{ "pdf_info": [ { "preproc_blocks": [ { "type": "title", "bbox": [ 135, 97, 474, 133 ], "lines": [ { "bbox": [ 138, 96, 477, 117 ], "spans": [ { "bbox": [ 138, 96, 477, 117 ], "score": 1.0, "content": "CAMEL: Communicative Agents for “Mind”", "type": "text" } ], "index": 0 }, { "bbox": [ 132, 114, 475, 135 ], "spans": [ { "bbox": [ 132, 114, 475, 135 ], "score": 1.0, "content": "Exploration of Large Language Model Society", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 229, 136, 379, 149 ], "lines": [ { "bbox": [ 227, 133, 381, 151 ], "spans": [ { "bbox": [ 227, 133, 381, 151 ], "score": 1.0, "content": "https://www.camel-ai.org", "type": "text" } ], "index": 2 } ], "index": 2 }, { "type": "title", "bbox": [ 119, 181, 326, 193 ], "lines": [ { "bbox": [ 117, 180, 327, 195 ], "spans": [ { "bbox": [ 117, 180, 327, 195 ], "score": 1.0, "content": "Guohao Li⇤ Hasan Abed Al Kader Hammoud", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 344, 181, 493, 193 ], "lines": [ { "bbox": [ 344, 181, 493, 195 ], "spans": [ { "bbox": [ 344, 181, 396, 195 ], "score": 1.0, "content": "Hani Itani*", "type": "text" }, { "bbox": [ 410, 181, 493, 195 ], "score": 1.0, "content": "Dmitrii Khizbullin", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "title", "bbox": [ 266, 210, 342, 221 ], "lines": [ { "bbox": [ 265, 208, 343, 223 ], "spans": [ { "bbox": [ 265, 208, 343, 223 ], "score": 1.0, "content": "Bernard Ghanem", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 177, 229, 434, 240 ], "lines": [ { "bbox": [ 175, 226, 435, 244 ], "spans": [ { "bbox": [ 175, 226, 435, 244 ], "score": 1.0, "content": "King Abdullah University of Science and Technology (KAUST)", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "title", "bbox": [ 283, 268, 328, 281 ], "lines": [ { "bbox": [ 282, 268, 330, 282 ], "spans": [ { "bbox": [ 282, 268, 330, 282 ], "score": 1.0, "content": "Abstract", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 143, 293, 469, 478 ], "lines": [ { "bbox": [ 142, 293, 469, 305 ], "spans": [ { "bbox": [ 142, 293, 469, 305 ], "score": 1.0, "content": "The rapid advancement of chat-based language models has led to remarkable", "type": "text" } ], "index": 8 }, { "bbox": [ 140, 303, 470, 317 ], "spans": [ { "bbox": [ 140, 303, 470, 317 ], "score": 1.0, "content": "progress in complex task-solving. However, their success heavily relies on human", "type": "text" } ], "index": 9 }, { "bbox": [ 141, 314, 471, 328 ], "spans": [ { "bbox": [ 141, 314, 471, 328 ], "score": 1.0, "content": "input to guide the conversation, which can be challenging and time-consuming.", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 325, 471, 338 ], "spans": [ { "bbox": [ 141, 325, 471, 338 ], "score": 1.0, "content": "This paper explores the potential of building scalable techniques to facilitate au-", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 337, 469, 349 ], "spans": [ { "bbox": [ 141, 337, 469, 349 ], "score": 1.0, "content": "tonomous cooperation among communicative agents, and provides insight into", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 348, 469, 360 ], "spans": [ { "bbox": [ 141, 348, 469, 360 ], "score": 1.0, "content": "their “cognitive” processes. To address the challenges of achieving autonomous", "type": "text" } ], "index": 13 }, { "bbox": [ 141, 360, 470, 370 ], "spans": [ { "bbox": [ 141, 360, 470, 370 ], "score": 1.0, "content": "cooperation, we propose a novel communicative agent framework named role-", "type": "text" } ], "index": 14 }, { "bbox": [ 142, 370, 469, 382 ], "spans": [ { "bbox": [ 142, 370, 469, 382 ], "score": 1.0, "content": "playing . Our approach involves using inception prompting to guide chat agents", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 380, 469, 392 ], "spans": [ { "bbox": [ 141, 380, 469, 392 ], "score": 1.0, "content": "toward task completion while maintaining consistency with human intentions. We", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 390, 470, 405 ], "spans": [ { "bbox": [ 141, 390, 470, 405 ], "score": 1.0, "content": "showcase how role-playing can be used to generate conversational data for studying", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 402, 470, 415 ], "spans": [ { "bbox": [ 141, 402, 470, 415 ], "score": 1.0, "content": "the behaviors and capabilities of a society of agents, providing a valuable resource", "type": "text" } ], "index": 18 }, { "bbox": [ 141, 413, 471, 426 ], "spans": [ { "bbox": [ 141, 413, 471, 426 ], "score": 1.0, "content": "for investigating conversational language models. In particular, we conduct com-", "type": "text" } ], "index": 19 }, { "bbox": [ 141, 423, 471, 437 ], "spans": [ { "bbox": [ 141, 423, 471, 437 ], "score": 1.0, "content": "prehensive studies on instruction-following cooperation in multi-agent settings.", "type": "text" } ], "index": 20 }, { "bbox": [ 141, 434, 471, 447 ], "spans": [ { "bbox": [ 141, 434, 471, 447 ], "score": 1.0, "content": "Our contributions include introducing a novel communicative agent framework,", "type": "text" } ], "index": 21 }, { "bbox": [ 141, 446, 470, 459 ], "spans": [ { "bbox": [ 141, 446, 470, 459 ], "score": 1.0, "content": "offering a scalable approach for studying the cooperative behaviors and capabili-", "type": "text" } ], "index": 22 }, { "bbox": [ 141, 457, 469, 469 ], "spans": [ { "bbox": [ 141, 457, 469, 469 ], "score": 1.0, "content": "ties of multi-agent systems, and open-sourcing our library to support research on", "type": "text" } ], "index": 23 }, { "bbox": [ 141, 468, 463, 479 ], "spans": [ { "bbox": [ 141, 468, 463, 479 ], "score": 1.0, "content": "communicative agents and beyond: https://github.com/camel-ai/camel.", "type": "text" } ], "index": 24 } ], "index": 16 }, { "type": "title", "bbox": [ 107, 497, 190, 511 ], "lines": [ { "bbox": [ 105, 496, 192, 514 ], "spans": [ { "bbox": [ 105, 496, 192, 514 ], "score": 1.0, "content": "1 Introduction", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 105, 522, 503, 545 ], "lines": [ { "bbox": [ 103, 520, 505, 536 ], "spans": [ { "bbox": [ 103, 520, 505, 536 ], "score": 1.0, "content": "“What magical trick makes us intelligent? The trick is that there is no trick. The power of intelligence", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 532, 381, 546 ], "spans": [ { "bbox": [ 105, 532, 381, 546 ], "score": 1.0, "content": "stems from our vast diversity, not from any single, perfect principle.”", "type": "text" } ], "index": 27 } ], "index": 26.5 }, { "type": "text", "bbox": [ 326, 550, 503, 561 ], "lines": [ { "bbox": [ 326, 549, 505, 562 ], "spans": [ { "bbox": [ 326, 549, 505, 562 ], "score": 1.0, "content": "- Marvin Minsky, The Society of Mind, p. 308", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 107, 566, 505, 697 ], "lines": [ { "bbox": [ 105, 564, 506, 579 ], "spans": [ { "bbox": [ 105, 564, 506, 579 ], "score": 1.0, "content": "Confronted with the complexities of real-world tasks, solving them often requires multiple steps.", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 576, 505, 590 ], "spans": [ { "bbox": [ 105, 576, 505, 590 ], "score": 1.0, "content": "The rapid progress of chat-based large-scale language models (LLMs) has yielded remarkable", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 587, 505, 601 ], "spans": [ { "bbox": [ 105, 587, 505, 601 ], "score": 1.0, "content": "achievements in complex task-solving [82, 84, 116, 89, 5, 10, 122, 13]. Nevertheless, it is worth", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 599, 505, 611 ], "spans": [ { "bbox": [ 106, 599, 505, 611 ], "score": 1.0, "content": "noting that their success is heavily reliant on human input to guide the conversation in the right", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 609, 505, 622 ], "spans": [ { "bbox": [ 105, 609, 505, 622 ], "score": 1.0, "content": "direction. This reliance necessitates users to provide relevant and precise prompts based on their", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 621, 505, 633 ], "spans": [ { "bbox": [ 105, 621, 505, 633 ], "score": 1.0, "content": "intentions and the chat agent’s feedback. This can be challenging, time-consuming, and sometimes", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 631, 506, 645 ], "spans": [ { "bbox": [ 106, 631, 506, 645 ], "score": 1.0, "content": "impossible. Crafting effective prompts often demands a deep understanding and expertise of a", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 642, 506, 655 ], "spans": [ { "bbox": [ 105, 642, 506, 655 ], "score": 1.0, "content": "particular domain of knowledge. Consider an individual who lacks trading expertise; they would find", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 654, 505, 666 ], "spans": [ { "bbox": [ 106, 654, 505, 666 ], "score": 1.0, "content": "it difficult to create suitable prompts for directing a chat agent to develop a trading application. This", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 664, 505, 677 ], "spans": [ { "bbox": [ 105, 664, 505, 677 ], "score": 1.0, "content": "predicament is raising a crucial question: can we replace human intervention with an autonomous", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 676, 505, 686 ], "spans": [ { "bbox": [ 106, 676, 505, 686 ], "score": 1.0, "content": "communicative agent capable of steering the conversation toward task completion with minimal", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 686, 506, 699 ], "spans": [ { "bbox": [ 106, 686, 506, 699 ], "score": 1.0, "content": "human supervision? To tackle this issue, it is crucial to conduct more research exploring the potential,", "type": "text" } ], "index": 40 } ], "index": 34.5 } ], "page_idx": 0, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 732, 385, 742 ], "lines": [ { "bbox": [ 105, 730, 386, 744 ], "spans": [ { "bbox": [ 105, 730, 386, 744 ], "score": 1.0, "content": "37th Conference on Neural Information Processing Systems (NeurIPS 2023).", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 119, 704, 190, 714 ], "lines": [ { "bbox": [ 118, 703, 192, 716 ], "spans": [ { "bbox": [ 118, 703, 192, 716 ], "score": 1.0, "content": "⇤Equal contribution", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 135, 97, 474, 133 ], "lines": [ { "bbox": [ 138, 96, 477, 117 ], "spans": [ { "bbox": [ 138, 96, 477, 117 ], "score": 1.0, "content": "CAMEL: Communicative Agents for “Mind”", "type": "text" } ], "index": 0 }, { "bbox": [ 132, 114, 475, 135 ], "spans": [ { "bbox": [ 132, 114, 475, 135 ], "score": 1.0, "content": "Exploration of Large Language Model Society", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 229, 136, 379, 149 ], "lines": [ { "bbox": [ 227, 133, 381, 151 ], "spans": [ { "bbox": [ 227, 133, 381, 151 ], "score": 1.0, "content": "https://www.camel-ai.org", "type": "text" } ], "index": 2 } ], "index": 2, "bbox_fs": [ 227, 133, 381, 151 ] }, { "type": "title", "bbox": [ 119, 181, 326, 193 ], "lines": [ { "bbox": [ 117, 180, 327, 195 ], "spans": [ { "bbox": [ 117, 180, 327, 195 ], "score": 1.0, "content": "Guohao Li⇤ Hasan Abed Al Kader Hammoud", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 344, 181, 493, 193 ], "lines": [ { "bbox": [ 344, 181, 493, 195 ], "spans": [ { "bbox": [ 344, 181, 396, 195 ], "score": 1.0, "content": "Hani Itani*", "type": "text" }, { "bbox": [ 410, 181, 493, 195 ], "score": 1.0, "content": "Dmitrii Khizbullin", "type": "text" } ], "index": 4 } ], "index": 4, "bbox_fs": [ 344, 181, 493, 195 ] }, { "type": "title", "bbox": [ 266, 210, 342, 221 ], "lines": [ { "bbox": [ 265, 208, 343, 223 ], "spans": [ { "bbox": [ 265, 208, 343, 223 ], "score": 1.0, "content": "Bernard Ghanem", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 177, 229, 434, 240 ], "lines": [ { "bbox": [ 175, 226, 435, 244 ], "spans": [ { "bbox": [ 175, 226, 435, 244 ], "score": 1.0, "content": "King Abdullah University of Science and Technology (KAUST)", "type": "text" } ], "index": 6 } ], "index": 6, "bbox_fs": [ 175, 226, 435, 244 ] }, { "type": "title", "bbox": [ 283, 268, 328, 281 ], "lines": [ { "bbox": [ 282, 268, 330, 282 ], "spans": [ { "bbox": [ 282, 268, 330, 282 ], "score": 1.0, "content": "Abstract", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 143, 293, 469, 478 ], "lines": [ { "bbox": [ 142, 293, 469, 305 ], "spans": [ { "bbox": [ 142, 293, 469, 305 ], "score": 1.0, "content": "The rapid advancement of chat-based language models has led to remarkable", "type": "text" } ], "index": 8 }, { "bbox": [ 140, 303, 470, 317 ], "spans": [ { "bbox": [ 140, 303, 470, 317 ], "score": 1.0, "content": "progress in complex task-solving. However, their success heavily relies on human", "type": "text" } ], "index": 9 }, { "bbox": [ 141, 314, 471, 328 ], "spans": [ { "bbox": [ 141, 314, 471, 328 ], "score": 1.0, "content": "input to guide the conversation, which can be challenging and time-consuming.", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 325, 471, 338 ], "spans": [ { "bbox": [ 141, 325, 471, 338 ], "score": 1.0, "content": "This paper explores the potential of building scalable techniques to facilitate au-", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 337, 469, 349 ], "spans": [ { "bbox": [ 141, 337, 469, 349 ], "score": 1.0, "content": "tonomous cooperation among communicative agents, and provides insight into", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 348, 469, 360 ], "spans": [ { "bbox": [ 141, 348, 469, 360 ], "score": 1.0, "content": "their “cognitive” processes. To address the challenges of achieving autonomous", "type": "text" } ], "index": 13 }, { "bbox": [ 141, 360, 470, 370 ], "spans": [ { "bbox": [ 141, 360, 470, 370 ], "score": 1.0, "content": "cooperation, we propose a novel communicative agent framework named role-", "type": "text" } ], "index": 14 }, { "bbox": [ 142, 370, 469, 382 ], "spans": [ { "bbox": [ 142, 370, 469, 382 ], "score": 1.0, "content": "playing . Our approach involves using inception prompting to guide chat agents", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 380, 469, 392 ], "spans": [ { "bbox": [ 141, 380, 469, 392 ], "score": 1.0, "content": "toward task completion while maintaining consistency with human intentions. We", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 390, 470, 405 ], "spans": [ { "bbox": [ 141, 390, 470, 405 ], "score": 1.0, "content": "showcase how role-playing can be used to generate conversational data for studying", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 402, 470, 415 ], "spans": [ { "bbox": [ 141, 402, 470, 415 ], "score": 