index": 81 + } + ], + "index": 72.5 + }, + { + "type": "title", + "bbox": [ + 306, + 570, + 361, + 583 + ], + "lines": [ + { + "bbox": [ + 304, + 569, + 363, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 363, + 585 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 82 + } + ], + "index": 82 + }, + { + "type": "text", + "bbox": [ + 305, + 590, + 526, + 773 + ], + "lines": [ + { + "bbox": [ + 304, + 588, + 527, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 527, + 603 + ], + "score": 1.0, + "content": "Tolga Bolukbasi, Kai-Wei Chang, James Y Zou,", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 315, + 600, + 526, + 612 + ], + "spans": [ + { + "bbox": [ + 315, + 600, + 526, + 612 + ], + "score": 1.0, + "content": "Venkatesh Saligrama, and Adam T Kalai. 2016. Man", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 314, + 612, + 527, + 624 + ], + "spans": [ + { + "bbox": [ + 314, + 612, + 527, + 624 + ], + "score": 1.0, + "content": "is to computer programmer as woman is to home-", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 315, + 622, + 526, + 635 + ], + "spans": [ + { + "bbox": [ + 315, + 622, + 526, + 635 + ], + "score": 1.0, + "content": "maker? debiasing word embeddings. Advances in", + "type": "text" + } + ], + "index": 86 + }, + { + "bbox": [ + 314, + 632, + 526, + 646 + ], + "spans": [ + { + "bbox": [ + 314, + 632, + 526, + 646 + ], + "score": 1.0, + "content": "Neural Information Processing Systems (NeurIPS),", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 315, + 644, + 331, + 656 + ], + "spans": [ + { + "bbox": [ + 315, + 644, + 331, + 656 + ], + "score": 1.0, + "content": "29.", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 304, + 664, + 527, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 527, + 677 + ], + "score": 1.0, + "content": "Shikha Bordia and Samuel Bowman. 2019. Identify-", + "type": "text" + } + ], + "index": 89 + }, + { + "bbox": [ + 315, + 676, + 526, + 688 + ], + "spans": [ + { + "bbox": [ + 315, + 676, + 526, + 688 + ], + "score": 1.0, + "content": "ing and reducing gender bias in word-level language", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 315, + 686, + 527, + 699 + ], + "spans": [ + { + "bbox": [ + 315, + 686, + 527, + 699 + ], + "score": 1.0, + "content": "models. In Proceedings of the North American Chap-", + "type": "text" + } + ], + "index": 91 + }, + { + "bbox": [ + 315, + 698, + 525, + 710 + ], + "spans": [ + { + "bbox": [ + 315, + 698, + 525, + 710 + ], + "score": 1.0, + "content": "ter of the Association for Computational Linguistics", + "type": "text" + } + ], + "index": 92 + }, + { + "bbox": [ + 315, + 709, + 524, + 721 + ], + "spans": [ + { + "bbox": [ + 315, + 709, + 524, + 721 + ], + "score": 1.0, + "content": "(NAACL): Student Research Workshop, pages 7–15.", + "type": "text" + } + ], + "index": 93 + }, + { + "bbox": [ + 304, + 729, + 525, + 741 + ], + "spans": [ + { + "bbox": [ + 304, + 729, + 525, + 741 + ], + "score": 1.0, + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 315, + 740, + 525, + 751 + ], + "spans": [ + { + "bbox": [ + 315, + 740, + 525, + 751 + ], + "score": 1.0, + "content": "Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 315, + 751, + 525, + 763 + ], + "spans": [ + { + "bbox": [ + 315, + 751, + 525, + 763 + ], + "score": 1.0, + "content": "Neelakantan, Pranav Shyam, Girish Sastry, Amanda", + "type": "text" + } + ], + "index": 96 + }, + { + "bbox": [ + 315, + 761, + 525, + 774 + ], + "spans": [ + { + "bbox": [ + 315, + 761, + 525, + 774 + ], + "score": 1.0, + "content": "Askell, et al. 2020. Language models are few-shot", + "type": "text" + } + ], + "index": 97 + } + ], + "index": 90 + } + ], + "page_idx": 8, + "page_size": [ + 595, + 841 + ], + "discarded_blocks": [], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 70, + 72, + 290, + 165 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 68, + 71, + 291, + 165 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 70, + 176, + 187, + 189 + ], + "lines": [ + { + "bbox": [ + 67, + 174, + 189, + 192 + ], + "spans": [ + { + "bbox": [ + 67, + 174, + 189, + 192 + ], + "score": 1.0, + "content": "8 Acknowledgements", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 69, + 198, + 290, + 250 + ], + "lines": [ + { + "bbox": [ + 69, + 198, + 290, + 210 + ], + "spans": [ + { + "bbox": [ + 69, + 198, + 290, + 210 + ], + "score": 1.0, + "content": "This research is supported by the National Research", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 211, + 291, + 225 + ], + "spans": [ + { + "bbox": [ + 68, + 211, + 291, + 225 + ], + "score": 1.0, + "content": "Foundation, Singapore under its AI Singapore Pro-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 225, + 292, + 236 + ], + "spans": [ + { + "bbox": [ + 68, + 225, + 292, + 236 + ], + "score": 1.0, + "content": "gramme (AISG Award No: AISG2-PhD/2022-01-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 237, + 109, + 252 + ], + "spans": [ + { + "bbox": [ + 68, + 237, + 109, + 252 + ], + "score": 1.0, + "content": "034[T]).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 68, + 198, + 292, + 252 + ] + }, + { + "type": "title", + "bbox": [ + 70, + 261, + 129, + 275 + ], + "lines": [ + { + "bbox": [ + 69, + 261, + 131, + 276 + ], + "spans": [ + { + "bbox": [ + 69, + 261, + 131, + 276 + ], + "score": 1.0, + "content": "Limitations", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 69, + 280, + 290, + 703 + ], + "lines": [ + { + "bbox": [ + 68, + 282, + 290, + 297 + ], + "spans": [ + { + "bbox": [ + 68, + 282, + 290, + 297 + ], + "score": 1.0, + "content": "While our study is a step toward the promising", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 68, + 296, + 291, + 311 + ], + "spans": [ + { + "bbox": [ + 68, + 296, + 291, + 311 + ], + "score": 1.0, + "content": "open direction of LLM-based human personality", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 310, + 291, + 324 + ], + "spans": [ + { + "bbox": [ + 69, + 310, + 291, + 324 + ], + "score": 1.0, + "content": "and psychology assessment, it possesses limitations", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 324, + 291, + 337 + ], + "spans": [ + { + "bbox": [ + 69, + 324, + 291, + 337 + ], + "score": 1.0, + "content": "and opportunities when applied to the real world.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 338, + 290, + 349 + ], + "spans": [ + { + "bbox": [ + 69, + 338, + 290, + 349 + ], + "score": 1.0, + "content": "First, our work focuses on ChatGPT model series", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 69, + 351, + 290, + 363 + ], + "spans": [ + { + "bbox": [ + 69, + 351, + 290, + 363 + ], + "score": 1.0, + "content": "and the experiments are conducted on a limited", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 68, + 364, + 290, + 377 + ], + "spans": [ + { + "bbox": [ + 68, + 364, + 290, + 377 + ], + "score": 1.0, + "content": "number of LLMs. Our framework is also scalable", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 68, + 378, + 290, + 390 + ], + "spans": [ + { + "bbox": [ + 68, + 378, + 290, + 390 + ], + "score": 1.0, + "content": "to be applied to other LLMs such as LLaMA, while", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 69, + 392, + 291, + 405 + ], + "spans": [ + { + "bbox": [ + 69, + 392, + 291, + 405 + ], + "score": 1.0, + "content": "its performance remains to be further explored. Sec-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 69, + 406, + 290, + 418 + ], + "spans": [ + { + "bbox": [ + 69, + 406, + 290, + 418 + ], + "score": 1.0, + "content": "ond, although most independent testings of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 68, + 418, + 291, + 432 + ], + "spans": [ + { + "bbox": [ + 68, + 418, + 291, + 432 + ], + "score": 1.0, + "content": "LLM under the same standard setting yield sim-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 431, + 291, + 446 + ], + "spans": [ + { + "bbox": [ + 68, + 431, + 291, + 446 + ], + "score": 1.0, + "content": "ilar assessments, the experimental setting (e.g.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 68, + 446, + 290, + 458 + ], + "spans": [ + { + "bbox": [ + 68, + 446, + 290, + 458 + ], + "score": 1.0, + "content": "hyper-parameters) or testing number can be further", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 68, + 459, + 290, + 472 + ], + "spans": [ + { + "bbox": [ + 68, + 459, + 290, + 472 + ], + "score": 1.0, + "content": "customized to test the reliability of LLMs under", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 68, + 471, + 291, + 488 + ], + "spans": [ + { + "bbox": [ + 68, + 471, + 291, + 488 + ], + "score": 1.0, + "content": "extreme cases. We will leverage the upcoming", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 68, + 486, + 291, + 500 + ], + "spans": [ + { + "bbox": [ + 68, + 486, + 291, + 500 + ], + "score": 1.0, + "content": "API that supports controllable hyper-parameters to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 68, + 498, + 292, + 514 + ], + "spans": [ + { + "bbox": [ + 68, + 498, + 292, + 514 + ], + "score": 1.0, + "content": "better evaluate GPT models. Third, the represen-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 69, + 513, + 291, + 527 + ], + "spans": [ + { + "bbox": [ + 69, + 513, + 291, + 527 + ], + "score": 1.0, + "content": "tations of different genders might be insufficient.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 68, + 527, + 292, + 540 + ], + "spans": [ + { + "bbox": [ + 68, + 527, + 292, + 540 + ], + "score": 1.0, + "content": "For example, the subjects “Ladies” and “Gentle-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 69, + 541, + 291, + 554 + ], + "spans": [ + { + "bbox": [ + 69, + 541, + 291, + 554 + ], + "score": 1.0, + "content": "men” also have different genders, while they can", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 68, + 554, + 291, + 567 + ], + "spans": [ + { + "bbox": [ + 68, + 554, + 291, + 567 + ], + "score": 1.0, + "content": "be viewed as groups that differ from “Men” and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 67, + 566, + 291, + 581 + ], + "spans": [ + { + "bbox": [ + 67, + 566, + 291, + 581 + ], + "score": 1.0, + "content": "“Women”. As the focus of this work is to devise", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 68, + 581, + 290, + 594 + ], + "spans": [ + { + "bbox": [ + 68, + 581, + 290, + 594 + ], + "score": 1.0, + "content": "a general evaluation framework, we will further", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 69, + 595, + 291, + 608 + ], + "spans": [ + { + "bbox": [ + 69, + 595, + 291, + 608 + ], + "score": 1.0, + "content": "explore the assessment of more diverse subjects in", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 68, + 607, + 291, + 622 + ], + "spans": [ + { + "bbox": [ + 68, + 607, + 291, + 622 + ], + "score": 1.0, + "content": "future works. Last, despite the popularity of MBTI", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 68, + 621, + 291, + 635 + ], + "spans": [ + { + "bbox": [ + 68, + 621, + 291, + 635 + ], + "score": 1.0, + "content": "in different areas, its scientific validity is still under", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 68, + 635, + 291, + 649 + ], + "spans": [ + { + "bbox": [ + 68, + 635, + 291, + 649 + ], + "score": 1.0, + "content": "exploration. In our work, MBTI is adopted as a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 68, + 649, + 290, + 662 + ], + "spans": [ + { + "bbox": [ + 68, + 649, + 290, + 662 + ], + "score": 1.0, + "content": "representative personality measure to help LLMs", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 69, + 662, + 291, + 676 + ], + "spans": [ + { + "bbox": [ + 69, + 662, + 291, + 676 + ], + "score": 1.0, + "content": "conduct quantitative evaluations. We will explore", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 68, + 676, + 290, + 689 + ], + "spans": [ + { + "bbox": [ + 68, + 676, + 290, + 689 + ], + "score": 1.0, + "content": "other tests such as Big Five Inventory (BFI) (John", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 69, + 690, + 259, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 690, + 259, + 703 + ], + "score": 1.0, + "content": "et al., 1999) under our scalable framework.