1.0, "content": "the behaviors and capabilities of a society of agents, providing a valuable resource", "type": "text" } ], "index": 18 }, { "bbox": [ 141, 413, 471, 426 ], "spans": [ { "bbox": [ 141, 413, 471, 426 ], "score": 1.0, "content": "for investigating conversational language models. In particular, we conduct com-", "type": "text" } ], "index": 19 }, { "bbox": [ 141, 423, 471, 437 ], "spans": [ { "bbox": [ 141, 423, 471, 437 ], "score": 1.0, "content": "prehensive studies on instruction-following cooperation in multi-agent settings.", "type": "text" } ], "index": 20 }, { "bbox": [ 141, 434, 471, 447 ], "spans": [ { "bbox": [ 141, 434, 471, 447 ], "score": 1.0, "content": "Our contributions include introducing a novel communicative agent framework,", "type": "text" } ], "index": 21 }, { "bbox": [ 141, 446, 470, 459 ], "spans": [ { "bbox": [ 141, 446, 470, 459 ], "score": 1.0, "content": "offering a scalable approach for studying the cooperative behaviors and capabili-", "type": "text" } ], "index": 22 }, { "bbox": [ 141, 457, 469, 469 ], "spans": [ { "bbox": [ 141, 457, 469, 469 ], "score": 1.0, "content": "ties of multi-agent systems, and open-sourcing our library to support research on", "type": "text" } ], "index": 23 }, { "bbox": [ 141, 468, 463, 479 ], "spans": [ { "bbox": [ 141, 468, 463, 479 ], "score": 1.0, "content": "communicative agents and beyond: https://github.com/camel-ai/camel.", "type": "text" } ], "index": 24 } ], "index": 16, "bbox_fs": [ 140, 293, 471, 479 ] }, { "type": "title", "bbox": [ 107, 497, 190, 511 ], "lines": [ { "bbox": [ 105, 496, 192, 514 ], "spans": [ { "bbox": [ 105, 496, 192, 514 ], "score": 1.0, "content": "1 Introduction", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 105, 522, 503, 545 ], "lines": [ { "bbox": [ 103, 520, 505, 536 ], "spans": [ { "bbox": [ 103, 520, 505, 536 ], "score": 1.0, "content": "“What magical trick makes us intelligent? The trick is that there is no trick. The power of intelligence", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 532, 381, 546 ], "spans": [ { "bbox": [ 105, 532, 381, 546 ], "score": 1.0, "content": "stems from our vast diversity, not from any single, perfect principle.”", "type": "text" } ], "index": 27 } ], "index": 26.5, "bbox_fs": [ 103, 520, 505, 546 ] }, { "type": "text", "bbox": [ 326, 550, 503, 561 ], "lines": [ { "bbox": [ 326, 549, 505, 562 ], "spans": [ { "bbox": [ 326, 549, 505, 562 ], "score": 1.0, "content": "- Marvin Minsky, The Society of Mind, p. 308", "type": "text" } ], "index": 28 } ], "index": 28, "bbox_fs": [ 326, 549, 505, 562 ] }, { "type": "text", "bbox": [ 107, 566, 505, 697 ], "lines": [ { "bbox": [ 105, 564, 506, 579 ], "spans": [ { "bbox": [ 105, 564, 506, 579 ], "score": 1.0, "content": "Confronted with the complexities of real-world tasks, solving them often requires multiple steps.", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 576, 505, 590 ], "spans": [ { "bbox": [ 105, 576, 505, 590 ], "score": 1.0, "content": "The rapid progress of chat-based large-scale language models (LLMs) has yielded remarkable", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 587, 505, 601 ], "spans": [ { "bbox": [ 105, 587, 505, 601 ], "score": 1.0, "content": "achievements in complex task-solving [82, 84, 116, 89, 5, 10, 122, 13]. Nevertheless, it is worth", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 599, 505, 611 ], "spans": [ { "bbox": [ 106, 599, 505, 611 ], "score": 1.0, "content": "noting that their success is heavily reliant on human input to guide the conversation in the right", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 609, 505, 622 ], "spans": [ { "bbox": [ 105, 609, 505, 622 ], "score": 1.0, "content": "direction. This reliance necessitates users to provide relevant and precise prompts based on their", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 621, 505, 633 ], "spans": [ { "bbox": [ 105, 621, 505, 633 ], "score": 1.0, "content": "intentions and the chat agent’s feedback. This can be challenging, time-consuming, and sometimes", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 631, 506, 645 ], "spans": [ { "bbox": [ 106, 631, 506, 645 ], "score": 1.0, "content": "impossible. Crafting effective prompts often demands a deep understanding and expertise of a", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 642, 506, 655 ], "spans": [ { "bbox": [ 105, 642, 506, 655 ], "score": 1.0, "content": "particular domain of knowledge. Consider an individual who lacks trading expertise; they would find", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 654, 505, 666 ], "spans": [ { "bbox": [ 106, 654, 505, 666 ], "score": 1.0, "content": "it difficult to create suitable prompts for directing a chat agent to develop a trading application. This", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 664, 505, 677 ], "spans": [ { "bbox": [ 105, 664, 505, 677 ], "score": 1.0, "content": "predicament is raising a crucial question: can we replace human intervention with an autonomous", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 676, 505, 686 ], "spans": [ { "bbox": [ 106, 676, 505, 686 ], "score": 1.0, "content": "communicative agent capable of steering the conversation toward task completion with minimal", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 686, 506, 699 ], "spans": [ { "bbox": [ 106, 686, 506, 699 ], "score": 1.0, "content": "human supervision? To tackle this issue, it is crucial to conduct more research exploring the potential,", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 73, 505, 86 ], "spans": [ { "bbox": [ 105, 73, 505, 86 ], "score": 1.0, "content": "capabilities, and limitations of communicative agents that operate entirely on their own to complete", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 105, 83, 505, 96 ], "spans": [ { "bbox": [ 105, 83, 505, 96 ], "score": 1.0, "content": "tasks. Understanding how multiple agents interact with each other is important for anticipating the", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 105, 95, 506, 108 ], "spans": [ { "bbox": [ 105, 95, 506, 108 ], "score": 1.0, "content": "future of artificial intelligence. The dynamics of collaborating or competing agents play a key role in", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 105, 105, 363, 118 ], "spans": [ { "bbox": [ 105, 105, 363, 118 ], "score": 1.0, "content": "determining the success of AI systems [6, 26, 27, 84, 99, 9, 10].", "type": "text", "cross_page": true } ], "index": 3 } ], "index": 34.5, "bbox_fs": [ 105, 564, 506, 699 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 73, 505, 117 ], "lines": [ { "bbox": [ 105, 73, 505, 86 ], "spans": [ { "bbox": [ 105, 73, 505, 86 ], "score": 1.0, "content": "capabilities, and limitations of communicative agents that operate entirely on their own to complete", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 83, 505, 96 ], "spans": [ { "bbox": [ 105, 83, 505, 96 ], "score": 1.0, "content": "tasks. Understanding how multiple agents interact with each other is important for anticipating the", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 95, 506, 108 ], "spans": [ { "bbox": [ 105, 95, 506, 108 ], "score": 1.0, "content": "future of artificial intelligence. The dynamics of collaborating or competing agents play a key role in", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 105, 363, 118 ], "spans": [ { "bbox": [ 105, 105, 363, 118 ], "score": 1.0, "content": "determining the success of AI systems [6, 26, 27, 84, 99, 9, 10].", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "text", "bbox": [ 107, 122, 505, 231 ], "lines": [ { "bbox": [ 106, 122, 504, 134 ], "spans": [ { "bbox": [ 106, 122, 504, 134 ], "score": 1.0, "content": "This paper explores the potential of building scalable techniques to facilitate autonomous cooperation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 132, 504, 145 ], "spans": [ { "bbox": [ 105, 132, 504, 145 ], "score": 1.0, "content": "among communicative agents and provide insight into their “cognitive” processes. Several challenges", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 144, 505, 156 ], "spans": [ { "bbox": [ 105, 144, 505, 156 ], "score": 1.0, "content": "arise when asking a society of agents to autonomously cooperate on completing tasks. Examples we", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 154, 505, 168 ], "spans": [ { "bbox": [ 105, 154, 505, 168 ], "score": 1.0, "content": "encountered in our preliminary analysis include role flipping, assistant repeating instructions, flake", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 165, 505, 178 ], "spans": [ { "bbox": [ 105, 165, 505, 178 ], "score": 1.0, "content": "replies, and infinite loop of messages. Therefore, it is critical to investigate ways to align these models", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 176, 504, 189 ], "spans": [ { "bbox": [ 105, 176, 504, 189 ], "score": 1.0, "content": "with human intentions and to explore means enabling their effective cooperation. To address these", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 187, 505, 200 ], "spans": [ { "bbox": [ 105, 187, 505, 200 ], "score": 1.0, "content": "issues, we propose a novel cooperative agent framework named role-playing to automate cooperation", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 198, 505, 211 ], "spans": [ { "bbox": [ 105, 198, 505, 211 ], "score": 1.0, "content": "between communicative agents. Specifically, our proposed approach involves using role-playing with", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 209, 505, 222 ], "spans": [ { "bbox": [ 105, 209, 505, 222 ], "score": 1.0, "content": "inception prompting to autonomously guide the communicative agents toward task completion. Only", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 220, 501, 234 ], "spans": [ { "bbox": [ 105, 220, 501, 234 ], "score": 1.0, "content": "a preliminary idea is needed from human to guide the conversations toward complex task-solving.", "type": "text" } ], "index": 13 } ], "index": 8.5 }, { "type": "text", "bbox": [ 107, 236, 505, 466 ], "lines": [ { "bbox": [ 106, 237, 505, 248 ], "spans": [ { "bbox": [ 106, 237, 505, 248 ], "score": 1.0, "content": "Our library, which we make publicly available, provides modular functionality, and includes imple-", "type": "text" } ], "index": 14 }, { "bbox": [ 104, 246, 505, 261 ], "spans": [ { "bbox": [ 104, 246, 505, 261 ], "score": 1.0, "content": "mentations of different agents, examples of well-crafted prompts, and data explorers. We hope our", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 258, 505, 271 ], "spans": [ { "bbox": [ 105, 258, 505, 271 ], "score": 1.0, "content": "library serves as a ground for future research in various areas such as multi-agent systems, cooperative", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 269, 505, 282 ], "spans": [ { "bbox": [ 106, 269, 505, 282 ], "score": 1.0, "content": "AI, game theory simulations, social analysis, AI ethics, AI alignment, and beyond. In addition, our", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 280, 506, 294 ], "spans": [ { "bbox": [ 105, 280, 506, 294 ], "score": 1.0, "content": "role-playing method provides a highly scalable way to generate conversational data for studying the", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 290, 505, 304 ], "spans": [ { "bbox": [ 105, 290, 505, 304 ], "score": 1.0, "content": "behaviors and capabilities of chat agents. We showcase how role-playing can be used to let chat agents", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 302, 505, 314 ], "spans": [ { "bbox": [ 105, 302, 505, 314 ], "score": 1.0, "content": "communicate with each other for task completion and record their conversations for behavior analysis", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 311, 506, 327 ], "spans": [ { "bbox": [ 105, 311, 506, 327 ], "score": 1.0, "content": "and capability understanding. In particular, we consider two cooperative scenarios of role-playing", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 322, 505, 337 ], "spans": [ { "bbox": [ 105, 322, 505, 337 ], "score": 1.0, "content": "and generate two large conversational, task-oriented, and instruction-following datasets: AI Society", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 335, 505, 347 ], "spans": [ { "bbox": [ 106, 335, 505, 347 ], "score": 1.0, "content": "and Code. We also use our framework to collect two single-turn question-answer datasets, Math and", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 345, 506, 358 ], "spans": [ { "bbox": [ 105, 345, 506, 358 ], "score": 1.0, "content": "Science, for LLM ability emergence study. Furthermore, we generate a Misalignment dataset that is a", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 356, 506, 369 ], "spans": [ { "bbox": [ 106, 356, 506, 369 ], "score": 1.0, "content": "simulation of possible malicious applications which demonstrate the potential risks of an unaligned", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 367, 506, 380 ], "spans": [ { "bbox": [ 105, 367, 506, 380 ], "score": 1.0, "content": "autonomous agent system. The datasets offer a valuable resource for investigating conversational", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 378, 506, 391 ], "spans": [ { "bbox": [ 106, 378, 506, 391 ], "score": 1.0, "content": "language models, enabling them to comprehend and react to human language more effectively. Fur-", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 388, 506, 403 ], "spans": [ { "bbox": [ 105, 388, 506, 403 ], "score": 1.0, "content": "thermore, our role-playing offers a scalable method of creating conversational instruction-following", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 399, 505, 413 ], "spans": [ { "bbox": [ 105, 399, 505, 413 ], "score": 1.0, "content": "data, which can potentially enhance the development of more advanced language models. We show", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 410, 505, 424 ], "spans": [ { "bbox": [ 105, 410, 505, 424 ], "score": 1.0, "content": "that solutions derived from our role-playing framework outperform those generated in a single shot", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 421, 505, 436 ], "spans": [ { "bbox": [ 105, 421, 505, 436 ], "score": 1.0, "content": "by gpt-3.5-turbo [82] in both GPT4 and human evaluations. We also study knowledge emergence", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 433, 505, 446 ], "spans": [ { "bbox": [ 106, 433, 505, 446 ], "score": 1.0, "content": "in LLMs by fine-tuning LLaMA [117] on progressively growing datasets generated through our", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 442, 505, 457 ], "spans": [ { "bbox": [ 105, 442, 505, 457 ], "score": 1.0, "content": "framework. Additionally, we evaluate our code generation capabilities through benchmarking our", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 454, 330, 467 ], "spans": [ { "bbox": [ 106, 454, 330, 467 ], "score": 1.0, "content": "final model on HumanEval [18] and HumanEval+ [69].", "type": "text" } ], "index": 34 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 471, 505, 581 ], "lines": [ { "bbox": [ 106, 470, 506, 483 ], "spans": [ { "bbox": [ 106, 470, 506, 483 ], "score": 1.0, "content": "Contributions. Our contributions are fourfold: (1) We introduce a novel cooperative agent frame-", "type": "text" } ], "index": 35 }, { "bbox": [ 104, 480, 506, 496 ], "spans": [ { "bbox": [ 104, 480, 506, 496 ], "score": 1.0, "content": "work, role-playing , that allows communicative agents to collaborate autonomously toward completing", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 493, 505, 505 ], "spans": [ { "bbox": [ 106, 493, 505, 505 ], "score": 1.0, "content": "tasks while requiring minimal human intervention; (2) Our framework offers a scalable approach", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 504, 505, 516 ], "spans": [ { "bbox": [ 106, 504, 505, 516 ], "score": 1.0, "content": "for studying the cooperative behaviors and capabilities of multi-agent systems. It illuminates the", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 514, 506, 528 ], "spans": [ { "bbox": [ 105, 514, 506, 528 ], "score": 1.0, "content": "challenges of achieving autonomous cooperation, and provides strategies for addressing them. We", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 526, 506, 537 ], "spans": [ { "bbox": [ 106, 526, 506, 537 ], "score": 1.0, "content": "showcase the potential power of multi-agent collaboration for complex-task solving; (3) We demon-", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 537, 506, 549 ], "spans": [ { "bbox": [ 105, 537, 506, 549 ], "score": 1.0, "content": "strate the significant emergence of LLM training abilities by utilizing the datasets we have collected", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 547, 507, 560 ], "spans": [ { "bbox": [ 105, 547, 507, 560 ], "score": 1.0, "content": "from simulating four distinct agent collaboration scenarios; (4) We have open-sourced our library,", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 558, 506, 571 ], "spans": [ { "bbox": [ 105, 558, 506, 571 ], "score": 1.0, "content": "containing implementations of various agents, data generation pipelines, data analysis tools, and", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 569, 415, 582 ], "spans": [ { "bbox": [ 106, 569, 415, 582 ], "score": 1.0, "content": "collected datasets, to support research on communicative agents and beyond.", "type": "text" } ], "index": 44 } ], "index": 39.5 }, { "type": "title", "bbox": [ 107, 592, 197, 605 ], "lines": [ { "bbox": [ 105, 591, 198, 607 ], "spans": [ { "bbox": [ 105, 591, 198, 607 ], "score": 1.0, "content": "2 Related Work", "type": "text" } ], "index": 45 } ], "index": 45 }, { "type": "text", "bbox": [ 107, 612, 505, 722 ], "lines": [ { "bbox": [ 106, 612, 506, 625 ], "spans": [ { "bbox": [ 106, 612, 506, 625 ], "score": 1.0, "content": "Communicative Agents. Communication between agents has been studied for a long time [76,", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 622, 506, 636 ], "spans": [ { "bbox": [ 105, 622, 506, 636 ], "score": 1.0, "content": "77]. There are many ways to facilitate communication between agents, and with agents [29, 90,", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 633, 506, 647 ], "spans": [ { "bbox": [ 105, 633, 506, 647 ], "score": 1.0, "content": "97]. Among these, natural language is considered the most natural form of communication [97].", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 644, 506, 658 ], "spans": [ { "bbox": [ 105, 644, 506, 658 ], "score": 1.0, "content": "By enabling agents to function as communicators themselves, they become capable of solving", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 654, 506, 669 ], "spans": [ { "bbox": [ 105, 654, 506, 669 ], "score": 1.0, "content": "complex tasks [113, 85, 72, 3, 30, 111, 79, 41, 28, 102, 80, 106, 35, 49, 2, 51, 1, 55, 50, 65, 92].", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 666, 506, 680 ], "spans": [ { "bbox": [ 105, 666, 506, 680 ], "score": 1.0, "content": "Communication between AI agents can occur in a competitive setting [115, 108] or a cooperative", "type": "text" } ], "index": 51 }, { "bbox": [ 106, 678, 506, 690 ], "spans": [ { "bbox": [ 106, 678, 506, 690 ], "score": 1.0, "content": "setting [40, 27, 11, 137, 70]. Cooperative AI refers to artificial intelligence systems that are designed", "type": "text" } ], "index": 52 }, { "bbox": [ 106, 689, 505, 701 ], "spans": [ { "bbox": [ 106, 689, 505, 701 ], "score": 1.0, "content": "to work together with humans and other AI systems to achieve common goals [24, 125]. Cooperative", "type": "text" } ], "index": 53 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 506, 712 ], "score": 1.0, "content": "AI systems take into account the needs and capabilities of other agents in the system and actively seek", "type": "text" } ], "index": 54 }, { "bbox": [ 105, 709, 505, 723 ], "spans": [ { "bbox": [ 105, 709, 505, 723 ], "score": 1.0, "content": "to collaborate and coordinate their actions with them, which has many potential benefits, including", "type": "text" } ], "index": 55 } ], "index": 50.5 } ], "page_idx": 1, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 741, 309, 750 ], "lines": [ { "bbox": [ 301, 740, 310, 753 ], "spans": [ { "bbox": [ 301, 740, 310, 753 ], "score": 1.0, "content": "2", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 73, 505, 117 ], "lines": [], "index": 1.5, "bbox_fs": [ 105, 73, 506, 118 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 122, 505, 231 ], "lines": [ { "bbox": [ 106, 122, 504, 134 ], "spans": [ { "bbox": [ 106, 122, 504, 134 ], "score": 1.0, "content": "This paper explores the potential of building scalable techniques to facilitate autonomous cooperation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 132, 504, 145 ], "spans": [ { "bbox": [ 105, 132, 504, 145 ], "score": 1.0, "content": "among communicative agents and provide insight into their “cognitive” processes. Several challenges", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 144, 505, 156 ], "spans": [ { "bbox": [ 105, 144, 505, 156 ], "score": 1.0, "content": "arise when asking a society of agents to autonomously cooperate on completing tasks. Examples we", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 154, 505, 168 ], "spans": [ { "bbox": [ 105, 154, 505, 168 ], "score": 1.0, "content": "encountered in our preliminary analysis include role flipping, assistant repeating instructions, flake", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 165, 505, 178 ], "spans": [ { "bbox": [ 105, 165, 505, 178 ], "score": 1.0, "content": "replies, and infinite loop of messages. Therefore, it is critical to investigate ways to align these models", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 176, 504, 189 ], "spans": [ { "bbox": [ 105, 176, 504, 189 ], "score": 1.0, "content": "with human intentions and to explore means enabling their effective cooperation. To address these", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 187, 505, 200 ], "spans": [ { "bbox": [ 105, 187, 505, 200 ], "score": 1.0, "content": "issues, we propose a novel cooperative agent framework named role-playing to automate cooperation", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 198, 505, 211 ], "spans": [ { "bbox": [ 105, 198, 505, 211 ], "score": 1.0, "content": "between communicative agents. Specifically, our proposed approach involves using role-playing with", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 209, 505, 222 ], "spans": [ { "bbox": [ 105, 209, 505, 222 ], "score": 1.0, "content": "inception prompting to autonomously guide the communicative agents toward task completion. Only", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 220, 501, 234 ], "spans": [ { "bbox": [ 105, 220, 501, 234 ], "score": 1.0, "content": "a preliminary idea is needed from human to guide the conversations toward complex task-solving.", "type": "text" } ], "index": 13 } ], "index": 8.5, "bbox_fs": [ 105, 122, 505, 234 ] }, { "type": "text", "bbox": [ 107, 236, 505, 466 ], "lines": [ { "bbox": [ 106, 237, 505, 248 ], "spans": [ { "bbox": [ 106, 237, 505, 248 ], "score": 1.0, "content": "Our library, which we make publicly available, provides modular functionality, and includes imple-", "type": "text" } ], "index": 14 }, { "bbox": [ 104, 246, 505, 261 ], "spans": [ { "bbox": [ 104, 246, 505, 261 ], "score": 1.0, "content": "mentations of different agents, examples of well-crafted prompts, and data explorers. We hope our", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 258, 505, 271 ], "spans": [ { "bbox": [ 105, 258, 505, 271 ], "score": 1.0, "content": "library serves as a ground for future research in various areas such as multi-agent systems, cooperative", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 269, 505, 282 ], "spans": [ { "bbox": [ 106, 269, 505, 282 ], "score": 1.0, "content": "AI, game theory simulations, social analysis, AI ethics, AI alignment, and beyond. In addition, our", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 280, 506, 294 ], "spans": [ { "bbox": [ 105, 280, 506, 294 ], "score": 1.0, "content": "role-playing method provides a highly scalable way to generate conversational data for studying the", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 290, 505, 304 ], "spans": [ { "bbox": [ 105, 290, 505, 304 ], "score": 1.0, "content": "behaviors and capabilities of chat agents. We showcase how role-playing can be used to let chat agents", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 302, 505, 314 ], "spans": [ { "bbox": [ 105, 302, 505, 314 ], "score": 1.0, "content": "communicate with each other for task completion and record their conversations for behavior analysis", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 311, 506, 327 ], "spans": [ { "bbox": [ 105, 311, 506, 327 ], "score": 1.0, "content": "and capability understanding. In particular, we consider two cooperative scenarios of role-playing", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 322, 505, 337 ], "spans": [ { "bbox": [ 105, 322, 505, 337 ], "score": 1.0, "content": "and generate two large conversational, task-oriented, and instruction-following datasets: AI Society", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 335, 505, 347 ], "spans": [ { "bbox": [ 106, 335, 505, 347 ], "score": 1.0, "content": "and Code. We also use our framework to collect two single-turn question-answer datasets, Math and", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 345, 506, 358 ], "spans": [ { "bbox": [ 105, 345, 506, 358 ], "score": 1.0, "content": "Science, for LLM ability emergence study. Furthermore, we generate a Misalignment dataset that is a", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 356, 506, 369 ], "spans": [ { "bbox": [ 106, 356, 506, 369 ], "score": 1.0, "content": "simulation of possible malicious applications which demonstrate the potential risks of an unaligned", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 367, 506, 380 ], "spans": [ { "bbox": [ 105, 367, 506, 380 ], "score": 1.0, "content": "autonomous agent system. The datasets offer a valuable resource for investigating conversational", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 378, 506, 391 ], "spans": [ { "bbox": [ 106, 378, 506, 391 ], "score": 1.0, "content": "language models, enabling them to comprehend and react to human language more effectively. Fur-", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 388, 506, 403 ], "spans": [ { "bbox": [ 105, 388, 506, 403 ], "score": 1.0, "content": "thermore, our role-playing offers a scalable method of creating conversational instruction-following", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 399, 505, 413 ], "spans": [ { "bbox": [ 105, 399, 505, 413 ], "score": 1.0, "content": "data, which can potentially enhance the development of more advanced language models. We show", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 410, 505, 424 ], "spans": [ { "bbox": [ 105, 410, 505, 424 ], "score": 1.0, "content": "that solutions derived from our role-playing framework outperform those generated in a single shot", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 421, 505, 436 ], "spans": [ { "bbox": [ 105, 421, 505, 436 ], "score": 1.0, "content": "by gpt-3.5-turbo [82] in both GPT4 and human evaluations. We also study knowledge emergence", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 433, 505, 446 ], "spans": [ { "bbox": [ 106, 433, 505, 446 ], "score": 1.0, "content": "in LLMs by fine-tuning LLaMA [117] on progressively growing datasets generated through our", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 442, 505, 457 ], "spans": [ { "bbox": [ 105, 442, 505, 457 ], "score": 1.0, "content": "framework. Additionally, we evaluate our code generation capabilities through benchmarking our", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 454, 330, 467 ], "spans": [ { "bbox": [ 106, 454, 330, 467 ], "score": 1.0, "content": "final model on HumanEval [18] and HumanEval+ [69].", "type": "text" } ], "index": 34 } ], "index": 24, "bbox_fs": [ 104, 237, 506, 467 ] }, { "type": "text", "bbox": [ 107, 471, 505, 581 ], "lines": [ { "bbox": [ 106, 470, 506, 483 ], "spans": [ { "bbox": [ 106, 470, 506, 483 ], "score": 1.0, "content": "Contributions. Our contributions are fourfold: (1) We introduce a novel cooperative agent frame-", "type": "text" } ], "index": 35 }, { "bbox": [ 104, 480, 506, 496 ], "spans": [ { "bbox": [ 104, 480, 506, 496 ], "score": 1.0, "content": "work, role-playing , that allows communicative agents to collaborate autonomously toward completing", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 493, 505, 505 ], "spans": [ { "bbox": [ 106, 493, 505, 505 ], "score": 1.0, "content": "tasks while requiring minimal human intervention; (2) Our framework offers a scalable approach", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 504, 505, 516 ], "spans": [ { "bbox": [ 106, 504, 505, 516 ], "score": 