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 28, + "bbox_fs": [ + 67, + 282, + 292, + 703 + ] + }, + { + "type": "title", + "bbox": [ + 70, + 712, + 183, + 726 + ], + "lines": [ + { + "bbox": [ + 68, + 711, + 184, + 727 + ], + "spans": [ + { + "bbox": [ + 68, + 711, + 184, + 727 + ], + "score": 1.0, + "content": "Ethics Considerations", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 69, + 734, + 289, + 774 + ], + "lines": [ + { + "bbox": [ + 68, + 732, + 291, + 748 + ], + "spans": [ + { + "bbox": [ + 68, + 732, + 291, + 748 + ], + "score": 1.0, + "content": "Misuse Potential. Due to the exploratory nature", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 68, + 747, + 291, + 761 + ], + "spans": [ + { + "bbox": [ + 68, + 747, + 291, + 761 + ], + "score": 1.0, + "content": "of our study, one should not directly use, generalize", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 68, + 761, + 290, + 774 + ], + "spans": [ + { + "bbox": [ + 68, + 761, + 290, + 774 + ], + "score": 1.0, + "content": "or match the assessment results (e.g., personality", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 72, + 527, + 84 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 527, + 84 + ], + "score": 1.0, + "content": "types of different professions) with certain real-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 85, + 525, + 98 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 525, + 98 + ], + "score": 1.0, + "content": "world populations. Otherwise, the misuse of the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 303, + 99, + 526, + 112 + ], + "spans": [ + { + "bbox": [ + 303, + 99, + 526, + 112 + ], + "score": 1.0, + "content": "proposed framework and LLM’s assessments might", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 112, + 526, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 112, + 526, + 126 + ], + "score": 1.0, + "content": "lead to unrealistic conclusions and even negative", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 127, + 525, + 138 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 525, + 138 + ], + "score": 1.0, + "content": "societal impacts (e.g., discrimination) on certain", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 140, + 525, + 152 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 525, + 152 + ], + "score": 1.0, + "content": "groups of people. Our framework must not be used", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 153, + 495, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 495, + 167 + ], + "score": 1.0, + "content": "for any ethically questionable applications.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 46, + "bbox_fs": [ + 68, + 732, + 291, + 774 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 72, + 525, + 166 + ], + "lines": [], + "index": 51, + "bbox_fs": [ + 303, + 72, + 527, + 167 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 305, + 174, + 525, + 295 + ], + "lines": [ + { + "bbox": [ + 304, + 172, + 528, + 188 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 528, + 188 + ], + "score": 1.0, + "content": "Biases. The LLMs used in our study are pre-", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 304, + 189, + 525, + 200 + ], + "spans": [ + { + "bbox": [ + 304, + 189, + 525, + 200 + ], + "score": 1.0, + "content": "trained on the large-scale datasets or Internet texts", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 201, + 525, + 214 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 525, + 214 + ], + "score": 1.0, + "content": "that may contain different biases or unsafe (e.g.,", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 214, + 527, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 527, + 228 + ], + "score": 1.0, + "content": "toxic) contents. Despite with human fine-tuning,", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 229, + 527, + 241 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 527, + 241 + ], + "score": 1.0, + "content": "the model could still generate some biased personal-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 242, + 525, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 525, + 255 + ], + "score": 1.0, + "content": "ity assessments that might not match the prevailing", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 256, + 527, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 527, + 268 + ], + "score": 1.0, + "content": "societal conceptions or values. Thus, the assess-", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 304, + 271, + 525, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 525, + 281 + ], + "score": 1.0, + "content": "ment results of LLMs via our framework must be", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 304, + 282, + 477, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 477, + 296 + ], + "score": 1.0, + "content": "further reviewed before generalization.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 59, + "bbox_fs": [ + 304, + 172, + 528, + 296 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 303, + 525, + 546 + ], + "lines": [ + { + "bbox": [ + 304, + 303, + 527, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 527, + 317 + ], + "score": 1.0, + "content": "Broader Impact. Our study reveals the possi-", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 317, + 525, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 525, + 331 + ], + "score": 1.0, + "content": "bility of applying LLMs to automatically analyze", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 332, + 525, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 525, + 345 + ], + "score": 1.0, + "content": "human psychology such as personalities, and opens", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 303, + 345, + 525, + 357 + ], + "spans": [ + { + "bbox": [ + 303, + 345, + 525, + 357 + ], + "score": 1.0, + "content": "a new avenue to learn about their perceptions and", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 359, + 525, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 525, + 370 + ], + "score": 1.0, + "content": "assessments on humans, so as to better understand", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 371, + 526, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 526, + 385 + ], + "score": 1.0, + "content": "LLMs’ potential thinking modes, response moti-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 304, + 386, + 525, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 525, + 398 + ], + "score": 1.0, + "content": "vations, and communication principles. This can", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 399, + 526, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 526, + 412 + ], + "score": 1.0, + "content": "help speed up the development of more reliable,", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 413, + 525, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 525, + 425 + ], + "score": 1.0, + "content": "human-friendly, and trustworthy LLMs, as well", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 304, + 426, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 525, + 439 + ], + "score": 1.0, + "content": "as facilitate the future research of AI psychology", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 304, + 439, + 526, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 526, + 453 + ], + "score": 1.0, + "content": "and sociology. Our work suggests that LLMs such", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 452, + 527, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 527, + 466 + ], + "score": 1.0, + "content": "as InstructGPT may have biases on different gen-", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 304, + 466, + 525, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 525, + 479 + ], + "score": 1.0, + "content": "ders, which could incur societal and ethical risks", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 481, + 527, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 527, + 493 + ], + "score": 1.0, + "content": "in their applications. Based on our study, we ad-", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 303, + 493, + 526, + 507 + ], + "spans": [ + { + "bbox": [ + 303, + 493, + 526, + 507 + ], + "score": 1.0, + "content": "vocate introducing more human-like psychology", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 304, + 507, + 525, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 525, + 521 + ], + "score": 1.0, + "content": "and personality testings into the design and training", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 304, + 520, + 526, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 526, + 534 + ], + "score": 1.0, + "content": "of LLMs, so as to improve model safety and user", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 304, + 535, + 357, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 357, + 548 + ], + "score": 1.0, + "content": "experience.", + "type": "text" + } + ], + "index": 81 + } + ], + "index": 72.5, + "bbox_fs": [ + 303, + 303, + 527, + 548 + ] + }, + { + "type": "title", + "bbox": [ + 306, + 570, + 361, + 583 + ], + "lines": [ + { + "bbox": [ + 304, + 569, + 363, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 363, + 585 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 82 + } + ], + "index": 82 + }, + { + "type": "text", + "bbox": [ + 305, + 590, + 526, + 773 + ], + "lines": [ + { + "bbox": [ + 304, + 588, + 527, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 527, + 603 + ], + "score": 1.0, + "content": "Tolga Bolukbasi, Kai-Wei Chang, James Y Zou,", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 315, + 600, + 526, + 612 + ], + "spans": [ + { + "bbox": [ + 315, + 600, + 526, + 612 + ], + "score": 1.0, + "content": "Venkatesh Saligrama, and Adam T Kalai. 2016. Man", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 314, + 612, + 527, + 624 + ], + "spans": [ + { + "bbox": [ + 314, + 612, + 527, + 624 + ], + "score": 1.0, + "content": "is to computer programmer as woman is to home-", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 315, + 622, + 526, + 635 + ], + "spans": [ + { + "bbox": [ + 315, + 622, + 526, + 635 + ], + "score": 1.0, + "content": "maker? debiasing word embeddings. Advances in", + "type": "text" + } + ], + "index": 86 + }, + { + "bbox": [ + 314, + 632, + 526, + 646 + ], + "spans": [ + { + "bbox": [ + 314, + 632, + 526, + 646 + ], + "score": 1.0, + "content": "Neural Information Processing Systems (NeurIPS),", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 315, + 644, + 331, + 656 + ], + "spans": [ + { + "bbox": [ + 315, + 644, + 331, + 656 + ], + "score": 1.0, + "content": "29.", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 304, + 664, + 527, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 527, + 677 + ], + "score": 1.0, + "content": "Shikha Bordia and Samuel Bowman. 2019. Identify-", + "type": "text" + } + ], + "index": 89 + }, + { + "bbox": [ + 315, + 676, + 526, + 688 + ], + "spans": [ + { + "bbox": [ + 315, + 676, + 526, + 688 + ], + "score": 1.0, + "content": "ing and reducing gender bias in word-level language", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 315, + 686, + 527, + 699 + ], + "spans": [ + { + "bbox": [ + 315, + 686, + 527, + 699 + ], + "score": 1.0, + "content": "models. In Proceedings of the North American Chap-", + "type": "text" + } + ], + "index": 91 + }, + { + "bbox": [ + 315, + 698, + 525, + 710 + ], + "spans": [ + { + "bbox": [ + 315, + 698, + 525, + 710 + ], + "score": 1.0, + "content": "ter of the Association for Computational Linguistics", + "type": "text" + } + ], + "index": 92 + }, + { + "bbox": [ + 315, + 709, + 524, + 721 + ], + "spans": [ + { + "bbox": [ + 315, + 709, + 524, + 721 + ], + "score": 1.0, + "content": "(NAACL): Student Research Workshop, pages 7–15.", + "type": "text" + } + ], + "index": 93 + }, + { + "bbox": [ + 304, + 729, + 525, + 741 + ], + "spans": [ + { + "bbox": [ + 304, + 729, + 525, + 741 + ], + "score": 1.0, + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 315, + 740, + 525, + 751 + ], + "spans": [ + { + "bbox": [ + 315, + 740, + 525, + 751 + ], + "score": 1.0, + "content": "Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 315, + 751, + 525, + 763 + ], + "spans": [ + { + "bbox": [ + 315, + 751, + 525, + 763 + ], + "score": 1.0, + "content": "Neelakantan, Pranav Shyam, Girish Sastry, Amanda", + "type": "text" + } + ], + "index": 96 + }, + { + "bbox": [ + 315, + 761, + 525, + 774 + ], + "spans": [ + { + "bbox": [ + 315, + 761, + 525, + 774 + ], + "score": 1.0, + "content": "Askell, et al. 2020. Language models are few-shot", + "type": "text" + } + ], + "index": 97 + }, + { + "bbox": [ + 79, + 70, + 291, + 86 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 291, + 86 + ], + "score": 1.0, + "content": "learners. Advances in Neural Information Processing", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 79, + 84, + 222, + 95 + ], + "spans": [ + { + "bbox": [ + 79, + 84, + 222, + 95 + ], + "score": 1.0, + "content": "Systems (NeurIPS), 33:1877–1901.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 90, + "bbox_fs": [ + 304, + 588, + 527, + 774 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 80, + 73, + 289, + 95 + ], + "lines": [ + { + "bbox": [ + 79, + 70, + 291, + 86 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 291, + 86 + ], + "score": 1.0, + "content": "learners. Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 79, + 84, + 222, + 95 + ], + "spans": [ + { + "bbox": [ + 79, + 84, + 222, + 95 + ], + "score": 1.0, + "content": "Systems (NeurIPS), 33:1877–1901.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 70, + 103, + 289, + 168 + ], + "lines": [ + { + "bbox": [ + 69, + 102, + 291, + 114 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 291, + 114 + ], + "score": 1.0, + "content": "Sébastien Bubeck, Varun Chandrasekaran, Ronen El-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 113, + 291, + 126 + ], + "spans": [ + { + "bbox": [ + 79, + 113, + 291, + 126 + ], + "score": 1.0, + "content": "dan, Johannes Gehrke, Eric Horvitz, Ece Kamar,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 125, + 291, + 136 + ], + "spans": [ + { + "bbox": [ + 80, + 125, + 291, + 136 + ], + "score": 1.0, + "content": "Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lund-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 136, + 291, + 147 + ], + "spans": [ + { + "bbox": [ + 79, + 136, + 291, + 147 + ], + "score": 1.0, + "content": "berg, et al. 2023. Sparks of artificial general intelli-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 146, + 291, + 159 + ], + "spans": [ + { + "bbox": [ + 78, + 146, + 291, + 159 + ], + "score": 1.0, + "content": "gence: Early experiments with GPT-4. arXiv preprint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 158, + 158, + 169 + ], + "spans": [ + { + "bbox": [ + 79, + 158, + 158, + 169 + ], + "score": 1.0, + "content": "arXiv:2303.12712.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 69, + 177, + 290, + 210 + ], + "lines": [ + { + "bbox": [ + 69, + 177, + 291, + 189 + ], + "spans": [ + { + "bbox": [ + 69, + 177, + 291, + 189 + ], + "score": 1.0, + "content": "Graham Caron and Shashank Srivastava. 2022. Identi-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 79, + 188, + 291, + 200 + ], + "spans": [ + { + "bbox": [ + 79, + 188, + 291, + 200 + ], + "score": 1.0, + "content": "fying and manipulating the personality traits of lan-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 79, + 199, + 278, + 210 + ], + "spans": [ + { + "bbox": [ + 79, + 199, + 278, + 210 + ], + "score": 1.0, + "content": "guage models. arXiv preprint arXiv:2212.10276.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 69, + 218, + 289, + 295 + ], + "lines": [ + { + "bbox": [ + 68, + 218, + 290, + 230 + ], + "spans": [ + { + "bbox": [ + 68, + 218, + 290, + 230 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 79, + 228, + 291, + 241 + ], + "spans": [ + { + "bbox": [ + 79, + 228, + 291, + 241 + ], + "score": 1.0, + "content": "Kristina Toutanova. 2019. BERT: Pre-training of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 79, + 240, + 291, + 253 + ], + "spans": [ + { + "bbox": [ + 79, + 240, + 291, + 253 + ], + "score": 1.0, + "content": "deep bidirectional transformers for language under-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 80, + 251, + 290, + 263 + ], + "spans": [ + { + "bbox": [ + 80, + 251, + 290, + 263 + ], + "score": 1.0, + "content": "standing. In Proceedings of the 2019 Conference", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 80, + 262, + 290, + 274 + ], + "spans": [ + { + "bbox": [ + 80, + 262, + 290, + 274 + ], + "score": 1.0, + "content": "of the North American Chapter of the Association", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 78, + 272, + 291, + 286 + ], + "spans": [ + { + "bbox": [ + 78, + 272, + 291, + 286 + ], + "score": 1.0, + "content": "for Computational Linguistics: Human Language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 78, + 283, + 273, + 296 + ], + "spans": [ + { + "bbox": [ + 78, + 283, + 273, + 296 + ], + "score": 1.0, + "content": "Technologies (NAACL-HLT), pages 4171–4186.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 69, + 303, + 291, + 336 + ], + "lines": [ + { + "bbox": [ + 69, + 303, + 291, + 315 + ], + "spans": [ + { + "bbox": [ + 69, + 303, + 291, + 315 + ], + "score": 1.0, + "content": "John M Digman. 1990. Personality structure: Emer-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 78, + 313, + 292, + 327 + ], + "spans": [ + { + "bbox": [ + 78, + 313, + 292, + 327 + ], + "score": 1.0, + "content": "gence of the five-factor model. Annual review of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 78, + 325, + 194, + 337 + ], + "spans": [ + { + "bbox": [ + 78, + 325, + 194, + 337 + ], + "score": 1.0, + "content": "psychology, 41(1):417–440.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 69, + 345, + 290, + 366 + ], + "lines": [ + { + "bbox": [ + 68, + 343, + 292, + 358 + ], + "spans": [ + { + "bbox": [ + 68, + 343, + 292, + 358 + ], + "score": 1.0, + "content": "Hans Jurgen Eysenck. 2012. A model for personality.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 79, + 356, + 230, + 367 + ], + "spans": [ + { + "bbox": [ + 79, + 356, + 230, + 367 + ], + "score": 1.0, + "content": "Springer Science & Business Media.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 69, + 375, + 290, + 419 + ], + "lines": [ + { + "bbox": [ + 69, + 374, + 291, + 387 + ], + "spans": [ + { + "bbox": [ + 69, + 374, + 291, + 387 + ], + "score": 1.0, + "content": "Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 79, + 385, + 292, + 397 + ], + "spans": [ + { + "bbox": [ + 79, + 385, + 292, + 397 + ], + "score": 1.0, + "content": "Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2022.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 78, + 396, + 291, + 410 + ], + "spans": [ + { + "bbox": [ + 78, + 396, + 291, + 410 + ], + "score": 1.0, + "content": "Large language models can self-improve. arXiv", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 79, + 408, + 194, + 419 + ], + "spans": [ + { + "bbox": [ + 79, + 408, + 194, + 419 + ], + "score": 1.0, + "content": "preprint arXiv:2210.11610.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 69, + 427, + 290, + 471 + ], + "lines": [ + { + "bbox": [ + 68, + 426, + 292, + 439 + ], + "spans": [ + { + "bbox": [ + 68, + 426, + 292, + 439 + ], + "score": 1.0, + "content": "Guangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wen-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 78, + 437, + 292, + 450 + ], + "spans": [ + { + "bbox": [ + 78, + 437, + 292, + 450 + ], + "score": 1.0, + "content": "juan Han, Chi Zhang, and Yixin Zhu. 2022. MPI:", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 79, + 449, + 290, + 461 + ], + "spans": [ + { + "bbox": [ + 79, + 449, + 290, + 461 + ], + "score": 1.0, + "content": "Evaluating and inducing personality in pre-trained", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 79, + 460, + 291, + 472 + ], + "spans": [ + { + "bbox": [ + 79, + 460, + 291, + 472 + ], + "score": 1.0, + "content": "language models. arXiv preprint arXiv:2206.07550.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 69, + 479, + 289, + 523 + ], + "lines": [ + { + "bbox": [ + 69, + 479, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 69, + 479, + 290, + 491 + ], + "score": 1.0, + "content": "Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 79, + 489, + 290, + 503 + ], + "spans": [ + { + "bbox": [ + 79, + 489, + 290, + 503 + ], + "score": 1.0, + "content": "Neubig. 2020. How can we know what language", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 80, + 501, + 290, + 513 + ], + "spans": [ + { + "bbox": [ + 80, + 501, + 290, + 513 + ], + "score": 1.0, + "content": "models know? Transactions of the Association for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 80, + 512, + 239, + 524 + ], + "spans": [ + { + "bbox": [ + 80, + 512, + 239, + 524 + ], + "score": 1.0, + "content": "Computational Linguistics, 8:423–438.