1.0, "content": "for studying the cooperative behaviors and capabilities of multi-agent systems. It illuminates the", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 514, 506, 528 ], "spans": [ { "bbox": [ 105, 514, 506, 528 ], "score": 1.0, "content": "challenges of achieving autonomous cooperation, and provides strategies for addressing them. We", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 526, 506, 537 ], "spans": [ { "bbox": [ 106, 526, 506, 537 ], "score": 1.0, "content": "showcase the potential power of multi-agent collaboration for complex-task solving; (3) We demon-", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 537, 506, 549 ], "spans": [ { "bbox": [ 105, 537, 506, 549 ], "score": 1.0, "content": "strate the significant emergence of LLM training abilities by utilizing the datasets we have collected", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 547, 507, 560 ], "spans": [ { "bbox": [ 105, 547, 507, 560 ], "score": 1.0, "content": "from simulating four distinct agent collaboration scenarios; (4) We have open-sourced our library,", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 558, 506, 571 ], "spans": [ { "bbox": [ 105, 558, 506, 571 ], "score": 1.0, "content": "containing implementations of various agents, data generation pipelines, data analysis tools, and", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 569, 415, 582 ], "spans": [ { "bbox": [ 106, 569, 415, 582 ], "score": 1.0, "content": "collected datasets, to support research on communicative agents and beyond.", "type": "text" } ], "index": 44 } ], "index": 39.5, "bbox_fs": [ 104, 470, 507, 582 ] }, { "type": "title", "bbox": [ 107, 592, 197, 605 ], "lines": [ { "bbox": [ 105, 591, 198, 607 ], "spans": [ { "bbox": [ 105, 591, 198, 607 ], "score": 1.0, "content": "2 Related Work", "type": "text" } ], "index": 45 } ], "index": 45 }, { "type": "text", "bbox": [ 107, 612, 505, 722 ], "lines": [ { "bbox": [ 106, 612, 506, 625 ], "spans": [ { "bbox": [ 106, 612, 506, 625 ], "score": 1.0, "content": "Communicative Agents. Communication between agents has been studied for a long time [76,", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 622, 506, 636 ], "spans": [ { "bbox": [ 105, 622, 506, 636 ], "score": 1.0, "content": "77]. There are many ways to facilitate communication between agents, and with agents [29, 90,", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 633, 506, 647 ], "spans": [ { "bbox": [ 105, 633, 506, 647 ], "score": 1.0, "content": "97]. Among these, natural language is considered the most natural form of communication [97].", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 644, 506, 658 ], "spans": [ { "bbox": [ 105, 644, 506, 658 ], "score": 1.0, "content": "By enabling agents to function as communicators themselves, they become capable of solving", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 654, 506, 669 ], "spans": [ { "bbox": [ 105, 654, 506, 669 ], "score": 1.0, "content": "complex tasks [113, 85, 72, 3, 30, 111, 79, 41, 28, 102, 80, 106, 35, 49, 2, 51, 1, 55, 50, 65, 92].", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 666, 506, 680 ], "spans": [ { "bbox": [ 105, 666, 506, 680 ], "score": 1.0, "content": "Communication between AI agents can occur in a competitive setting [115, 108] or a cooperative", "type": "text" } ], "index": 51 }, { "bbox": [ 106, 678, 506, 690 ], "spans": [ { "bbox": [ 106, 678, 506, 690 ], "score": 1.0, "content": "setting [40, 27, 11, 137, 70]. Cooperative AI refers to artificial intelligence systems that are designed", "type": "text" } ], "index": 52 }, { "bbox": [ 106, 689, 505, 701 ], "spans": [ { "bbox": [ 106, 689, 505, 701 ], "score": 1.0, "content": "to work together with humans and other AI systems to achieve common goals [24, 125]. Cooperative", "type": "text" } ], "index": 53 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 506, 712 ], "score": 1.0, "content": "AI systems take into account the needs and capabilities of other agents in the system and actively seek", "type": "text" } ], "index": 54 }, { "bbox": [ 105, 709, 505, 723 ], "spans": [ { "bbox": [ 105, 709, 505, 723 ], "score": 1.0, "content": "to collaborate and coordinate their actions with them, which has many potential benefits, including", "type": "text" } ], "index": 55 }, { "bbox": [ 105, 72, 505, 86 ], "spans": [ { "bbox": [ 105, 72, 505, 86 ], "score": 1.0, "content": "increased efficiency, improved decision-making, and the ability to tackle complex problems that are", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 106, 84, 505, 95 ], "spans": [ { "bbox": [ 106, 84, 505, 95 ], "score": 1.0, "content": "beyond the reach of any single agent. However, designing effective cooperative AI systems is still an", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 105, 93, 506, 108 ], "spans": [ { "bbox": [ 105, 93, 506, 108 ], "score": 1.0, "content": "active area of research, as it requires addressing a range of technical, ethical, and social challenges", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 105, 105, 505, 119 ], "spans": [ { "bbox": [ 105, 105, 505, 119 ], "score": 1.0, "content": "[27]. Our work enables communicative agents to engage in a conversation and cooperate with each", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 105, 116, 506, 129 ], "spans": [ { "bbox": [ 105, 116, 506, 129 ], "score": 1.0, "content": "other to solve assigned tasks. The agents, each assigned a distinct role, are expected to apply their", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 105, 127, 321, 139 ], "spans": [ { "bbox": [ 105, 127, 321, 139 ], "score": 1.0, "content": "expertise and knowledge to solve their common task.", "type": "text", "cross_page": true } ], "index": 5 } ], "index": 50.5, "bbox_fs": [ 105, 612, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 73, 505, 138 ], "lines": [ { "bbox": [ 105, 72, 505, 86 ], "spans": [ { "bbox": [ 105, 72, 505, 86 ], "score": 1.0, "content": "increased efficiency, improved decision-making, and the ability to tackle complex problems that are", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 84, 505, 95 ], "spans": [ { "bbox": [ 106, 84, 505, 95 ], "score": 1.0, "content": "beyond the reach of any single agent. However, designing effective cooperative AI systems is still an", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 93, 506, 108 ], "spans": [ { "bbox": [ 105, 93, 506, 108 ], "score": 1.0, "content": "active area of research, as it requires addressing a range of technical, ethical, and social challenges", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 105, 505, 119 ], "spans": [ { "bbox": [ 105, 105, 505, 119 ], "score": 1.0, "content": "[27]. Our work enables communicative agents to engage in a conversation and cooperate with each", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 116, 506, 129 ], "spans": [ { "bbox": [ 105, 116, 506, 129 ], "score": 1.0, "content": "other to solve assigned tasks. The agents, each assigned a distinct role, are expected to apply their", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 127, 321, 139 ], "spans": [ { "bbox": [ 105, 127, 321, 139 ], "score": 1.0, "content": "expertise and knowledge to solve their common task.", "type": "text" } ], "index": 5 } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 143, 506, 471 ], "lines": [ { "bbox": [ 106, 144, 506, 156 ], "spans": [ { "bbox": [ 106, 144, 506, 156 ], "score": 1.0, "content": "Instructional LLMs and Prompt Engineering. LLMs are trained on diverse text data and excel", "type": "text" } ], "index": 6 }, { "bbox": [ 104, 153, 507, 168 ], "spans": [ { "bbox": [ 104, 153, 507, 168 ], "score": 1.0, "content": "in text completion, with various downstream NLP applications [12, 22, 47, 131, 117]. However,", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 164, 506, 179 ], "spans": [ { "bbox": [ 105, 164, 506, 179 ], "score": 1.0, "content": "InstructGPT suggests that LLMs may not align with user intent, proposing reinforcement learning", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 177, 506, 189 ], "spans": [ { "bbox": [ 106, 177, 506, 189 ], "score": 1.0, "content": "from human feedback (RLHF) [23] and Instruction Fine-Tuning (IFT) [121] to improve LLMs’", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 187, 506, 200 ], "spans": [ { "bbox": [ 106, 187, 506, 200 ], "score": 1.0, "content": "relevance and appropriateness to user instructions. Special types of instruction or prompting methods", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 198, 505, 210 ], "spans": [ { "bbox": [ 106, 198, 505, 210 ], "score": 1.0, "content": ", such as Chain-of-Thought (CoT) [123], zero-shot-CoT [61], and ReAct [126], have recently been", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 207, 506, 223 ], "spans": [ { "bbox": [ 105, 207, 506, 223 ], "score": 1.0, "content": "developed to enhance the performance of LLMs on reasoning, arithmetic and decision making", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 219, 506, 232 ], "spans": [ { "bbox": [ 105, 219, 506, 232 ], "score": 1.0, "content": "tasks [134, 118, 52, 73, 31, 103, 43, 64, 132, 46, 133, 105, 128, 25, 81, 109]. These techniques", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 231, 506, 243 ], "spans": [ { "bbox": [ 106, 231, 506, 243 ], "score": 1.0, "content": "underpin the impressive capabilities of recent dialogue LLMs [106, 116, 36, 9, 82, 13], which", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 241, 506, 254 ], "spans": [ { "bbox": [ 106, 241, 506, 254 ], "score": 1.0, "content": "aim to simulate human-like conversations and provide personalized and interactive experiences for", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 253, 506, 265 ], "spans": [ { "bbox": [ 105, 253, 506, 265 ], "score": 1.0, "content": "users, exhibiting the behavior of conversational AI agents [33]. However, generating instruction", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 262, 506, 277 ], "spans": [ { "bbox": [ 105, 262, 506, 277 ], "score": 1.0, "content": "datasets is a crucial challenge in building instruct-based LLMs, with existing datasets ranging", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 274, 506, 287 ], "spans": [ { "bbox": [ 105, 274, 506, 287 ], "score": 1.0, "content": "from crowdsourced to generated. Hand-crafted instruction instances are available in [120], while", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 286, 506, 298 ], "spans": [ { "bbox": [ 105, 286, 506, 298 ], "score": 1.0, "content": "leveraging previously crowdsourced NLP datasets is a less labor-intensive curation approach [121,", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 296, 506, 309 ], "spans": [ { "bbox": [ 106, 296, 506, 309 ], "score": 1.0, "content": "71, 78, 53]. LLMs have been explored for data generation in [101, 63, 68, 114], and Self-Instruct", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 307, 506, 320 ], "spans": [ { "bbox": [ 106, 307, 506, 320 ], "score": 1.0, "content": "[119] proposes a semi-automated process for instruction instance generation. Unnatural-Instruction", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 318, 506, 331 ], "spans": [ { "bbox": [ 106, 318, 506, 331 ], "score": 1.0, "content": "[48] collects instruction instances by prompting a language model with only three seed examples and", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 329, 506, 341 ], "spans": [ { "bbox": [ 105, 329, 506, 341 ], "score": 1.0, "content": "paraphrasing the generated instances to expand the dataset. There is also a large chunk of work that has", "type": "text" } ], "index": 23 }, { "bbox": [ 104, 339, 506, 353 ], "spans": [ { "bbox": [ 104, 339, 506, 353 ], "score": 1.0, "content": "proposed methods for automatic dataset creation [67, 57, 19, 75, 20, 98, 59, 96, 129, 62, 130, 86, 8].", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 350, 506, 364 ], "spans": [ { "bbox": [ 105, 350, 506, 364 ], "score": 1.0, "content": "Another important challenge is prompt engineering. The quality of the prompt used to guide LLMs", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 362, 505, 374 ], "spans": [ { "bbox": [ 105, 362, 505, 374 ], "score": 1.0, "content": "significantly affects its performance [91, 12, 66]. While LMs pre-trained on large data can implicitly", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 373, 505, 384 ], "spans": [ { "bbox": [ 106, 373, 505, 384 ], "score": 1.0, "content": "learn tasks with few-shot prompting, hand-crafted prompts may not always suffice. Automated", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 384, 506, 396 ], "spans": [ { "bbox": [ 105, 384, 506, 396 ], "score": 1.0, "content": "prompt generation methods have been proposed, such as gradient-guided search [104], mining-based", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 395, 506, 407 ], "spans": [ { "bbox": [ 106, 395, 506, 407 ], "score": 1.0, "content": "and paraphrasing-based techniques [54], a meta-prompt [93], and automatic instruction selection and", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 405, 506, 418 ], "spans": [ { "bbox": [ 105, 405, 506, 418 ], "score": 1.0, "content": "generation [136]. In this work, we introduce a conversational LLM auto-prompting method called", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 415, 507, 430 ], "spans": [ { "bbox": [ 105, 415, 507, 430 ], "score": 1.0, "content": "Inception Prompting, which enables agents to prompt each other to solve tasks through Role-Playing.", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 426, 506, 440 ], "spans": [ { "bbox": [ 105, 426, 506, 440 ], "score": 1.0, "content": "The AI user continuously provides instructions to the AI assistant for task-solving. This enables us to", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 439, 505, 450 ], "spans": [ { "bbox": [ 106, 439, 505, 450 ], "score": 1.0, "content": "save the streaming instruction-solution pairs and create diverse, instructional, conversational, and", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 449, 505, 461 ], "spans": [ { "bbox": [ 106, 449, 505, 461 ], "score": 1.0, "content": "task-oriented datasets. These datasets can be used to analyze the behavior and capabilities of LLMs", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 461, 418, 472 ], "spans": [ { "bbox": [ 106, 461, 418, 472 ], "score": 1.0, "content": "and for future research for fine-tuning LLMs with conversational instructions.", "type": "text" } ], "index": 35 } ], "index": 20.5 }, { "type": "text", "bbox": [ 107, 476, 505, 575 ], "lines": [ { "bbox": [ 106, 476, 504, 488 ], "spans": [ { "bbox": [ 106, 476, 504, 488 ], "score": 1.0, "content": "AI Alignment. AI alignment is a field that aims to ensure that AI systems adhere to their intended", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 487, 506, 499 ], "spans": [ { "bbox": [ 105, 487, 506, 499 ], "score": 1.0, "content": "goals, interests, and values, as envisioned by their designers [4, 39, 110, 32, 38, 74, 10]. The first", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 498, 505, 510 ], "spans": [ { "bbox": [ 105, 498, 505, 510 ], "score": 1.0, "content": "attempt at AI alignment was made through the \"Three Laws of Robotics,\" which was introduced", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 509, 505, 522 ], "spans": [ { "bbox": [ 105, 509, 505, 522 ], "score": 1.0, "content": "by Isaac Asimov in his science fiction stories [6]. Developing aligned AI systems is crucial for", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 520, 506, 533 ], "spans": [ { "bbox": [ 105, 520, 506, 533 ], "score": 1.0, "content": "achieving desired objectives while avoiding unintended consequences. Research in AI alignment", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 530, 505, 544 ], "spans": [ { "bbox": [ 105, 530, 505, 544 ], "score": 1.0, "content": "focuses on discouraging AI models from producing false, offensive, deceptive, or manipulative", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 541, 506, 555 ], "spans": [ { "bbox": [ 105, 541, 506, 555 ], "score": 1.0, "content": "information that could result in various harms [56, 112, 42, 37]. Achieving a high level of alignment", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 552, 506, 565 ], "spans": [ { "bbox": [ 105, 552, 506, 565 ], "score": 1.0, "content": "requires researchers to grapple with complex ethical, philosophical, and technical issues. We conduct", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 563, 506, 577 ], "spans": [ { "bbox": [ 105, 563, 506, 577 ], "score": 1.0, "content": "extensive experiments to study different role-playing situations, which probe the alignment of LLMs.", "type": "text" } ], "index": 44 } ], "index": 40 }, { "type": "title", "bbox": [ 107, 585, 192, 599 ], "lines": [ { "bbox": [ 104, 582, 194, 603 ], "spans": [ { "bbox": [ 104, 582, 194, 603 ], "score": 1.0, "content": "3 Methodology", "type": "text" } ], "index": 45 } ], "index": 45 }, { "type": "text", "bbox": [ 107, 604, 505, 648 ], "lines": [ { "bbox": [ 105, 603, 505, 617 ], "spans": [ { "bbox": [ 105, 603, 505, 617 ], "score": 1.0, "content": "In this paper, we focus on studying communicative agents under cooperative settings where they", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 614, 505, 628 ], "spans": [ { "bbox": [ 105, 614, 505, 628 ], "score": 1.0, "content": "share common interests. In particular, we study the assistant-user scenario, where a preliminary", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 626, 506, 639 ], "spans": [ { "bbox": [ 105, 626, 506, 639 ], "score": 1.0, "content": "idea is given at the start. Agents will conceptualize the idea into a specific task and complete it", "type": "text" } ], "index": 48 }, { "bbox": [ 106, 637, 258, 649 ], "spans": [ { "bbox": [ 106, 637, 258, 649 ], "score": 1.0, "content": "autonomously through conversations.", "type": "text" } ], "index": 49 } ], "index": 47.5 }, { "type": "title", "bbox": [ 107, 662, 237, 675 ], "lines": [ { "bbox": [ 105, 661, 239, 676 ], "spans": [ { "bbox": [ 105, 661, 239, 676 ], "score": 1.0, "content": "3.1 Role-playing Framework", "type": "text" } ], "index": 50 } ], "index": 50 }, { "type": "text", "bbox": [ 105, 683, 504, 705 ], "lines": [ { "bbox": [ 103, 681, 506, 696 ], "spans": [ { "bbox": [ 103, 681, 506, 696 ], "score": 1.0, "content": "“What’s the most resilient parasite? An Idea. A single idea from the human mind can build cities. An", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 694, 454, 706 ], "spans": [ { "bbox": [ 105, 694, 454, 706 ], "score": 1.0, "content": "idea can transform the world and rewrite all the rules. Which is why I have to steal it.”", "type": "text" } ], "index": 52 } ], "index": 51.5 }, { "type": "text", "bbox": [ 406, 711, 499, 722 ], "lines": [ { "bbox": [ 405, 709, 500, 724 ], "spans": [ { "bbox": [ 405, 709, 500, 724 ], "score": 1.0, "content": "- Dom Cobb, Inception", "type": "text" } ], "index": 53 } ], "index": 53 } ], "page_idx": 2, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 741, 309, 750 ], "lines": [ { "bbox": [ 301, 740, 309, 752 ], "spans": [ { "bbox": [ 301, 740, 309, 752 ], "score": 1.0, "content": "3", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 73, 505, 138 ], "lines": [], "index": 2.5, "bbox_fs": [ 105, 72, 506, 139 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 143, 506, 471 ], "lines": [ { "bbox": [ 106, 144, 506, 156 ], "spans": [ { "bbox": [ 106, 144, 506, 156 ], "score": 1.0, "content": "Instructional LLMs and Prompt Engineering. LLMs are trained on diverse text data and excel", "type": "text" } ], "index": 6 }, { "bbox": [ 104, 153, 507, 168 ], "spans": [ { "bbox": [ 104, 153, 507, 168 ], "score": 1.0, "content": "in text completion, with various downstream NLP applications [12, 22, 47, 131, 117]. However,", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 164, 506, 179 ], "spans": [ { "bbox": [ 105, 164, 506, 179 ], "score": 1.0, "content": "InstructGPT suggests that LLMs may not align with user intent, proposing reinforcement learning", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 177, 506, 189 ], "spans": [ { "bbox": [ 106, 177, 506, 189 ], "score": 1.0, "content": "from human feedback (RLHF) [23] and Instruction Fine-Tuning (IFT) [121] to improve LLMs’", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 187, 506, 200 ], "spans": [ { "bbox": [ 106, 187, 506, 200 ], "score": 1.0, "content": "relevance and appropriateness to user instructions. Special types of instruction or prompting methods", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 198, 505, 210 ], "spans": [ { "bbox": [ 106, 198, 505, 210 ], "score": 1.0, "content": ", such as Chain-of-Thought (CoT) [123], zero-shot-CoT [61], and ReAct [126], have recently been", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 207, 506, 223 ], "spans": [ { "bbox": [ 105, 207, 506, 223 ], "score": 1.0, "content": "developed to enhance the performance of LLMs on reasoning, arithmetic and decision making", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 219, 506, 232 ], "spans": [ { "bbox": [ 105, 219, 506, 232 ], "score": 1.0, "content": "tasks [134, 118, 52, 73, 31, 103, 43, 64, 132, 46, 133, 105, 128, 25, 81, 109]. These techniques", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 231, 506, 243 ], "spans": [ { "bbox": [ 106, 231, 506, 243 ], "score": 1.0, "content": "underpin the impressive capabilities of recent dialogue LLMs [106, 116, 36, 9, 82, 13], which", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 241, 506, 254 ], "spans": [ { "bbox": [ 106, 241, 506, 254 ], "score": 1.0, "content": "aim to simulate human-like conversations and provide personalized and interactive experiences for", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 253, 506, 265 ], "spans": [ { "bbox": [ 105, 253, 506, 265 ], "score": 1.0, "content": "users, exhibiting the behavior of conversational AI agents [33]. However, generating instruction", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 262, 506, 277 ], "spans": [ { "bbox": [ 105, 262, 506, 277 ], "score": 1.0, "content": "datasets is a crucial challenge in building instruct-based LLMs, with existing datasets ranging", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 274, 506, 287 ], "spans": [ { "bbox": [ 105, 274, 506, 287 ], "score": 1.0, "content": "from crowdsourced to generated. Hand-crafted instruction instances are available in [120], while", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 286, 506, 298 ], "spans": [ { "bbox": [ 105, 286, 506, 298 ], "score": 1.0, "content": "leveraging previously crowdsourced NLP datasets is a less labor-intensive curation approach [121,", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 296, 506, 309 ], "spans": [ { "bbox": [ 106, 296, 506, 309 ], "score": 1.0, "content": "71, 78, 53]. LLMs have been explored for data generation in [101, 63, 68, 114], and Self-Instruct", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 307, 506, 320 ], "spans": [ { "bbox": [ 106, 307, 506, 320 ], "score": 1.0, "content": "[119] proposes a semi-automated process for instruction instance generation. Unnatural-Instruction", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 318, 506, 331 ], "spans": [ { "bbox": [ 106, 318, 506, 331 ], "score": 1.0, "content": "[48] collects instruction instances by prompting a language model with only three seed examples and", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 329, 506, 341 ], "spans": [ { "bbox": [ 105, 329, 506, 341 ], "score": 1.0, "content": "paraphrasing the generated instances to expand the dataset. There is also a large chunk of work that has", "type": "text" } ], "index": 23 }, { "bbox": [ 104, 339, 506, 353 ], "spans": [ { "bbox": [ 104, 339, 506, 353 ], "score": 1.0, "content": "proposed methods for automatic dataset creation [67, 57, 19, 75, 20, 98, 59, 96, 129, 62, 130, 86, 8].", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 350, 506, 364 ], "spans": [ { "bbox": [ 105, 350, 506, 364 ], "score": 1.0, "content": "Another important challenge is prompt engineering. The quality of the prompt used to guide LLMs", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 362, 505, 374 ], "spans": [ { "bbox": [ 105, 362, 505, 374 ], "score": 1.0, "content": "significantly affects its performance [91, 12, 66]. While LMs pre-trained on large data can implicitly", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 373, 505, 384 ], "spans": [ { "bbox": [ 106, 373, 505, 384 ], "score": 1.0, "content": "learn tasks with few-shot prompting, hand-crafted prompts may not always suffice. Automated", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 384, 506, 396 ], "spans": [ { "bbox": [ 105, 384, 506, 396 ], "score": 1.0, "content": "prompt generation methods have been proposed, such as gradient-guided search [104], mining-based", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 395, 506, 407 ], "spans": [ { "bbox": [ 106, 395, 506, 407 ], "score": 1.0, "content": "and paraphrasing-based techniques [54], a meta-prompt [93], and automatic instruction selection and", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 405, 506, 418 ], "spans": [ { "bbox": [ 105, 405, 506, 418 ], "score": 1.0, "content": "generation [136]. In this work, we introduce a conversational LLM auto-prompting method called", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 415, 507, 430 ], "spans": [ { "bbox": [ 105, 415, 507, 430 ], "score": 1.0, "content": "Inception Prompting, which enables agents to prompt each other to solve tasks through Role-Playing.", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 426, 506, 440 ], "spans": [ { "bbox": [ 105, 426, 506, 440 ], "score": 1.0, "content": "The AI user continuously provides instructions to the AI assistant for task-solving. This enables us to", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 439, 505, 450 ], "spans": [ { "bbox": [ 106, 439, 505, 450 ], "score": 1.0, "content": "save the streaming instruction-solution pairs and create diverse, instructional, conversational, and", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 449, 505, 461 ], "spans": [ { "bbox": [ 106, 449, 505, 461 ], "score": 1.0, "content": "task-oriented datasets. These datasets can be used to analyze the behavior and capabilities of LLMs", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 461, 418, 472 ], "spans": [ { "bbox": [ 106, 461, 418, 472 ], "score": 1.0, "content": "and for future research for fine-tuning LLMs with conversational instructions.", "type": "text" } ], "index": 35 } ], "index": 20.5, "bbox_fs": [ 104, 144, 507, 472 ] }, { "type": "text", "bbox": [ 107, 476, 505, 575 ], "lines": [ { "bbox": [ 106, 476, 504, 488 ], "spans": [ { "bbox": [ 106, 476, 504, 488 ], "score": 1.0, "content": "AI Alignment. AI alignment is a field that aims to ensure that AI systems adhere to their intended", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 487, 506, 499 ], "spans": [ { "bbox": [ 105, 487, 506, 499 ], "score": 1.0, "content": "goals, interests, and values, as envisioned by their designers [4, 39, 110, 32, 38, 74, 10]. The first", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 498, 505, 510 ], "spans": [ { "bbox": [ 105, 498, 505, 510 ], "score": 1.0, "content": "attempt at AI alignment was made through the \"Three Laws of Robotics,\" which was introduced", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 509, 505, 522 ], "spans": [ { "bbox": [ 105, 509, 505, 522 ], "score": 1.0, "content": "by Isaac Asimov in his science fiction stories [6]. Developing aligned AI systems is crucial for", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 520, 506, 533 ], "spans": [ { "bbox": [ 105, 520, 506, 533 ], "score": 1.0, "content": "achieving desired objectives while avoiding unintended consequences. Research in AI alignment", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 530, 505, 544 ], "spans": [ { "bbox": [ 105, 530, 505, 544 ], "score": 1.0, "content": "focuses on discouraging AI models from producing false, offensive, deceptive, or manipulative", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 541, 506, 555 ], "spans": [ { "bbox": [ 105, 541, 506, 555 ], "score": 1.0, "content": "information that could result in various harms [56, 112, 42, 37]. Achieving a high level of alignment", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 552, 506, 565 ], "spans": [ { "bbox": [ 105, 552, 506, 565 ], "score": 1.0, "content": "requires researchers to grapple with complex ethical, philosophical, and technical issues. We conduct", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 563, 506, 577 ], "spans": [ { "bbox": [ 105, 563, 506, 577 ], "score": 1.0, "content": "extensive experiments to study different role-playing situations, which probe the alignment of LLMs.", "type": "text" } ], "index": 44 } ], "index": 40, "bbox_fs": [ 105, 476, 506, 577 ] }, { "type": "title", "bbox": [ 107, 585, 192, 599 ], "lines": [ { "bbox": [ 104, 582, 194, 603 ], "spans": [ { "bbox": [ 104, 582, 194, 603 ], "score": 1.0, "content": "3 Methodology", "type": "text" } ], "index": 45 } ], "index": 45 }, { "type": "text", "bbox": [ 107, 604, 505, 648 ], "lines": [ { "bbox": [ 105, 603, 505, 617 ], "spans": [ { "bbox": [ 105, 603, 505, 617 ], "score": 1.0, "content": "In this paper, we focus on studying communicative agents under cooperative settings where they", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 614, 505, 628 ], "spans": [ { "bbox": [ 105, 614, 505, 628 ], "score": 1.0, "content": "share common interests. In particular, we study the assistant-user scenario, where a preliminary", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 626, 506, 639 ], "spans": [ { "bbox": [ 105, 626, 506, 639 ], "score": 1.0, "content": "idea is given at the start. Agents will conceptualize the idea into a specific task and complete it", "type": "text" } ], "index": 48 }, { "bbox": [ 106, 637, 258, 649 ], "spans": [ { "bbox": [ 106, 637, 258, 649 ], "score": 1.0, "content": "autonomously through conversations.", "type": "text" } ], "index": 49 } ], "index": 47.5, "bbox_fs": [ 105, 603, 506, 649 ] }, { "type": "title", "bbox": [ 107, 662, 237, 675 ], "lines": [ { "bbox": [ 105, 661, 239, 676 ], "spans": [ { "bbox": [ 105, 661, 239, 676 ], "score": 1.0, "content": "3.1 Role-playing Framework", "type": "text" } ], "index": 50 } ], "index": 50 }, { "type": "text", "bbox": [ 105, 683, 504, 705 ], "lines": [ { "bbox": [ 103, 681, 506, 696 ], "spans": [ { "bbox": [ 103, 681, 506, 696 ], "score": 1.0, "content": "“What’s the most resilient parasite? An Idea. A single idea from the human mind can build cities. An", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 694, 454, 706 ], "spans": [ { "bbox": [ 105, 694, 454, 706 ], "score": 1.0, "content": "idea can transform the world and rewrite all the rules. Which is why I have to steal it.”", "type": "text" } ], "index": 52 } ], "index": 51.5, "bbox_fs": [ 103, 681, 506, 706 ] }, { "type": "text", "bbox": [ 406, 711, 499, 722 ], "lines": [ { "bbox": [ 405, 709, 500, 724 ], "spans": [ { "bbox": [ 405, 709, 500, 724 ], "score": 1.0, "content": "- Dom Cobb, Inception", "type": "text" } ], "index": 53 } ], "index": 53, "bbox_fs": [ 405, 709, 500, 724 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 143, 70, 468, 254 ], "blocks": [ { "type": "image_body", "bbox": [ 143, 70, 468, 254 ], "group_id": 0, "lines": [ { "bbox": [ 143, 70, 468, 254 ], "spans": [ { "bbox": [ 143, 70, 468, 254 ], "score": 0.976, "type": "image", "image_path": "6abc6059b41ace03cab1cbc2501d41463903922476103d46fbfb309436a5b8fc.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 143, 70, 468, 131.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 143, 131.33333333333334, 468, 192.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 143, 192.66666666666669, 468, 254.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 259, 505, 336 ], "group_id": 0, "lines": [ { "bbox": [ 105, 258, 507, 273 ], "spans": [ { "bbox": [ 105, 258, 507, 273 ], "score": 1.0, "content": "Figure 1: CAMEL Role-Playing Framework. Our role-playing setup starts with the human user", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 270, 506, 283 ], "spans": [ { "bbox": [ 105, 270, 506, 283 ], "score": 1.0, "content": "having an idea they want to implement, e.g. develop a trading bot for the stock market. The roles", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 280, 506, 294 ], "spans": [ { "bbox": [ 105, 280, 506, 294 ], "score": 1.0, "content": "involved in this task would be an AI assistant agent who is a python programmer and an AI user", "type": "text" } ], "index": 5 }, { "bbox": [ 104, 291, 505, 306 ], "spans": [ { "bbox": [ 104, 291, 505, 306 ], "score": 1.0, "content": "agent who is a stock trader. The task is made more specific using our task specifier agent, leading", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 303, 505, 316 ], "spans": [ { "bbox": [ 105, 303, 505, 316 ], "score": 1.0, "content": "to a well-defined task for the assistant to solve. Both AI user and AI assistant are provided with", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 314, 505, 326 ], "spans": [ { "bbox": [ 106, 314, 505, 326 ], "score": 1.0, "content": "the specified task, after which they collaboratively communicate by chatting with each other in an", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 325, 332, 338 ], "spans": [ { "bbox": [ 105, 325, 332, 338 ], "score": 1.0, "content": "instruction-following fashion to solve the specified task.", "type": "text" } ], "index": 9 } ], "index": 6 } ], "index": 3.5 }, { "type": "text", "bbox": [ 106, 361, 505, 461 ], "lines": [ { "bbox": [ 106, 362, 505, 374 ], "spans": [ { "bbox": [ 106, 362, 505, 374 ], "score": 1.0, "content": "Our proposed framework is a novel role-playing approach for studying multiple communicative", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 372, 506, 385 ], "spans": [ { "bbox": [ 105, 372, 506, 385 ], "score": 1.0, "content": "agents. Specifically, we concentrate on task-oriented role-playing that involves one AI assistant and", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 384, 505, 396 ], "spans": [ { "bbox": [ 106, 384, 505, 396 ], "score": 1.0, "content": "one AI user. After the multi-agent system receives a preliminary idea and the role assignment from", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 394, 506, 407 ], "spans": [ { "bbox": [ 105, 394, 506, 407 ], "score": 1.0, "content": "human users, a task-specifier agent will provide a detailed description to make the idea specific.", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 404, 505, 418 ], "spans": [ { "bbox": [ 105, 404, 505, 418 ], "score": 1.0, "content": "Afterwards, the AI assistant and AI user will cooperate on completing the specified task through", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 416, 504, 428 ], "spans": [ { "bbox": [ 106, 416, 504, 428 ], "score": 1.0, "content": "multi-turn conversations until the AI user determines the task is done. The AI user is responsible for", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 427, 505, 439 ], "spans": [ { "bbox": [ 105, 427, 505, 439 ], "score": 1.0, "content": "giving instructions to the AI assistant and directing the conversation toward task completion. On the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 437, 505, 450 ], "spans": [ { "bbox": [ 105, 437, 505, 450 ], "score": 1.0, "content": "other hand, the AI assistant is designed to follow the instructions from the AI user and respond with", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 449, 417, 462 ], "spans": [ { "bbox": [ 105, 449, 417, 462 ], "score": 1.0, "content": "specific solutions. The whole role-playing framework is depicted in Figure 1.", "type": "text" } ], "index": 18 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 465, 505, 651 ], "lines": [ { "bbox": [ 106, 465, 505, 477 ], "spans": [ { "bbox": [ 106, 465, 505, 477 ], "score": 1.0, "content": "Human Input and Task Specifying. The role-playing session will be instantiated from an idea and", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 476, 505, 489 ], "spans": [ { "bbox": [ 105, 476, 505, 489 ], "score": 1.0, "content": "selected roles by humans. As an example in Figure 1, a human has a preliminary idea to develop", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 488, 505, 499 ], "spans": [ { "bbox": [ 105, 488, 505, 499 ], "score": 1.0, "content": "a trading bot for the stock market. Humans may or may not have the knowledge about how the", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 497, 505, 510 ], "spans": [ { "bbox": [ 105, 497, 505, 510 ], "score": 1.0, "content": "idea can be realized. What is needed is only to designate the potential roles that can implement the", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 507, 506, 522 ], "spans": [ { "bbox": [ 105, 507, 506, 522 ], "score": 1.0, "content": "idea. For instance, a Python Programmer could collaborate with a Stock Trader to realize the idea", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 518, 506, 533 ], "spans": [ { "bbox": [ 105, 518, 506, 533 ], "score": 1.0, "content": "of developing a trading bot for the stock market. After the idea and roles are determined, the task", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 531, 506, 543 ], "spans": [ { "bbox": [ 105, 531, 506, 543 ], "score": 1.0, "content": "specifier agent will brainstorm a specific task that the AI Assistant role can help with the AI user role", "type": "text" } ], "index": 25 }, { "bbox": [ 104, 541, 506, 555 ], "spans": [ { "bbox": [ 104, 541, 506, 555 ], "score": 1.0, "content": "to complete based on the input idea. An example of a specified task in this scenario could be: develop", "type": "text" } ], "index": 26 }, { "bbox": [ 104, 552, 506, 566 ], "spans": [ { "bbox": [ 104, 552, 506, 566 ], "score": 1.0, "content": "a trading bot with a sentiment analysis tool that can monitor social media platforms for positive or", "type": "text" } ], "index": 27 }, { "bbox": [ 104, 564, 506, 576 ], "spans": [ { "bbox": [ 104, 564, 506, 576 ], "score": 1.0, "content": "negative comments about a particular stock, and execute trades based on sentiment analysis results.", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 574, 505, 587 ], "spans": [ { "bbox": [ 105, 574, 505, 587 ], "score": 1.0, "content": "The main motivation for introducing a task specifier is that conversational agents usually require", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 585, 506, 598 ], "spans": [ { "bbox": [ 105, 585, 506, 598 ], "score": 1.0, "content": "a concrete task prompt for realizing the task which might be challenging or time-consuming for a", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 596, 506, 609 ], "spans": [ { "bbox": [ 105, 596, 506, 609 ], "score": 1.0, "content": "non-domain expert. Therefore, the task specifier agent serves as an enhanced imagination module", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 607, 506, 620 ], "spans": [ { "bbox": [ 105, 607, 506, 620 ], "score": 1.0, "content": "for the idea implementation. Please note that, when studying our framework at a large scale for AI", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 618, 506, 630 ], "spans": [ { "bbox": [ 105, 618, 506, 630 ], "score": 1.0, "content": "society and Code scenarios, we generate roles and ideas automatically by prompting LLMs instead of", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 628, 507, 643 ], "spans": [ { "bbox": [ 105, 628, 507, 643 ], "score": 1.0, "content": "relying on human inputs. For our generated Math and Science datasets we generated problem topics,", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 640, 348, 653 ], "spans": [ { "bbox": [ 105, 640, 348, 653 ], "score": 1.0, "content": "subtopics, and problems automatically by prompting LLMs.", "type": "text" } ], "index": 35 } ], "index": 27 }, { "type": "text", "bbox": [ 107, 656, 505, 722 ], "lines": [ { "bbox": [ 105, 655, 505, 668 ], "spans": [ { "bbox": [ 105, 655, 505, 668 ], "score": 1.0, "content": "AI Assistant-User Role Assignment. After the task specification, The AI assistant role and the AI", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 667, 505, 680 ], "spans": [ { "bbox": [ 105, 667, 505, 680 ], "score": 1.0, "content": "user role will be assigned to the user agent and the assistant agent correspondingly to complete the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 677, 506, 691 ], "spans": [ { "bbox": [ 105, 677, 506, 691 ], "score": 1.0, "content": "specified task. In practice, a system message is passed to each agent declaring their role. We refer", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 688, 505, 702 ], "spans": [ { "bbox": [ 105, 688, 283, 702 ], "score": 1.0, "content": "to the assistant system prompt/message by", "type": "text" }, { "bbox": [ 283, 689, 298, 700 ], "score": 0.9, "content": "\\mathcal { P } _ { A }", "type": "inline_equation" }, { "bbox": [ 299, 688, 394, 702 ], "score": 1.0, "content": "and that of the user by", "type": "text" }, { "bbox": [ 395, 689, 409, 700 ], "score": 0.89, "content": "\\mathcal { P } _ { \\mathcal { U } }", "type": "inline_equation" }, { "bbox": [ 409, 688, 505, 702 ], "score": 1.0, "content": ". The system messages", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 700, 506, 712 ], "spans": [ { "bbox": [ 105, 700, 344, 712 ], "score": 1.0, "content": "are passed to the agents before the conversations start. Let", "type": "text" }, { "bbox": [ 345, 700, 357, 711 ], "score": 0.88, "content": "\\mathcal { F } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 358, 700, 376, 712 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 376, 700, 389, 711 ], "score": 0.89, "content": "\\mathcal { F } _ { 2 }", "type": "inline_equation" }, { "bbox": [ 389, 700, 506, 712 ], "score": 1.0, "content": "denote two large-scale auto-", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 711, 505, 723 ], "spans": [ { "bbox": [ 106, 711, 505, 723 ], "score": 1.0, "content": "regressive language models [82]. When the system message is passed to those models respectively, we", "type": "text" } ], "index": 41 } ], "index": 38.5 } ], "page_idx": 3, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 742, 308, 750 ], "lines": [ { "bbox": [ 301, 741, 310, 752 ], "spans": [ { "bbox": [ 301, 741, 310, 752 ], "score": 1.0, "content": "", "type": "text", "height": 11, "width": 9 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 143, 70, 468, 254 ], "blocks": [ { "type": "image_body", "bbox": [ 143, 70, 468, 254 ], "group_id": 0, "lines": [ { "bbox": [ 143, 70, 468, 254 ], "spans": [ { "bbox": [ 143, 70, 468, 254 ], "score": 0.976, "type": "image", "image_path": "6abc6059b41ace03cab1cbc2501d41463903922476103d46fbfb309436a5b8fc.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 143, 70, 468, 131.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 143, 131.33333333333334, 468, 192.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 143, 192.66666666666669, 468, 254.