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 69, + 531, + 290, + 564 + ], + "lines": [ + { + "bbox": [ + 68, + 530, + 292, + 544 + ], + "spans": [ + { + "bbox": [ + 68, + 530, + 292, + 544 + ], + "score": 1.0, + "content": "Oliver P John, Sanjay Srivastava, et al. 1999. The big-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 80, + 543, + 291, + 554 + ], + "spans": [ + { + "bbox": [ + 80, + 543, + 291, + 554 + ], + "score": 1.0, + "content": "five trait taxonomy: History, measurement, and theo-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 80, + 554, + 162, + 565 + ], + "spans": [ + { + "bbox": [ + 80, + 554, + 162, + 565 + ], + "score": 1.0, + "content": "retical perspectives.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 69, + 573, + 290, + 616 + ], + "lines": [ + { + "bbox": [ + 69, + 572, + 291, + 585 + ], + "spans": [ + { + "bbox": [ + 69, + 572, + 291, + 585 + ], + "score": 1.0, + "content": "Saketh Reddy Karra, Son Nguyen, and Theja Tula-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 79, + 583, + 290, + 596 + ], + "spans": [ + { + "bbox": [ + 79, + 583, + 290, + 596 + ], + "score": 1.0, + "content": "bandhula. 2022. AI personification: Estimating", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 80, + 595, + 290, + 606 + ], + "spans": [ + { + "bbox": [ + 80, + 595, + 290, + 606 + ], + "score": 1.0, + "content": "the personality of language models. arXiv preprint", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 79, + 605, + 158, + 616 + ], + "spans": [ + { + "bbox": [ + 79, + 605, + 158, + 616 + ], + "score": 1.0, + "content": "arXiv:2204.12000.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 69, + 625, + 290, + 658 + ], + "lines": [ + { + "bbox": [ + 69, + 624, + 291, + 637 + ], + "spans": [ + { + "bbox": [ + 69, + 624, + 291, + 637 + ], + "score": 1.0, + "content": "Michal Kosinski. 2023. Theory of mind may have spon-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 79, + 636, + 290, + 648 + ], + "spans": [ + { + "bbox": [ + 79, + 636, + 290, + 648 + ], + "score": 1.0, + "content": "taneously emerged in large language models. arXiv", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 79, + 648, + 193, + 658 + ], + "spans": [ + { + "bbox": [ + 79, + 648, + 193, + 658 + ], + "score": 1.0, + "content": "preprint arXiv:2302.02083.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 69, + 666, + 290, + 710 + ], + "lines": [ + { + "bbox": [ + 68, + 665, + 291, + 679 + ], + "spans": [ + { + "bbox": [ + 68, + 665, + 291, + 679 + ], + "score": 1.0, + "content": "Xingxuan Li, Yutong Li, Linlin Liu, Lidong Bing, and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 80, + 677, + 291, + 689 + ], + "spans": [ + { + "bbox": [ + 80, + 677, + 291, + 689 + ], + "score": 1.0, + "content": "Shafiq Joty. 2022. Is GPT-3 a psychopath? evalu-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 80, + 688, + 291, + 700 + ], + "spans": [ + { + "bbox": [ + 80, + 688, + 291, + 700 + ], + "score": 1.0, + "content": "ating large language models from a psychological", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 79, + 699, + 268, + 711 + ], + "spans": [ + { + "bbox": [ + 79, + 699, + 268, + 711 + ], + "score": 1.0, + "content": "perspective. arXiv preprint arXiv:2212.10529.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + }, + { + "type": "text", + "bbox": [ + 69, + 718, + 290, + 773 + ], + "lines": [ + { + "bbox": [ + 68, + 716, + 293, + 731 + ], + "spans": [ + { + "bbox": [ + 68, + 716, + 293, + 731 + ], + "score": 1.0, + "content": "Marilù Miotto, Nicola Rossberg, and Bennett Kleinberg.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 79, + 728, + 291, + 741 + ], + "spans": [ + { + "bbox": [ + 79, + 728, + 291, + 741 + ], + "score": 1.0, + "content": "2022. Who is GPT-3? An exploration of person-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 79, + 739, + 290, + 752 + ], + "spans": [ + { + "bbox": [ + 79, + 739, + 290, + 752 + ], + "score": 1.0, + "content": "ality, values and demographics. In Conference on", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 79, + 749, + 290, + 765 + ], + "spans": [ + { + "bbox": [ + 79, + 749, + 290, + 765 + ], + "score": 1.0, + "content": "Empirical Methods in Natural Language Processing", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 79, + 761, + 166, + 775 + ], + "spans": [ + { + "bbox": [ + 79, + 761, + 166, + 775 + ], + "score": 1.0, + "content": "(EMNLP) Workshop.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 51 + }, + { + "type": "text", + "bbox": [ + 304, + 73, + 525, + 117 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 527, + 86 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 527, + 86 + ], + "score": 1.0, + "content": "Shima Rahimi Moghaddam and Christopher J Honey.", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 315, + 83, + 525, + 96 + ], + "spans": [ + { + "bbox": [ + 315, + 83, + 525, + 96 + ], + "score": 1.0, + "content": "2023. Boosting theory-of-mind performance in large", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 315, + 95, + 525, + 107 + ], + "spans": [ + { + "bbox": [ + 315, + 95, + 525, + 107 + ], + "score": 1.0, + "content": "language models via prompting. arXiv preprint", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 315, + 106, + 394, + 117 + ], + "spans": [ + { + "bbox": [ + 315, + 106, + 394, + 117 + ], + "score": 1.0, + "content": "arXiv:2304.11490.", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 55.5 + }, + { + "type": "text", + "bbox": [ + 304, + 126, + 524, + 148 + ], + "lines": [ + { + "bbox": [ + 303, + 124, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 303, + 124, + 526, + 140 + ], + "score": 1.0, + "content": "Isabel Briggs Myers. 1962. The Myers-Briggs Type", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 315, + 137, + 422, + 149 + ], + "spans": [ + { + "bbox": [ + 315, + 137, + 422, + 149 + ], + "score": 1.0, + "content": "Indicator: Manual (1962).", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 58.5 + }, + { + "type": "text", + "bbox": [ + 304, + 158, + 525, + 202 + ], + "lines": [ + { + "bbox": [ + 304, + 158, + 527, + 170 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 527, + 170 + ], + "score": 1.0, + "content": "Isabel Briggs Myers and Mary H. McCaulley. 1985.", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 315, + 169, + 525, + 181 + ], + "spans": [ + { + "bbox": [ + 315, + 169, + 525, + 181 + ], + "score": 1.0, + "content": "Manual: A guide to the development and use of the", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 315, + 180, + 526, + 192 + ], + "spans": [ + { + "bbox": [ + 315, + 180, + 526, + 192 + ], + "score": 1.0, + "content": "Myers-Briggs Type Indicator. Consulting Psycholo-", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 314, + 191, + 363, + 203 + ], + "spans": [ + { + "bbox": [ + 314, + 191, + 363, + 203 + ], + "score": 1.0, + "content": "gists Press.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 61.5 + }, + { + "type": "text", + "bbox": [ + 304, + 212, + 525, + 288 + ], + "lines": [ + { + "bbox": [ + 304, + 212, + 527, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 527, + 224 + ], + "score": 1.0, + "content": "Moin Nadeem, Anna Bethke, and Siva Reddy. 2021.", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 315, + 223, + 526, + 235 + ], + "spans": [ + { + "bbox": [ + 315, + 223, + 526, + 235 + ], + "score": 1.0, + "content": "StereoSet: Measuring stereotypical bias in pretrained", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 315, + 234, + 527, + 245 + ], + "spans": [ + { + "bbox": [ + 315, + 234, + 527, + 245 + ], + "score": 1.0, + "content": "language models. In Proceedings of the Annual Meet-", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 315, + 244, + 526, + 257 + ], + "spans": [ + { + "bbox": [ + 315, + 244, + 526, + 257 + ], + "score": 1.0, + "content": "ing of the Association for Computational Linguistics", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 315, + 255, + 526, + 267 + ], + "spans": [ + { + "bbox": [ + 315, + 255, + 526, + 267 + ], + "score": 1.0, + "content": "and the International Joint Conference on Natural", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 314, + 266, + 527, + 279 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 527, + 279 + ], + "score": 1.0, + "content": "Language Processing (ACL-IJCNLP), pages 5356–", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 314, + 276, + 342, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 276, + 342, + 289 + ], + "score": 1.0, + "content": "5371.", + "type": "text" + } + ], + "index": 70 + } + ], + "index": 67 + }, + { + "type": "text", + "bbox": [ + 304, + 298, + 525, + 365 + ], + "lines": [ + { + "bbox": [ + 304, + 298, + 527, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 527, + 311 + ], + "score": 1.0, + "content": "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida,", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 314, + 308, + 527, + 322 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 527, + 322 + ], + "score": 1.0, + "content": "Carroll Wainwright, Pamela Mishkin, Chong Zhang,", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 315, + 319, + 527, + 332 + ], + "spans": [ + { + "bbox": [ + 315, + 319, + 527, + 332 + ], + "score": 1.0, + "content": "Sandhini Agarwal, Katarina Slama, Alex Gray, et al.", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 314, + 330, + 528, + 344 + ], + "spans": [ + { + "bbox": [ + 314, + 330, + 528, + 344 + ], + "score": 1.0, + "content": "2022. Training language models to follow instruc-", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 315, + 342, + 526, + 353 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 526, + 353 + ], + "score": 1.0, + "content": "tions with human feedback. In Advances in Neural", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 315, + 353, + 492, + 365 + ], + "spans": [ + { + "bbox": [ + 315, + 353, + 492, + 365 + ], + "score": 1.0, + "content": "Information Processing Systems (NeurIPS).", + "type": "text" + } + ], + "index": 76 + } + ], + "index": 73.5 + }, + { + "type": "text", + "bbox": [ + 304, + 374, + 525, + 418 + ], + "lines": [ + { + "bbox": [ + 304, + 373, + 527, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 527, + 386 + ], + "score": 1.0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan,", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 314, + 384, + 526, + 398 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 526, + 398 + ], + "score": 1.0, + "content": "Dario Amodei, Ilya Sutskever, et al. 2019. Language", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 316, + 396, + 525, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 525, + 407 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. OpenAI", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 315, + 406, + 368, + 419 + ], + "spans": [ + { + "bbox": [ + 315, + 406, + 368, + 419 + ], + "score": 1.0, + "content": "blog, 1(8):9.", + "type": "text" + } + ], + "index": 80 + } + ], + "index": 78.5 + }, + { + "type": "text", + "bbox": [ + 304, + 428, + 525, + 493 + ], + "lines": [ + { + "bbox": [ + 304, + 427, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 525, + 439 + ], + "score": 1.0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 313, + 437, + 527, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 527, + 452 + ], + "score": 1.0, + "content": "Lee, Sharan Narang, Michael Matena, Yanqi Zhou,", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 315, + 449, + 525, + 462 + ], + "spans": [ + { + "bbox": [ + 315, + 449, + 525, + 462 + ], + "score": 1.0, + "content": "Wei Li, and Peter J Liu. 2020. Exploring the limits", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 315, + 460, + 527, + 472 + ], + "spans": [ + { + "bbox": [ + 315, + 460, + 527, + 472 + ], + "score": 1.0, + "content": "of transfer learning with a unified text-to-text trans-", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 315, + 470, + 527, + 484 + ], + "spans": [ + { + "bbox": [ + 315, + 470, + 527, + 484 + ], + "score": 1.0, + "content": "former. The Journal of Machine Learning Research,", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 315, + 482, + 390, + 495 + ], + "spans": [ + { + "bbox": [ + 315, + 482, + 390, + 495 + ], + "score": 1.0, + "content": "21(1):5485–5551.", + "type": "text" + } + ], + "index": 86 + } + ], + "index": 83.5 + }, + { + "type": "text", + "bbox": [ + 304, + 504, + 525, + 537 + ], + "lines": [ + { + "bbox": [ + 304, + 503, + 526, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 526, + 516 + ], + "score": 1.0, + "content": "James M Schuerger. 2000. The sixteen personality fac-", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 315, + 515, + 525, + 526 + ], + "spans": [ + { + "bbox": [ + 315, + 515, + 525, + 526 + ], + "score": 1.0, + "content": "tor questionnaire (16PF). Testing and assessment in", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 315, + 526, + 459, + 537 + ], + "spans": [ + { + "bbox": [ + 315, + 526, + 459, + 537 + ], + "score": 1.0, + "content": "counseling practice, pages 73–110.", + "type": "text" + } + ], + "index": 89 + } + ], + "index": 88 + }, + { + "type": "text", + "bbox": [ + 304, + 546, + 525, + 623 + ], + "lines": [ + { + "bbox": [ + 304, + 545, + 526, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 526, + 558 + ], + "score": 1.0, + "content": "Emily Sheng, Kai-Wei Chang, Prem Natarajan, and", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 315, + 557, + 527, + 569 + ], + "spans": [ + { + "bbox": [ + 315, + 557, + 527, + 569 + ], + "score": 1.0, + "content": "Nanyun Peng. 2019. The woman worked as a babysit-", + "type": "text" + } + ], + "index": 91 + }, + { + "bbox": [ + 314, + 567, + 527, + 580 + ], + "spans": [ + { + "bbox": [ + 314, + 567, + 527, + 580 + ], + "score": 1.0, + "content": "ter: On biases in language generation. In Proceed-", + "type": "text" + } + ], + "index": 92 + }, + { + "bbox": [ + 315, + 579, + 526, + 591 + ], + "spans": [ + { + "bbox": [ + 315, + 579, + 526, + 591 + ], + "score": 1.0, + "content": "ings of the 2019 Conference on Empirical Methods", + "type": "text" + } + ], + "index": 93 + }, + { + "bbox": [ + 315, + 590, + 527, + 602 + ], + "spans": [ + { + "bbox": [ + 315, + 590, + 527, + 602 + ], + "score": 1.0, + "content": "in Natural Language Processing and the 9th Inter-", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 315, + 601, + 527, + 613 + ], + "spans": [ + { + "bbox": [ + 315, + 601, + 527, + 613 + ], + "score": 1.0, + "content": "national Joint Conference on Natural Language Pro-", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 315, + 612, + 503, + 623 + ], + "spans": [ + { + "bbox": [ + 315, + 612, + 503, + 623 + ], + "score": 1.0, + "content": "cessing (EMNLP-IJCNLP), pages 3407–3412.", + "type": "text" + } + ], + "index": 96 + } + ], + "index": 93 + }, + { + "type": "text", + "bbox": [ + 304, + 633, + 525, + 676 + ], + "lines": [ + { + "bbox": [ + 303, + 631, + 528, + 645 + ], + "spans": [ + { + "bbox": [ + 303, + 631, + 528, + 645 + ], + "score": 1.0, + "content": "Eva AM van Dis, Johan Bollen, Willem Zuidema,", + "type": "text" + } + ], + "index": 97 + }, + { + "bbox": [ + 314, + 642, + 527, + 655 + ], + "spans": [ + { + "bbox": [ + 314, + 642, + 527, + 655 + ], + "score": 1.0, + "content": "Robert van Rooij, and Claudi L Bockting. 2023.", + "type": "text" + } + ], + "index": 98 + }, + { + "bbox": [ + 315, + 653, + 527, + 667 + ], + "spans": [ + { + "bbox": [ + 315, + 653, + 527, + 667 + ], + "score": 1.0, + "content": "ChatGPT: Five priorities for research. Nature,", + "type": "text" + } + ], + "index": 99 + }, + { + "bbox": [ + 315, + 665, + 400, + 677 + ], + "spans": [ + { + "bbox": [ + 315, + 665, + 400, + 677 + ], + "score": 1.0, + "content": "614(7947):224–226.", + "type": "text" + } + ], + "index": 100 + } + ], + "index": 98.5 + }, + { + "type": "text", + "bbox": [ + 304, + 686, + 525, + 741 + ], + "lines": [ + { + "bbox": [ + 304, + 685, + 526, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 685, + 526, + 698 + ], + "score": 1.0, + "content": "Laura Weidinger, John Mellor, Maribeth Rauh, Conor", + "type": "text" + } + ], + "index": 101 + }, + { + "bbox": [ + 315, + 696, + 526, + 709 + ], + "spans": [ + { + "bbox": [ + 315, + 696, + 526, + 709 + ], + "score": 1.0, + "content": "Griffin, Jonathan Uesato, Po-Sen Huang, Myra", + "type": "text" + } + ], + "index": 102 + }, + { + "bbox": [ + 315, + 708, + 527, + 720 + ], + "spans": [ + { + "bbox": [ + 315, + 708, + 527, + 720 + ], + "score": 1.0, + "content": "Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh,", + "type": "text" + } + ], + "index": 103 + }, + { + "bbox": [ + 315, + 718, + 526, + 731 + ], + "spans": [ + { + "bbox": [ + 315, + 718, + 526, + 731 + ], + "score": 1.0, + "content": "et al. 2021. Ethical and social risks of harm from", + "type": "text" + } + ], + "index": 104 + }, + { + "bbox": [ + 314, + 730, + 527, + 743 + ], + "spans": [ + { + "bbox": [ + 314, + 730, + 527, + 743 + ], + "score": 1.0, + "content": "language models. arXiv preprint arXiv:2112.04359.", + "type": "text" + } + ], + "index": 105 + } + ], + "index": 103 + }, + { + "type": "text", + "bbox": [ + 305, + 751, + 525, + 773 + ], + "lines": [ + { + "bbox": [ + 304, + 750, + 525, + 763 + ], + "spans": [ + { + "bbox": [ + 304, + 750, + 525, + 763 + ], + "score": 1.0, + "content": "Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei", + "type": "text" + } + ], + "index": 106 + }, + { + "bbox": [ + 315, + 761, + 527, + 775 + ], + "spans": [ + { + "bbox": [ + 315, + 761, + 527, + 775 + ], + "score": 1.0, + "content": "Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. 2022.", + "type": "text" + } + ], + "index": 107 + } + ], + "index": 106.5 + } + ], + "page_idx": 9, + "page_size": [ + 595, + 841 + ], + "discarded_blocks": [], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 80, + 73, + 289, + 95 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 79, + 70, + 291, + 95 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 70, + 103, + 289, + 168 + ], + "lines": [ + { + "bbox": [ + 69, + 102, + 291, + 114 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 291, + 114 + ], + "score": 1.0, + "content": "Sébastien Bubeck, Varun Chandrasekaran, Ronen El-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 113, + 291, + 126 + ], + "spans": [ + { + "bbox": [ + 79, + 113, + 291, + 126 + ], + "score": 1.0, + "content": "dan, Johannes Gehrke, Eric Horvitz, Ece Kamar,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 125, + 291, + 136 + ], + "spans": [ + { + "bbox": [ + 80, + 125, + 291, + 136 + ], + "score": 1.0, + "content": "Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lund-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 136, + 291, + 147 + ], + "spans": [ + { + "bbox": [ + 79, + 136, + 291, + 147 + ], + "score": 1.0, + "content": "berg, et al. 2023. Sparks of artificial general intelli-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 146, + 291, + 159 + ], + "spans": [ + { + "bbox": [ + 78, + 146, + 291, + 159 + ], + "score": 1.0, + "content": "gence: Early experiments with GPT-4. arXiv preprint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 158, + 158, + 169 + ], + "spans": [ + { + "bbox": [ + 79, + 158, + 158, + 169 + ], + "score": 1.0, + "content": "arXiv:2303.12712.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5, + "bbox_fs": [ + 69, + 102, + 291, + 169 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 177, + 290, + 210 + ], + "lines": [ + { + "bbox": [ + 69, + 177, + 291, + 189 + ], + "spans": [ + { + "bbox": [ + 69, + 177, + 291, + 189 + ], + "score": 1.0, + "content": "Graham Caron and Shashank Srivastava. 2022. Identi-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 79, + 188, + 291, + 200 + ], + "spans": [ + { + "bbox": [ + 79, + 188, + 291, + 200 + ], + "score": 1.0, + "content": "fying and manipulating the personality traits of lan-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 79, + 199, + 278, + 210 + ], + "spans": [ + { + "bbox": [ + 79, + 199, + 278, + 210 + ], + "score": 1.0, + "content": "guage models. arXiv preprint arXiv:2212.10276.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 69, + 177, + 291, + 210 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 218, + 289, + 295 + ], + "lines": [ + { + "bbox": [ + 68, + 218, + 290, + 230 + ], + "spans": [ + { + "bbox": [ + 68, + 218, + 290, + 230 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 79, + 228, + 291, + 241 + ], + "spans": [ + { + "bbox": [ + 79, + 228, + 291, + 241 + ], + "score": 1.0, + "content": "Kristina Toutanova. 2019. BERT: Pre-training of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 79, + 240, + 291, + 253 + ], + "spans": [ + { + "bbox": [ + 79, + 240, + 291, + 253 + ], + "score": 1.0, + "content": "deep bidirectional transformers for language under-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 80, + 251, + 290, + 263 + ], + "spans": [ + { + "bbox": [ + 80, + 251, + 290, + 263 + ], + "score": 1.0, + "content": "standing. In Proceedings of the 2019 Conference", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 80, + 262, + 290, + 274 + ], + "spans": [ + { + "bbox": [ + 80, + 262, + 290, + 274 + ], + "score": 1.0, + "content": "of the North American Chapter of the Association", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 78, + 272, + 291, + 286 + ], + "spans": [ + { + "bbox": [ + 78, + 272, + 291, + 286 + ], + "score": 1.0, + "content": "for Computational Linguistics: Human Language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 78, + 283, + 273, + 296 + ], + "spans": [ + { + "bbox": [ + 78, + 283, + 273, + 296 + ], + "score": 1.0, + "content": "Technologies (NAACL-HLT), pages 4171–4186.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14, + "bbox_fs": [ + 68, + 218, + 291, + 296 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 303, + 291, + 336 + ], + "lines": [ + { + "bbox": [ + 69, + 303, + 291, + 315 + ], + "spans": [ + { + "bbox": [ + 69, + 303, + 291, + 315 + ], + "score": 1.0, + "content": "John M Digman. 