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 259, 505, 336 ], "group_id": 0, "lines": [ { "bbox": [ 105, 258, 507, 273 ], "spans": [ { "bbox": [ 105, 258, 507, 273 ], "score": 1.0, "content": "Figure 1: CAMEL Role-Playing Framework. Our role-playing setup starts with the human user", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 270, 506, 283 ], "spans": [ { "bbox": [ 105, 270, 506, 283 ], "score": 1.0, "content": "having an idea they want to implement, e.g. develop a trading bot for the stock market. The roles", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 280, 506, 294 ], "spans": [ { "bbox": [ 105, 280, 506, 294 ], "score": 1.0, "content": "involved in this task would be an AI assistant agent who is a python programmer and an AI user", "type": "text" } ], "index": 5 }, { "bbox": [ 104, 291, 505, 306 ], "spans": [ { "bbox": [ 104, 291, 505, 306 ], "score": 1.0, "content": "agent who is a stock trader. The task is made more specific using our task specifier agent, leading", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 303, 505, 316 ], "spans": [ { "bbox": [ 105, 303, 505, 316 ], "score": 1.0, "content": "to a well-defined task for the assistant to solve. Both AI user and AI assistant are provided with", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 314, 505, 326 ], "spans": [ { "bbox": [ 106, 314, 505, 326 ], "score": 1.0, "content": "the specified task, after which they collaboratively communicate by chatting with each other in an", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 325, 332, 338 ], "spans": [ { "bbox": [ 105, 325, 332, 338 ], "score": 1.0, "content": "instruction-following fashion to solve the specified task.", "type": "text" } ], "index": 9 } ], "index": 6 } ], "index": 3.5 }, { "type": "text", "bbox": [ 106, 361, 505, 461 ], "lines": [ { "bbox": [ 106, 362, 505, 374 ], "spans": [ { "bbox": [ 106, 362, 505, 374 ], "score": 1.0, "content": "Our proposed framework is a novel role-playing approach for studying multiple communicative", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 372, 506, 385 ], "spans": [ { "bbox": [ 105, 372, 506, 385 ], "score": 1.0, "content": "agents. Specifically, we concentrate on task-oriented role-playing that involves one AI assistant and", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 384, 505, 396 ], "spans": [ { "bbox": [ 106, 384, 505, 396 ], "score": 1.0, "content": "one AI user. After the multi-agent system receives a preliminary idea and the role assignment from", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 394, 506, 407 ], "spans": [ { "bbox": [ 105, 394, 506, 407 ], "score": 1.0, "content": "human users, a task-specifier agent will provide a detailed description to make the idea specific.", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 404, 505, 418 ], "spans": [ { "bbox": [ 105, 404, 505, 418 ], "score": 1.0, "content": "Afterwards, the AI assistant and AI user will cooperate on completing the specified task through", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 416, 504, 428 ], "spans": [ { "bbox": [ 106, 416, 504, 428 ], "score": 1.0, "content": "multi-turn conversations until the AI user determines the task is done. The AI user is responsible for", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 427, 505, 439 ], "spans": [ { "bbox": [ 105, 427, 505, 439 ], "score": 1.0, "content": "giving instructions to the AI assistant and directing the conversation toward task completion. On the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 437, 505, 450 ], "spans": [ { "bbox": [ 105, 437, 505, 450 ], "score": 1.0, "content": "other hand, the AI assistant is designed to follow the instructions from the AI user and respond with", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 449, 417, 462 ], "spans": [ { "bbox": [ 105, 449, 417, 462 ], "score": 1.0, "content": "specific solutions. The whole role-playing framework is depicted in Figure 1.", "type": "text" } ], "index": 18 } ], "index": 14, "bbox_fs": [ 105, 362, 506, 462 ] }, { "type": "text", "bbox": [ 106, 465, 505, 651 ], "lines": [ { "bbox": [ 106, 465, 505, 477 ], "spans": [ { "bbox": [ 106, 465, 505, 477 ], "score": 1.0, "content": "Human Input and Task Specifying. The role-playing session will be instantiated from an idea and", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 476, 505, 489 ], "spans": [ { "bbox": [ 105, 476, 505, 489 ], "score": 1.0, "content": "selected roles by humans. As an example in Figure 1, a human has a preliminary idea to develop", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 488, 505, 499 ], "spans": [ { "bbox": [ 105, 488, 505, 499 ], "score": 1.0, "content": "a trading bot for the stock market. Humans may or may not have the knowledge about how the", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 497, 505, 510 ], "spans": [ { "bbox": [ 105, 497, 505, 510 ], "score": 1.0, "content": "idea can be realized. What is needed is only to designate the potential roles that can implement the", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 507, 506, 522 ], "spans": [ { "bbox": [ 105, 507, 506, 522 ], "score": 1.0, "content": "idea. For instance, a Python Programmer could collaborate with a Stock Trader to realize the idea", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 518, 506, 533 ], "spans": [ { "bbox": [ 105, 518, 506, 533 ], "score": 1.0, "content": "of developing a trading bot for the stock market. After the idea and roles are determined, the task", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 531, 506, 543 ], "spans": [ { "bbox": [ 105, 531, 506, 543 ], "score": 1.0, "content": "specifier agent will brainstorm a specific task that the AI Assistant role can help with the AI user role", "type": "text" } ], "index": 25 }, { "bbox": [ 104, 541, 506, 555 ], "spans": [ { "bbox": [ 104, 541, 506, 555 ], "score": 1.0, "content": "to complete based on the input idea. An example of a specified task in this scenario could be: develop", "type": "text" } ], "index": 26 }, { "bbox": [ 104, 552, 506, 566 ], "spans": [ { "bbox": [ 104, 552, 506, 566 ], "score": 1.0, "content": "a trading bot with a sentiment analysis tool that can monitor social media platforms for positive or", "type": "text" } ], "index": 27 }, { "bbox": [ 104, 564, 506, 576 ], "spans": [ { "bbox": [ 104, 564, 506, 576 ], "score": 1.0, "content": "negative comments about a particular stock, and execute trades based on sentiment analysis results.", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 574, 505, 587 ], "spans": [ { "bbox": [ 105, 574, 505, 587 ], "score": 1.0, "content": "The main motivation for introducing a task specifier is that conversational agents usually require", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 585, 506, 598 ], "spans": [ { "bbox": [ 105, 585, 506, 598 ], "score": 1.0, "content": "a concrete task prompt for realizing the task which might be challenging or time-consuming for a", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 596, 506, 609 ], "spans": [ { "bbox": [ 105, 596, 506, 609 ], "score": 1.0, "content": "non-domain expert. Therefore, the task specifier agent serves as an enhanced imagination module", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 607, 506, 620 ], "spans": [ { "bbox": [ 105, 607, 506, 620 ], "score": 1.0, "content": "for the idea implementation. Please note that, when studying our framework at a large scale for AI", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 618, 506, 630 ], "spans": [ { "bbox": [ 105, 618, 506, 630 ], "score": 1.0, "content": "society and Code scenarios, we generate roles and ideas automatically by prompting LLMs instead of", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 628, 507, 643 ], "spans": [ { "bbox": [ 105, 628, 507, 643 ], "score": 1.0, "content": "relying on human inputs. For our generated Math and Science datasets we generated problem topics,", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 640, 348, 653 ], "spans": [ { "bbox": [ 105, 640, 348, 653 ], "score": 1.0, "content": "subtopics, and problems automatically by prompting LLMs.", "type": "text" } ], "index": 35 } ], "index": 27, "bbox_fs": [ 104, 465, 507, 653 ] }, { "type": "text", "bbox": [ 107, 656, 505, 722 ], "lines": [ { "bbox": [ 105, 655, 505, 668 ], "spans": [ { "bbox": [ 105, 655, 505, 668 ], "score": 1.0, "content": "AI Assistant-User Role Assignment. After the task specification, The AI assistant role and the AI", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 667, 505, 680 ], "spans": [ { "bbox": [ 105, 667, 505, 680 ], "score": 1.0, "content": "user role will be assigned to the user agent and the assistant agent correspondingly to complete the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 677, 506, 691 ], "spans": [ { "bbox": [ 105, 677, 506, 691 ], "score": 1.0, "content": "specified task. In practice, a system message is passed to each agent declaring their role. We refer", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 688, 505, 702 ], "spans": [ { "bbox": [ 105, 688, 283, 702 ], "score": 1.0, "content": "to the assistant system prompt/message by", "type": "text" }, { "bbox": [ 283, 689, 298, 700 ], "score": 0.9, "content": "\\mathcal { P } _ { A }", "type": "inline_equation" }, { "bbox": [ 299, 688, 394, 702 ], "score": 1.0, "content": "and that of the user by", "type": "text" }, { "bbox": [ 395, 689, 409, 700 ], "score": 0.89, "content": "\\mathcal { P } _ { \\mathcal { U } }", "type": "inline_equation" }, { "bbox": [ 409, 688, 505, 702 ], "score": 1.0, "content": ". The system messages", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 700, 506, 712 ], "spans": [ { "bbox": [ 105, 700, 344, 712 ], "score": 1.0, "content": "are passed to the agents before the conversations start. Let", "type": "text" }, { "bbox": [ 345, 700, 357, 711 ], "score": 0.88, "content": "\\mathcal { F } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 358, 700, 376, 712 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 376, 700, 389, 711 ], "score": 0.89, "content": "\\mathcal { F } _ { 2 }", "type": "inline_equation" }, { "bbox": [ 389, 700, 506, 712 ], "score": 1.0, "content": "denote two large-scale auto-", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 711, 505, 723 ], "spans": [ { "bbox": [ 106, 711, 505, 723 ], "score": 1.0, "content": "regressive language models [82]. When the system message is passed to those models respectively, we", "type": "text" } ], "index": 41 }, { "bbox": [ 104, 69, 508, 87 ], "spans": [ { "bbox": [ 104, 69, 134, 87 ], "score": 1.0, "content": "obtain", "type": "text", "cross_page": true }, { "bbox": [ 134, 71, 178, 84 ], "score": 0.93, "content": "\\mathcal { A } \\mathcal { F } _ { 1 } ^ { \\mathcal { P } _ { A } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 179, 69, 197, 87 ], "score": 1.0, "content": "and", "type": "text", "cross_page": true }, { "bbox": [ 197, 71, 240, 85 ], "score": 0.93, "content": "\\mathcal { U } \\mathcal { F } _ { 2 } ^ { \\mathcal { P } _ { \\mathcal { U } } }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 241, 69, 508, 87 ], "score": 1.0, "content": "which are referred to as the assistant and user agents respectively.", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 106, 83, 506, 96 ], "spans": [ { "bbox": [ 106, 83, 506, 96 ], "score": 1.0, "content": "In Figure 1, the AI assistant and the AI user are assigned the roles of a Python Programmer and a", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 105, 93, 506, 107 ], "spans": [ { "bbox": [ 105, 93, 506, 107 ], "score": 1.0, "content": "Stock Trader at the beginning of the role-playing session respectively. The AI user serves as a task", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 105, 105, 507, 118 ], "spans": [ { "bbox": [ 105, 105, 507, 118 ], "score": 1.0, "content": "planner, engaging in interactive planning to determine feasible steps for the AI assistant to execute.", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 106, 116, 505, 129 ], "spans": [ { "bbox": [ 106, 116, 505, 129 ], "score": 1.0, "content": "Meanwhile, the AI assistant acts as a task executor, offering solutions, executing planned steps, and", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 106, 127, 248, 140 ], "spans": [ { "bbox": [ 106, 127, 248, 140 ], "score": 1.0, "content": "providing responses to the AI user.", "type": "text", "cross_page": true } ], "index": 5 } ], "index": 38.5, "bbox_fs": [ 105, 655, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 106, 70, 506, 138 ], "lines": [ { "bbox": [ 104, 69, 508, 87 ], "spans": [ { "bbox": [ 104, 69, 134, 87 ], "score": 1.0, "content": "obtain", "type": "text" }, { "bbox": [ 134, 71, 178, 84 ], "score": 0.93, "content": "\\mathcal { A } \\mathcal { F } _ { 1 } ^ { \\mathcal { P } _ { A } }", "type": "inline_equation" }, { "bbox": [ 179, 69, 197, 87 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 197, 71, 240, 85 ], "score": 0.93, "content": "\\mathcal { U } \\mathcal { F } _ { 2 } ^ { \\mathcal { P } _ { \\mathcal { U } } }", "type": "inline_equation" }, { "bbox": [ 241, 69, 508, 87 ], "score": 1.0, "content": "which are referred to as the assistant and user agents respectively.", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 83, 506, 96 ], "spans": [ { "bbox": [ 106, 83, 506, 96 ], "score": 1.0, "content": "In Figure 1, the AI assistant and the AI user are assigned the roles of a Python Programmer and a", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 93, 506, 107 ], "spans": [ { "bbox": [ 105, 93, 506, 107 ], "score": 1.0, "content": "Stock Trader at the beginning of the role-playing session respectively. The AI user serves as a task", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 105, 507, 118 ], "spans": [ { "bbox": [ 105, 105, 507, 118 ], "score": 1.0, "content": "planner, engaging in interactive planning to determine feasible steps for the AI assistant to execute.", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 116, 505, 129 ], "spans": [ { "bbox": [ 106, 116, 505, 129 ], "score": 1.0, "content": "Meanwhile, the AI assistant acts as a task executor, offering solutions, executing planned steps, and", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 127, 248, 140 ], "spans": [ { "bbox": [ 106, 127, 248, 140 ], "score": 1.0, "content": "providing responses to the AI user.", "type": "text" } ], "index": 5 } ], "index": 2.5 }, { "type": "text", "bbox": [ 106, 143, 505, 210 ], "lines": [ { "bbox": [ 106, 144, 504, 155 ], "spans": [ { "bbox": [ 106, 144, 495, 155 ], "score": 1.0, "content": "Conversation Towards Task-Solving. After the role assignment is completed, the AI assistant", "type": "text" }, { "bbox": [ 495, 144, 504, 154 ], "score": 0.68, "content": "\\mathcal { A }", "type": "inline_equation" } ], "index": 6 }, { "bbox": [ 105, 154, 506, 167 ], "spans": [ { "bbox": [ 105, 154, 157, 167 ], "score": 1.0, "content": "and AI user", "type": "text" }, { "bbox": [ 157, 155, 167, 165 ], "score": 0.7, "content": "\\mathcal { U }", "type": "inline_equation" }, { "bbox": [ 167, 154, 506, 167 ], "score": 1.0, "content": "will collaborate in an instruction-following manner to accomplish