1990. Personality structure: Emer-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 78, + 313, + 292, + 327 + ], + "spans": [ + { + "bbox": [ + 78, + 313, + 292, + 327 + ], + "score": 1.0, + "content": "gence of the five-factor model. Annual review of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 78, + 325, + 194, + 337 + ], + "spans": [ + { + "bbox": [ + 78, + 325, + 194, + 337 + ], + "score": 1.0, + "content": "psychology, 41(1):417–440.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 69, + 303, + 292, + 337 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 345, + 290, + 366 + ], + "lines": [ + { + "bbox": [ + 68, + 343, + 292, + 358 + ], + "spans": [ + { + "bbox": [ + 68, + 343, + 292, + 358 + ], + "score": 1.0, + "content": "Hans Jurgen Eysenck. 2012. A model for personality.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 79, + 356, + 230, + 367 + ], + "spans": [ + { + "bbox": [ + 79, + 356, + 230, + 367 + ], + "score": 1.0, + "content": "Springer Science & Business Media.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 68, + 343, + 292, + 367 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 375, + 290, + 419 + ], + "lines": [ + { + "bbox": [ + 69, + 374, + 291, + 387 + ], + "spans": [ + { + "bbox": [ + 69, + 374, + 291, + 387 + ], + "score": 1.0, + "content": "Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 79, + 385, + 292, + 397 + ], + "spans": [ + { + "bbox": [ + 79, + 385, + 292, + 397 + ], + "score": 1.0, + "content": "Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2022.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 78, + 396, + 291, + 410 + ], + "spans": [ + { + "bbox": [ + 78, + 396, + 291, + 410 + ], + "score": 1.0, + "content": "Large language models can self-improve. arXiv", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 79, + 408, + 194, + 419 + ], + "spans": [ + { + "bbox": [ + 79, + 408, + 194, + 419 + ], + "score": 1.0, + "content": "preprint arXiv:2210.11610.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 69, + 374, + 292, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 427, + 290, + 471 + ], + "lines": [ + { + "bbox": [ + 68, + 426, + 292, + 439 + ], + "spans": [ + { + "bbox": [ + 68, + 426, + 292, + 439 + ], + "score": 1.0, + "content": "Guangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wen-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 78, + 437, + 292, + 450 + ], + "spans": [ + { + "bbox": [ + 78, + 437, + 292, + 450 + ], + "score": 1.0, + "content": "juan Han, Chi Zhang, and Yixin Zhu. 2022. MPI:", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 79, + 449, + 290, + 461 + ], + "spans": [ + { + "bbox": [ + 79, + 449, + 290, + 461 + ], + "score": 1.0, + "content": "Evaluating and inducing personality in pre-trained", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 79, + 460, + 291, + 472 + ], + "spans": [ + { + "bbox": [ + 79, + 460, + 291, + 472 + ], + "score": 1.0, + "content": "language models. arXiv preprint arXiv:2206.07550.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 68, + 426, + 292, + 472 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 479, + 289, + 523 + ], + "lines": [ + { + "bbox": [ + 69, + 479, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 69, + 479, + 290, + 491 + ], + "score": 1.0, + "content": "Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 79, + 489, + 290, + 503 + ], + "spans": [ + { + "bbox": [ + 79, + 489, + 290, + 503 + ], + "score": 1.0, + "content": "Neubig. 2020. How can we know what language", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 80, + 501, + 290, + 513 + ], + "spans": [ + { + "bbox": [ + 80, + 501, + 290, + 513 + ], + "score": 1.0, + "content": "models know? Transactions of the Association for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 80, + 512, + 239, + 524 + ], + "spans": [ + { + "bbox": [ + 80, + 512, + 239, + 524 + ], + "score": 1.0, + "content": "Computational Linguistics, 8:423–438.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 69, + 479, + 290, + 524 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 531, + 290, + 564 + ], + "lines": [ + { + "bbox": [ + 68, + 530, + 292, + 544 + ], + "spans": [ + { + "bbox": [ + 68, + 530, + 292, + 544 + ], + "score": 1.0, + "content": "Oliver P John, Sanjay Srivastava, et al. 1999. The big-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 80, + 543, + 291, + 554 + ], + "spans": [ + { + "bbox": [ + 80, + 543, + 291, + 554 + ], + "score": 1.0, + "content": "five trait taxonomy: History, measurement, and theo-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 80, + 554, + 162, + 565 + ], + "spans": [ + { + "bbox": [ + 80, + 554, + 162, + 565 + ], + "score": 1.0, + "content": "retical perspectives.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 68, + 530, + 292, + 565 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 573, + 290, + 616 + ], + "lines": [ + { + "bbox": [ + 69, + 572, + 291, + 585 + ], + "spans": [ + { + "bbox": [ + 69, + 572, + 291, + 585 + ], + "score": 1.0, + "content": "Saketh Reddy Karra, Son Nguyen, and Theja Tula-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 79, + 583, + 290, + 596 + ], + "spans": [ + { + "bbox": [ + 79, + 583, + 290, + 596 + ], + "score": 1.0, + "content": "bandhula. 2022. AI personification: Estimating", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 80, + 595, + 290, + 606 + ], + "spans": [ + { + "bbox": [ + 80, + 595, + 290, + 606 + ], + "score": 1.0, + "content": "the personality of language models. arXiv preprint", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 79, + 605, + 158, + 616 + ], + "spans": [ + { + "bbox": [ + 79, + 605, + 158, + 616 + ], + "score": 1.0, + "content": "arXiv:2204.12000.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 69, + 572, + 291, + 616 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 625, + 290, + 658 + ], + "lines": [ + { + "bbox": [ + 69, + 624, + 291, + 637 + ], + "spans": [ + { + "bbox": [ + 69, + 624, + 291, + 637 + ], + "score": 1.0, + "content": "Michal Kosinski. 2023. Theory of mind may have spon-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 79, + 636, + 290, + 648 + ], + "spans": [ + { + "bbox": [ + 79, + 636, + 290, + 648 + ], + "score": 1.0, + "content": "taneously emerged in large language models. arXiv", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 79, + 648, + 193, + 658 + ], + "spans": [ + { + "bbox": [ + 79, + 648, + 193, + 658 + ], + "score": 1.0, + "content": "preprint arXiv:2302.02083.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 69, + 624, + 291, + 658 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 666, + 290, + 710 + ], + "lines": [ + { + "bbox": [ + 68, + 665, + 291, + 679 + ], + "spans": [ + { + "bbox": [ + 68, + 665, + 291, + 679 + ], + "score": 1.0, + "content": "Xingxuan Li, Yutong Li, Linlin Liu, Lidong Bing, and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 80, + 677, + 291, + 689 + ], + "spans": [ + { + "bbox": [ + 80, + 677, + 291, + 689 + ], + "score": 1.0, + "content": "Shafiq Joty. 2022. Is GPT-3 a psychopath? evalu-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 80, + 688, + 291, + 700 + ], + "spans": [ + { + "bbox": [ + 80, + 688, + 291, + 700 + ], + "score": 1.0, + "content": "ating large language models from a psychological", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 79, + 699, + 268, + 711 + ], + "spans": [ + { + "bbox": [ + 79, + 699, + 268, + 711 + ], + "score": 1.0, + "content": "perspective. arXiv preprint arXiv:2212.10529.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5, + "bbox_fs": [ + 68, + 665, + 291, + 711 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 718, + 290, + 773 + ], + "lines": [ + { + "bbox": [ + 68, + 716, + 293, + 731 + ], + "spans": [ + { + "bbox": [ + 68, + 716, + 293, + 731 + ], + "score": 1.0, + "content": "Marilù Miotto, Nicola Rossberg, and Bennett Kleinberg.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 79, + 728, + 291, + 741 + ], + "spans": [ + { + "bbox": [ + 79, + 728, + 291, + 741 + ], + "score": 1.0, + "content": "2022. Who is GPT-3? An exploration of person-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 79, + 739, + 290, + 752 + ], + "spans": [ + { + "bbox": [ + 79, + 739, + 290, + 752 + ], + "score": 1.0, + "content": "ality, values and demographics. In Conference on", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 79, + 749, + 290, + 765 + ], + "spans": [ + { + "bbox": [ + 79, + 749, + 290, + 765 + ], + "score": 1.0, + "content": "Empirical Methods in Natural Language Processing", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 79, + 761, + 166, + 775 + ], + "spans": [ + { + "bbox": [ + 79, + 761, + 166, + 775 + ], + "score": 1.0, + "content": "(EMNLP) Workshop.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 51, + "bbox_fs": [ + 68, + 716, + 293, + 775 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 73, + 525, + 117 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 527, + 86 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 527, + 86 + ], + "score": 1.0, + "content": "Shima Rahimi Moghaddam and Christopher J Honey.", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 315, + 83, + 525, + 96 + ], + "spans": [ + { + "bbox": [ + 315, + 83, + 525, + 96 + ], + "score": 1.0, + "content": "2023. Boosting theory-of-mind performance in large", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 315, + 95, + 525, + 107 + ], + "spans": [ + { + "bbox": [ + 315, + 95, + 525, + 107 + ], + "score": 1.0, + "content": "language models via prompting. arXiv preprint", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 315, + 106, + 394, + 117 + ], + "spans": [ + { + "bbox": [ + 315, + 106, + 394, + 117 + ], + "score": 1.0, + "content": "arXiv:2304.11490.", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 55.5, + "bbox_fs": [ + 304, + 72, + 527, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 126, + 524, + 148 + ], + "lines": [ + { + "bbox": [ + 303, + 124, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 303, + 124, + 526, + 140 + ], + "score": 1.0, + "content": "Isabel Briggs Myers. 1962. The Myers-Briggs Type", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 315, + 137, + 422, + 149 + ], + "spans": [ + { + "bbox": [ + 315, + 137, + 422, + 149 + ], + "score": 1.0, + "content": "Indicator: Manual (1962).", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 58.5, + "bbox_fs": [ + 303, + 124, + 526, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 158, + 525, + 202 + ], + "lines": [ + { + "bbox": [ + 304, + 158, + 527, + 170 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 527, + 170 + ], + "score": 1.0, + "content": "Isabel Briggs Myers and Mary H. McCaulley. 1985.", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 315, + 169, + 525, + 181 + ], + "spans": [ + { + "bbox": [ + 315, + 169, + 525, + 181 + ], + "score": 1.0, + "content": "Manual: A guide to the development and use of the", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 315, + 180, + 526, + 192 + ], + "spans": [ + { + "bbox": [ + 315, + 180, + 526, + 192 + ], + "score": 1.0, + "content": "Myers-Briggs Type Indicator. Consulting Psycholo-", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 314, + 191, + 363, + 203 + ], + "spans": [ + { + "bbox": [ + 314, + 191, + 363, + 203 + ], + "score": 1.0, + "content": "gists Press.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 61.5, + "bbox_fs": [ + 304, + 158, + 527, + 203 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 212, + 525, + 288 + ], + "lines": [ + { + "bbox": [ + 304, + 212, + 527, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 527, + 224 + ], + "score": 1.0, + "content": "Moin Nadeem, Anna Bethke, and Siva Reddy. 2021.", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 315, + 223, + 526, + 235 + ], + "spans": [ + { + "bbox": [ + 315, + 223, + 526, + 235 + ], + "score": 1.0, + "content": "StereoSet: Measuring stereotypical bias in pretrained", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 315, + 234, + 527, + 245 + ], + "spans": [ + { + "bbox": [ + 315, + 234, + 527, + 245 + ], + "score": 1.0, + "content": "language models. In Proceedings of the Annual Meet-", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 315, + 244, + 526, + 257 + ], + "spans": [ + { + "bbox": [ + 315, + 244, + 526, + 257 + ], + "score": 1.0, + "content": "ing of the Association for Computational Linguistics", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 315, + 255, + 526, + 267 + ], + "spans": [ + { + "bbox": [ + 315, + 255, + 526, + 267 + ], + "score": 1.0, + "content": "and the International Joint Conference on Natural", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 314, + 266, + 527, + 279 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 527, + 279 + ], + "score": 1.0, + "content": "Language Processing (ACL-IJCNLP), pages 5356–", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 314, + 276, + 342, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 276, + 342, + 289 + ], + "score": 1.0, + "content": "5371.", + "type": "text" + } + ], + "index": 70 + } + ], + "index": 67, + "bbox_fs": [ + 304, + 212, + 527, + 289 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 298, + 525, + 365 + ], + "lines": [ + { + "bbox": [ + 304, + 298, + 527, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 527, + 311 + ], + "score": 1.0, + "content": "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida,", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 314, + 308, + 527, + 322 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 527, + 322 + ], + "score": 1.0, + "content": "Carroll Wainwright, Pamela Mishkin, Chong Zhang,", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 315, + 319, + 527, + 332 + ], + "spans": [ + { + "bbox": [ + 315, + 319, + 527, + 332 + ], + "score": 1.0, + "content": "Sandhini Agarwal, Katarina Slama, Alex Gray, et al.", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 314, + 330, + 528, + 344 + ], + "spans": [ + { + "bbox": [ + 314, + 330, + 528, + 344 + ], + "score": 1.0, + "content": "2022. Training language models to follow instruc-", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 315, + 342, + 526, + 353 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 526, + 353 + ], + "score": 1.0, + "content": "tions with human feedback. In Advances in Neural", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 315, + 353, + 492, + 365 + ], + "spans": [ + { + "bbox": [ + 315, + 353, + 492, + 365 + ], + "score": 1.0, + "content": "Information Processing Systems (NeurIPS).", + "type": "text" + } + ], + "index": 76 + } + ], + "index": 73.5, + "bbox_fs": [ + 304, + 298, + 528, + 365 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 374, + 525, + 418 + ], + "lines": [ + { + "bbox": [ + 304, + 373, + 527, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 527, + 386 + ], + "score": 1.0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan,", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 314, + 384, + 526, + 398 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 526, + 398 + ], + "score": 1.0, + "content": "Dario Amodei, Ilya Sutskever, et al. 2019. Language", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 316, + 396, + 525, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 525, + 407 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. OpenAI", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 315, + 406, + 368, + 419 + ], + "spans": [ + { + "bbox": [ + 315, + 406, + 368, + 419 + ], + "score": 1.0, + "content": "blog, 1(8):9.", + "type": "text" + } + ], + "index": 80 + } + ], + "index": 78.5, + "bbox_fs": [ + 304, + 373, + 527, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 428, + 525, + 493 + ], + "lines": [ + { + "bbox": [ + 304, + 427, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 525, + 439 + ], + "score": 1.0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine", + "type": "text" + } + ], + "index": 81 + }, + { + "bbox": [ + 313, + 437, + 527, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 527, + 452 + ], + "score": 1.0, + "content": "Lee, Sharan Narang, Michael Matena, Yanqi Zhou,", + "type": "text" + } + ], + "index": 82 + }, + { + "bbox": [ + 315, + 449, + 525, + 462 + ], + "spans": [ + { + "bbox": [ + 315, + 449, + 525, + 462 + ], + "score": 1.0, + "content": "Wei Li, and Peter J Liu. 2020. Exploring the limits", + "type": "text" + } + ], + "index": 83 + }, + { + "bbox": [ + 315, + 460, + 527, + 472 + ], + "spans": [ + { + "bbox": [ + 315, + 460, + 527, + 472 + ], + "score": 1.0, + "content": "of transfer learning with a unified text-to-text trans-", + "type": "text" + } + ], + "index": 84 + }, + { + "bbox": [ + 315, + 470, + 527, + 484 + ], + "spans": [ + { + "bbox": [ + 315, + 470, + 527, + 484 + ], + "score": 1.0, + "content": "former. The Journal of Machine Learning Research,", + "type": "text" + } + ], + "index": 85 + }, + { + "bbox": [ + 315, + 482, + 390, + 495 + ], + "spans": [ + { + "bbox": [ + 315, + 482, + 390, + 495 + ], + "score": 1.0, + "content": "21(1):5485–5551.", + "type": "text" + } + ], + "index": 86 + } + ], + "index": 83.5, + "bbox_fs": [ + 304, + 427, + 527, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 504, + 525, + 537 + ], + "lines": [ + { + "bbox": [ + 304, + 503, + 526, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 526, + 516 + ], + "score": 1.0, + "content": "James M Schuerger. 2000. The sixteen personality fac-", + "type": "text" + } + ], + "index": 87 + }, + { + "bbox": [ + 315, + 515, + 525, + 526 + ], + "spans": [ + { + "bbox": [ + 315, + 515, + 525, + 526 + ], + "score": 1.0, + "content": "tor questionnaire (16PF). Testing and assessment in", + "type": "text" + } + ], + "index": 88 + }, + { + "bbox": [ + 315, + 526, + 459, + 537 + ], + "spans": [ + { + "bbox": [ + 315, + 526, + 459, + 537 + ], + "score": 1.0, + "content": "counseling practice, pages 73–110.", + "type": "text" + } + ], + "index": 89 + } + ], + "index": 88, + "bbox_fs": [ + 304, + 503, + 526, + 537 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 546, + 525, + 623 + ], + "lines": [ + { + "bbox": [ + 304, + 545, + 526, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 526, + 558 + ], + "score": 1.0, + "content": "Emily Sheng, Kai-Wei Chang, Prem Natarajan, and", + "type": "text" + } + ], + "index": 90 + }, + { + "bbox": [ + 315, + 557, + 527, + 569 + ], + "spans": [ + { + "bbox": [ + 315, + 557, + 527, + 569 + ], + "score": 1.0, + "content": "Nanyun Peng. 2019. The woman worked as a babysit-", + "type": "text" + } + ], + "index": 91 + }, + { + "bbox": [ + 314, + 567, + 527, + 580 + ], + "spans": [ + { + "bbox": [ + 314, + 567, + 527, + 580 + ], + "score": 1.0, + "content": "ter: On biases in language generation. In Proceed-", + "type": "text" + } + ], + "index": 92 + }, + { + "bbox": [ + 315, + 579, + 526, + 591 + ], + "spans": [ + { + "bbox": [ + 315, + 579, + 526, + 591 + ], + "score": 1.0, + "content": "ings of the 2019 Conference on Empirical Methods", + "type": "text" + } + ], + "index": 93 + }, + { + "bbox": [ + 315, + 590, + 527, + 602 + ], + "spans": [ + { + "bbox": [ + 315, + 590, + 527, + 602 + ], + "score": 1.0, + "content": "in Natural Language Processing and the 9th Inter-", + "type": "text" + } + ], + "index": 94 + }, + { + "bbox": [ + 315, + 601, + 527, + 613 + ], + "spans": [ + { + "bbox": [ + 315, + 601, + 527, + 613 + ], + "score": 1.0, + "content": "national Joint Conference on Natural Language Pro-", + "type": "text" + } + ], + "index": 95 + }, + { + "bbox": [ + 315, + 612, + 503, + 623 + ], + "spans": [ + { + "bbox": [ + 315, + 612, + 503, + 623 + ], + "score": 1.0, + "content": "cessing (EMNLP-IJCNLP), pages 3407–3412.", + "type": "text" + } + ], + "index": 96 + } + ], + "index": 93, + "bbox_fs": [ + 304, + 545, + 527, + 623 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 633, + 525, + 676 + ], + "lines": [ + { + "bbox": [ + 303, + 631, + 528, + 645 + ], + "spans": [ + { + "bbox": [ + 303, + 631, + 528, + 645 + ], + "score": 1.0, + "content": "Eva AM van Dis, Johan Bollen, Willem Zuidema,", + "type": "text" + } + ], + "index": 97 + }, + { + "bbox": [ + 314, + 642, + 527, + 655 + ], + "spans": [ + { + "bbox": [ + 314, + 642, + 527, + 655 + ], + "score": 1.0, + "content": "Robert van Rooij, and Claudi L Bockting. 2023.", + "type": "text" + } + ], + "index": 98 + }, + { + "bbox": [ + 315, + 653, + 527, + 667 + ], + "spans": [ + { + "bbox": [ + 315, + 653, + 527, + 667 + ], + "score": 1.0, + "content": "ChatGPT: Five priorities for research. Nature,", + "type": "text" + } + ], + "index": 99 + }, + { + "bbox": [ + 315, + 665, + 400, + 677 + ], + "spans": [ + { + "bbox": [ + 315, + 665, + 400, + 677 + ], + "score": 1.0, + "content": "614(7947):224–226.", + "type": "text" + } + ], + "index": 100 + } + ], + "index": 98.5, + "bbox_fs": [ + 303, + 631, + 528, + 677 + ] + }, + { + "type": "text", + "bbox": [ + 304, + 686, + 525, + 741 + ], + "lines": [ + { + "bbox": [ + 304, + 685, + 526, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 685, + 526, + 698 + ], + "score": 1.0, + "content": "Laura Weidinger, John Mellor, Maribeth Rauh, Conor", + "type": "text" + } + ], + "index": 101 + }, + { + "bbox": [ + 315, + 696, + 526, + 709 + ], + "spans": [ + { + "bbox": [ + 315, + 696, + 526, + 709 + ], + "score": 1.0, + "content": "Griffin, Jonathan Uesato, Po-Sen Huang, Myra", + "type": "text" + } + ], + "index": 102 + }, + { + "bbox": [ + 315, + 708, + 527, + 720 + ], + "spans": [ + { + "bbox": [ + 315, + 708, + 527, + 720 + ], + "score": 1.0, + "content": "Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh,", + "type": "text" + } + ], + "index": 103 + }, + { + "bbox": [ + 315, + 718, + 526, + 731 + ], + "spans": [ + { + "bbox": [ + 315, + 718, + 526, + 731 + ], + "score": 1.0, + "content": "et al. 2021. Ethical and social risks of harm from", + "type": "text" + } + ], + "index": 104 + }, + { + "bbox": [ + 314, + 730, + 527, + 743 + ], + "spans": [ + { + "bbox": [ + 314, + 730, + 527, + 743 + ], + "score": 1.0, + "content": "language models. arXiv preprint arXiv:2112.04359.", + "type": "text" + } + ], + "index": 105 + } + ], + "index": 103, + "bbox_fs": [ + 304, + 685, + 527, + 743 + ] + }, + { + "type": "text", + "bbox": [ + 305, + 751, + 525, + 773 + ], + "lines": [ + { + "bbox": [ + 304, + 750, + 525, + 763 + ], + "spans": [ + { + "bbox": [ + 304, + 750, + 525, + 763 + ], + "score": 1.0, + "content": "Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei", + "type": "text" + } + ], + "index": 106 + }, + { + "bbox": [ + 315, + 761, + 527, + 775 + ], + "spans": [ + { + "bbox": [ + 315, + 761, + 527, + 775 + ], + "score": 1.0, + "content": "Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. 2022.", + "type": "text" + } + ], + "index": 107 + } + ], + "index": 106.5, + "bbox_fs": [ + 304, + 750, + 527, + 775 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 80, + 72, + 290, + 116 + ], + "lines": [ + { + "bbox": [ + 79, + 72, + 291, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 72, + 291, + 84 + ], + "score": 1.0, + "content": "An empirical study of GPT-3 for few-shot knowledge-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 79, + 83, + 290, + 96 + ], + "spans": [ + { + "bbox": [ + 79, + 83, + 290, + 96 + ], + "score": 1.0, + "content": "based VQA. In Proceedings of the AAAI Conference", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 79, + 94, + 290, + 107 + ], + "spans": [ + { + "bbox": [ + 79, + 94, + 290, + 107 + ], + "score": 1.0, + "content": "on Artificial Intelligence (AAAI), volume 36, pages", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 105, + 131, + 117 + ], + "spans": [ + { + "bbox": [ + 79, + 105, + 131, + 117 + ], + "score": 1.0, + "content": "3081–3089.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 69, + 125, + 290, + 169 + ], + "lines": [ + { + "bbox": [ + 69, + 124, + 291, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 124, + 291, + 138 + ], + "score": 1.0, + "content": "Ann Yuan, Andy Coenen, Emily Reif, and Daphne Ip-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 135, + 290, + 149 + ], + "spans": [ + { + "bbox": [ + 79, + 135, + 290, + 149 + ], + "score": 1.0, + "content": "polito. 2022. Wordcraft: Story writing with large", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 147, + 290, + 159 + ], + "spans": [ + { + "bbox": [ + 79, + 147, + 290, + 159 + ], + "score": 1.0, + "content": "language models. In 27th International Conference", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 158, + 230, + 169 + ], + "spans": [ + { + "bbox": [ + 80, + 158, + 230, + 169 + ], + "score": 1.0, + "content": "on Intelligent User Interfaces. ACM.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 69, + 178, + 290, + 200 + ], + "lines": [ + { + "bbox": [ + 69, + 178, + 291, + 190 + ], + "spans": [ + { + "bbox": [ + 69, + 178, + 291, + 190 + ], + "score": 1.0, + "content": "Xiaoming Zhai. 2022. ChatGPT user experience: Impli-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 79, + 189, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 79, + 189, + 286, + 201 + ], + "score": 1.0, + "content": "cations for education. Available at SSRN 4312418.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 69, + 209, + 290, + 264 + ], + "lines": [ + { + "bbox": [ + 69, + 209, + 290, + 221 + ], + "spans": [ + { + "bbox": [ + 69, + 209, + 290, + 221 + ], + "score": 1.0, + "content": "Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 219, + 292, + 232 + ], + "spans": [ + { + "bbox": [ + 79, + 219, + 292, + 232 + ], + "score": 1.0, + "content": "Sameer Singh. 2021. Calibrate before use: Im-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 79, + 230, + 292, + 243 + ], + "spans": [ + { + "bbox": [ + 79, + 230, + 292, + 243 + ], + "score": 1.0, + "content": "proving few-shot performance of language models.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 79, + 241, + 290, + 255 + ], + "spans": [ + { + "bbox": [ + 79, + 241, + 290, + 255 + ], + "score": 1.0, + "content": "In International Conference on Machine Learning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 79, + 253, + 232, + 264 + ], + "spans": [ + { + "bbox": [ + 79, + 253, + 232, + 264 + ], + "score": 1.0, + "content": "(ICML), pages 12697–12706. PMLR.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 70, + 273, + 290, + 317 + ], + "lines": [ + { + "bbox": [ + 69, + 273, + 290, + 285 + ], + "spans": [ + { + "bbox": [ + 69, + 273, + 290, + 285 + ], + "score": 1.0, + "content": "Terry Yue Zhuo, Yujin Huang, Chunyang Chen, and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 79, + 283, + 291, + 296 + ], + "spans": [ + { + "bbox": [ + 79, + 283, + 291, + 296 + ], + "score": 1.0, + "content": "Zhenchang Xing. 2023. Exploring AI ethics of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 294, + 291, + 308 + ], + "spans": [ + { + "bbox": [ + 79, + 294, + 291, + 308 + ], + "score": 1.0, + "content": "ChatGPT: A diagnostic analysis. arXiv preprint", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 79, + 306, + 159, + 317 + ], + "spans": [ + { + "bbox": [ + 79, + 306, + 159, + 317 + ], + "score": 1.0, + "content": "arXiv:2301.12867.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 69, + 326, + 290, + 348 + ], + "lines": [ + { + "bbox": [ + 69, + 325, + 291, + 338 + ], + "spans": [ + { + "bbox": [ + 69, + 325, + 291, + 338 + ], + "score": 1.0, + "content": "Mingyu Zong and Bhaskar Krishnamachari. 2022. A", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 79, + 337, + 291, + 349 + ], + "spans": [ + { + "bbox": [ + 79, + 337, + 291, + 349 + ], + "score": 1.0, + "content": "survey on GPT-3. arXiv preprint arXiv:2212.00857.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + } + ], + "page_idx": 10, + "page_size": [ + 595, + 841 + ], + "discarded_blocks": [], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 80, + 72, + 290, + 116 + ], + "lines": [ + { + "bbox": [ + 79, + 72, + 291, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 72, + 291, + 84 + ], + "score": 1.0, + "content": "An empirical study of GPT-3 for few-shot knowledge-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 79, + 83, + 290, + 96 + ], + "spans": [ + { + "bbox": [ + 79, + 83, + 290, + 96 + ], + "score": 1.0, + "content": "based VQA. In Proceedings of the AAAI Conference", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 79, + 94, + 290, + 107 + ], + "spans": [ + { + "bbox": [ + 79, + 94, + 290, + 107 + ], + "score": 1.0, + "content": "on Artificial Intelligence (AAAI), volume 36, pages", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 105, + 131, + 117 + ], + "spans": [ + { + "bbox": [ + 79, + 105, + 131, + 117 + ], + "score": 1.0, + "content": "3081–3089.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 79, + 72, + 291, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 125, + 290, + 169 + ], + "lines": [ + { + "bbox": [ + 69, + 124, + 291, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 124, + 291, + 138 + ], + "score": 1.0, + "content": "Ann Yuan, Andy Coenen, Emily Reif, and Daphne Ip-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 135, + 290, + 149 + ], + "spans": [ + { + "bbox": [ + 79, + 135, + 290, + 149 + ], + "score": 1.0, + "content": "polito. 2022. Wordcraft: Story writing with large", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 147, + 290, + 159 + ], + "spans": [ + { + "bbox": [ + 79, + 147, + 290, + 159 + ], + "score": 1.0, + "content": "language models. In 27th International Conference", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 158, + 230, + 169 + ], + "spans": [ + { + "bbox": [ + 80, + 158, + 230, + 169 + ], + "score": 1.0, + "content": "on Intelligent User Interfaces. ACM.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 69, + 124, + 291, + 169 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 178, + 290, + 200 + ], + "lines": [ + { + "bbox": [ + 69, + 178, + 291, + 190 + ], + "spans": [ + { + "bbox": [ + 69, + 178, + 291, + 190 + ], + "score": 1.0, + "content": "Xiaoming Zhai. 2022. ChatGPT user experience: Impli-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 79, + 189, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 79, + 189, + 286, + 201 + ], + "score": 1.0, + "content": "cations for education. Available at SSRN 4312418.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 69, + 178, + 291, + 201 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 209, + 290, + 264 + ], + "lines": [ + { + "bbox": [ + 69, + 209, + 290, + 221 + ], + "spans": [ + { + "bbox": [ + 69, + 209, + 290, + 221 + ], + "score": 1.0, + "content": "Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 219, + 292, + 232 + ], + "spans": [ + { + "bbox": [ + 79, + 219, + 292, + 232 + ], + "score": 1.0, + "content": "Sameer Singh. 2021. Calibrate before use: Im-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 79, + 230, + 292, + 243 + ], + "spans": [ + { + "bbox": [ + 79, + 230, + 292, + 243 + ], + "score": 1.0, + "content": "proving few-shot performance of language models.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 79, + 241, + 290, + 255 + ], + "spans": [ + { + "bbox": [ + 79, + 241, + 290, + 255 + ], + "score": 1.0, + "content": "In International Conference on Machine Learning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 79, + 253, + 232, + 264 + ], + "spans": [ + { + "bbox": [ + 79, + 253, + 232, + 264 + ], + "score": 1.0, + "content": "(ICML), pages 12697–12706. PMLR.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 69, + 209, + 292, + 264 + ] + }, + { + "type": "text", + "bbox": [ + 70, + 273, + 290, + 317 + ], + "lines": [ + { + "bbox": [ + 69, + 273, + 290, + 285 + ], + "spans": [ + { + "bbox": [ + 69, + 273, + 290, + 285 + ], + "score": 1.0, + "content": "Terry Yue Zhuo, Yujin Huang, Chunyang Chen, and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 79, + 283, + 291, + 296 + ], + "spans": [ + { + "bbox": [ + 79, + 283, + 291, + 296 + ], + "score": 1.0, + "content": "Zhenchang Xing. 2023. Exploring AI ethics of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 294, + 291, + 308 + ], + "spans": [ + { + "bbox": [ + 79, + 294, + 291, + 308 + ], + "score": 1.0, + "content": "ChatGPT: A diagnostic analysis. arXiv preprint", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 79, + 306, + 159, + 317 + ], + "spans": [ + { + "bbox": [ + 79, + 306, + 159, + 317 + ], + "score": 1.0, + "content": "arXiv:2301.12867.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 69, + 273, + 291, + 317 + ] + }, + { + "type": "text", + "bbox": [ + 69, + 326, + 290, + 348 + ], + "lines": [ + { + "bbox": [ + 69, + 325, + 291, + 338 + ], + "spans": [ + { + "bbox": [ + 69, + 325, + 291, + 338 + ], + "score": 1.0, + "content": "Mingyu Zong and Bhaskar Krishnamachari. 2022. A", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 79, + 337, + 291, + 349 + ], + "spans": [ + { + "bbox": [ + 79, + 337, + 291, + 349 + ], + "score": 1.0, + "content": "survey on GPT-3. arXiv preprint arXiv:2212.00857.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 69, + 325, + 291, + 349 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file