the task. In the", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 164, 506, 178 ], "spans": [ { "bbox": [ 105, 164, 506, 178 ], "score": 1.0, "content": "AI assistant-user scenario, the AI user is responsible for providing instructions, and the assistant", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 176, 506, 189 ], "spans": [ { "bbox": [ 105, 176, 506, 189 ], "score": 1.0, "content": "is expected to respond with a solution that fulfills the instructions. Formally, we denote the user", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 187, 506, 199 ], "spans": [ { "bbox": [ 105, 187, 250, 199 ], "score": 1.0, "content": "instruction message obtained at time", "type": "text" }, { "bbox": [ 250, 188, 255, 197 ], "score": 0.78, "content": "t", "type": "inline_equation" }, { "bbox": [ 256, 187, 267, 199 ], "score": 1.0, "content": "by", "type": "text" }, { "bbox": [ 268, 188, 278, 198 ], "score": 0.87, "content": "\\mathcal { T } _ { t }", "type": "inline_equation" }, { "bbox": [ 278, 187, 390, 199 ], "score": 1.0, "content": "and the assistant solution by", "type": "text" }, { "bbox": [ 390, 187, 401, 198 ], "score": 0.87, "content": "S _ { t }", "type": "inline_equation" }, { "bbox": [ 401, 187, 506, 199 ], "score": 1.0, "content": ". The set of conversational", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 198, 409, 211 ], "spans": [ { "bbox": [ 106, 198, 236, 211 ], "score": 1.0, "content": "messages obtained up until time", "type": "text" }, { "bbox": [ 237, 199, 242, 208 ], "score": 0.78, "content": "t", "type": "inline_equation" }, { "bbox": [ 242, 198, 409, 211 ], "score": 1.0, "content": "is denoted by Equation (1) shown below:", "type": "text" } ], "index": 11 } ], "index": 8.5 }, { "type": "interline_equation", "bbox": [ 210, 214, 401, 229 ], "lines": [ { "bbox": [ 210, 214, 401, 229 ], "spans": [ { "bbox": [ 210, 214, 401, 229 ], "score": 0.89, "content": "\\mathcal { M } _ { t } = \\{ ( \\mathbb { Z } _ { 0 } , S _ { 0 } ) , . . . , ( \\mathbb { Z } _ { t } , S _ { t } ) \\} = \\{ ( \\mathbb { Z } _ { i } , S _ { i } ) \\} | _ { i = 0 } ^ { t }", "type": "interline_equation", "image_path": "584a2523cce87559c9bcc0b164d41a42e1c2db642aaacb258252ec360bcd3b7b.jpg" } ] } ], "index": 12, "virtual_lines": [ { "bbox": [ 210, 214, 401, 229 ], "spans": [], "index": 12 } ] }, { "type": "text", "bbox": [ 106, 239, 505, 284 ], "lines": [ { "bbox": [ 106, 240, 505, 251 ], "spans": [ { "bbox": [ 106, 240, 198, 251 ], "score": 1.0, "content": "At the next time step,", "type": "text" }, { "bbox": [ 198, 240, 220, 250 ], "score": 0.88, "content": "t + 1", "type": "inline_equation" }, { "bbox": [ 221, 240, 273, 251 ], "score": 1.0, "content": ", the AI user", "type": "text" }, { "bbox": [ 273, 240, 282, 249 ], "score": 0.81, "content": "\\mathcal { U }", "type": "inline_equation" }, { "bbox": [ 283, 240, 469, 251 ], "score": 1.0, "content": "takes the historical conversation message set", "type": "text" }, { "bbox": [ 470, 240, 486, 250 ], "score": 0.89, "content": "\\mathcal { M } _ { t }", "type": "inline_equation" }, { "bbox": [ 487, 240, 505, 251 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 250, 506, 263 ], "spans": [ { "bbox": [ 105, 250, 211, 263 ], "score": 1.0, "content": "provides a new instruction", "type": "text" }, { "bbox": [ 212, 251, 232, 262 ], "score": 0.91, "content": "\\mathcal { T } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 232, 250, 474, 263 ], "score": 1.0, "content": ", as shown in Equation (2). The produced instruction message", "type": "text" }, { "bbox": [ 474, 251, 495, 262 ], "score": 0.91, "content": "\\mathcal { T } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 495, 250, 506, 263 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 261, 505, 273 ], "spans": [ { "bbox": [ 105, 261, 254, 273 ], "score": 1.0, "content": "then passed, along with message set", "type": "text" }, { "bbox": [ 255, 262, 271, 272 ], "score": 0.9, "content": "\\mathcal { M } _ { t }", "type": "inline_equation" }, { "bbox": [ 271, 261, 351, 273 ], "score": 1.0, "content": ", to the AI assistant", "type": "text" }, { "bbox": [ 351, 262, 360, 271 ], "score": 0.8, "content": "\\mathcal { A }", "type": "inline_equation" }, { "bbox": [ 360, 261, 505, 273 ], "score": 1.0, "content": ". The AI assistant will then respond", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 272, 304, 285 ], "spans": [ { "bbox": [ 105, 272, 217, 285 ], "score": 1.0, "content": "with a solution, denoted by", "type": "text" }, { "bbox": [ 217, 272, 237, 284 ], "score": 0.92, "content": "\\boldsymbol { S } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 238, 272, 304, 285 ], "score": 1.0, "content": "in Equation (3):", "type": "text" } ], "index": 16 } ], "index": 14.5 }, { "type": "interline_equation", "bbox": [ 172, 290, 461, 304 ], "lines": [ { "bbox": [ 172, 290, 461, 304 ], "spans": [ { "bbox": [ 172, 290, 461, 304 ], "score": 0.57, "content": "\\mathcal { T } _ { t + 1 } = \\mathcal { U } ( \\mathcal { M } t ) \\qquad ( 2 ) \\qquad \\mathcal { S } t + 1 = \\mathcal { A } ( \\mathcal { M } t , \\mathcal { I } t + 1 )", "type": "interline_equation", "image_path": "d8aeb1e12cd4bac821651717bfb612231d32faa6dd413caca233a48f894d24c4.jpg" } ] } ], "index": 17, "virtual_lines": [ { "bbox": [ 172, 290, 461, 304 ], "spans": [], "index": 17 } ] }, { "type": "text", "bbox": [ 107, 307, 505, 330 ], "lines": [ { "bbox": [ 105, 306, 505, 321 ], "spans": [ { "bbox": [ 105, 306, 218, 321 ], "score": 1.0, "content": "After obtaining the solution", "type": "text" }, { "bbox": [ 219, 308, 240, 319 ], "score": 0.91, "content": "\\boldsymbol { S } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 240, 306, 311, 321 ], "score": 1.0, "content": "to the instruction", "type": "text" }, { "bbox": [ 311, 308, 331, 319 ], "score": 0.91, "content": "\\mathcal { T } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 331, 306, 505, 321 ], "score": 1.0, "content": ", the message set is updated using Equation", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 317, 189, 332 ], "spans": [ { "bbox": [ 105, 317, 158, 332 ], "score": 1.0, "content": "(4) to obtain", "type": "text" }, { "bbox": [ 158, 319, 184, 330 ], "score": 0.92, "content": "\\mathcal { M } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 185, 317, 189, 332 ], "score": 1.0, "content": ":", "type": "text" } ], "index": 19 } ], "index": 18.5 }, { "type": "interline_equation", "bbox": [ 245, 345, 366, 358 ], "lines": [ { "bbox": [ 245, 345, 366, 358 ], "spans": [ { "bbox": [ 245, 345, 366, 358 ], "score": 0.93, "content": "\\mathcal { M } _ { t + 1 } \\mathcal { M } _ { t } \\cup ( \\mathcal { T } _ { t + 1 } , S _ { t + 1 } )", "type": "interline_equation", "image_path": "28d361b73b4a4a05c7bbcb3c2fc41fdf3966e24da4d895590ad198a0b7cad260.jpg" } ] } ], "index": 20, "virtual_lines": [ { "bbox": [ 245, 345, 366, 358 ], "spans": [], "index": 20 } ] }, { "type": "text", "bbox": [ 106, 366, 505, 443 ], "lines": [ { "bbox": [ 105, 365, 505, 378 ], "spans": [ { "bbox": [ 105, 365, 505, 378 ], "score": 1.0, "content": "Note that the formulation above not only models AI-AI communicative scenarios, but it can also be", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 377, 507, 389 ], "spans": [ { "bbox": [ 106, 377, 507, 389 ], "score": 1.0, "content": "easily extended to model human-AI communication or communication between more than two agents.", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 389, 506, 401 ], "spans": [ { "bbox": [ 106, 389, 506, 401 ], "score": 1.0, "content": "Specifically, we can use message-passing graphs to model communication between an arbitrary", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 398, 506, 412 ], "spans": [ { "bbox": [ 105, 398, 506, 412 ], "score": 1.0, "content": "number of agents. In Figure 1, we observe that the AI user initiates the installation and import of", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 410, 505, 421 ], "spans": [ { "bbox": [ 106, 410, 505, 421 ], "score": 1.0, "content": "essential Python libraries for sentiment analysis and stock trading by instructing the AI assistant", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 420, 506, 434 ], "spans": [ { "bbox": [ 105, 420, 506, 434 ], "score": 1.0, "content": "through conversations. This example is drawn from our experiments, and the entire conversation is", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 432, 213, 444 ], "spans": [ { "bbox": [ 106, 432, 213, 444 ], "score": 1.0, "content": "available in the Appendix.", "type": "text" } ], "index": 27 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 447, 505, 492 ], "lines": [ { "bbox": [ 106, 447, 505, 460 ], "spans": [ { "bbox": [ 106, 447, 505, 460 ], "score": 1.0, "content": "Critic-In-The-Loop. To enhance the controllability of the role-playing framework, we introduce", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 458, 507, 472 ], "spans": [ { "bbox": [ 105, 458, 507, 472 ], "score": 1.0, "content": "a critic agent capable of selecting proposals from or providing feedback to the role-playing agents.", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 469, 505, 483 ], "spans": [ { "bbox": [ 105, 469, 505, 483 ], "score": 1.0, "content": "This enables tree-search-like decision-making for task-solving. In practice, the critic can be either an", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 481, 501, 493 ], "spans": [ { "bbox": [ 106, 481, 501, 493 ], "score": 1.0, "content": "AI agent or a human. The detailed implementation and case studies can be found in the Appendix.", "type": "text" } ], "index": 31 } ], "index": 29.5 }, { "type": "title", "bbox": [ 107, 505, 219, 517 ], "lines": [ { "bbox": [ 105, 502, 221, 520 ], "spans": [ { "bbox": [ 105, 502, 221, 520 ], "score": 1.0, "content": "3.2 Inception Prompting", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 106, 524, 505, 722 ], "lines": [ { "bbox": [ 106, 525, 505, 538 ], "spans": [ { "bbox": [ 106, 525, 505, 538 ], "score": 1.0, "content": "Since prompt engineering is crucial to our role-playing framework, this section delves deeply into", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 536, 505, 549 ], "spans": [ { "bbox": [ 105, 536, 505, 549 ], "score": 1.0, "content": "our prompting techniques. Our prompt engineering occurs solely at the beginning of role-playing, for", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 548, 505, 559 ], "spans": [ { "bbox": [ 106, 548, 505, 559 ], "score": 1.0, "content": "task specification and role assignment. Once the conversation phase commences, the AI assistant", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 559, 505, 570 ], "spans": [ { "bbox": [ 106, 559, 505, 570 ], "score": 1.0, "content": "and AI user prompt each other automatically in a loop until termination. As such, we refer to our", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 569, 506, 582 ], "spans": [ { "bbox": [ 106, 569, 506, 582 ], "score": 1.0, "content": "technique as Inception Prompting. Our Inception prompt consists of three prompts: the task specifier", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 580, 506, 593 ], "spans": [ { "bbox": [ 106, 580, 138, 593 ], "score": 1.0, "content": "prompt", "type": "text" }, { "bbox": [ 138, 580, 154, 591 ], "score": 0.9, "content": "\\mathcal { P } _ { T }", "type": "inline_equation" }, { "bbox": [ 154, 580, 272, 593 ], "score": 1.0, "content": ", the assistant system prompt", "type": "text" }, { "bbox": [ 272, 580, 287, 591 ], "score": 0.92, "content": "\\mathcal { P } _ { A }", "type": "inline_equation" }, { "bbox": [ 288, 580, 406, 593 ], "score": 1.0, "content": ", and the user system prompt", "type": "text" }, { "bbox": [ 406, 580, 420, 591 ], "score": 0.89, "content": "\\mathcal { P } _ { \\mathcal { U } }", "type": "inline_equation" }, { "bbox": [ 421, 580, 506, 593 ], "score": 1.0, "content": ". As an example, we", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 591, 505, 603 ], "spans": [ { "bbox": [ 106, 591, 258, 603 ], "score": 1.0, "content": "consider the inception prompt of the", "type": "text" }, { "bbox": [ 259, 591, 271, 601 ], "score": 0.27, "content": "A I", "type": "inline_equation" }, { "bbox": [ 272, 591, 492, 603 ], "score": 1.0, "content": "Society scenario. The templates for these prompts of", "type": "text" }, { "bbox": [ 492, 591, 505, 601 ], "score": 0.52, "content": "A I", "type": "inline_equation" } ], "index": 39 }, { "bbox": [ 105, 601, 505, 614 ], "spans": [ { "bbox": [ 105, 601, 505, 614 ], "score": 1.0, "content": "Society role-playing are shown in Figure 2. The task specifier prompt contains information about", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 613, 506, 625 ], "spans": [ { "bbox": [ 106, 613, 506, 625 ], "score": 1.0, "content": "the roles of the AI assistant and AI user in the role-playing session. Therefore, the task specifier", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 623, 506, 636 ], "spans": [ { "bbox": [ 105, 623, 506, 636 ], "score": 1.0, "content": "agent can take a preliminary task/idea as input and generate a specific task using imagination. The AI", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 635, 506, 647 ], "spans": [ { "bbox": [ 106, 635, 204, 647 ], "score": 1.0, "content": "assistant system prompt", "type": "text" }, { "bbox": [ 204, 635, 219, 646 ], "score": 0.89, "content": "\\mathcal { P } _ { A }", "type": "inline_equation" }, { "bbox": [ 219, 635, 345, 647 ], "score": 1.0, "content": "and the AI user system prompt", "type": "text" }, { "bbox": [ 345, 635, 360, 646 ], "score": 0.89, "content": "\\mathcal { P } _ { \\mathcal { U } }", "type": "inline_equation" }, { "bbox": [ 360, 635, 506, 647 ], "score": 1.0, "content": "are mostly symmetrical and include", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 646, 506, 657 ], "spans": [ { "bbox": [ 106, 646, 506, 657 ], "score": 1.0, "content": "information about the assigned task and roles, communication protocols, termination conditions, and", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 657, 506, 668 ], "spans": [ { "bbox": [ 106, 657, 506, 668 ], "score": 1.0, "content": "constraints or requirements to avoid unwanted behaviors. The prompt designs for both roles are", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 667, 506, 680 ], "spans": [ { "bbox": [ 106, 667, 506, 680 ], "score": 1.0, "content": "crucial to achieve autonomous cooperation between agents. It is non-trivial to engineer prompts that", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 677, 505, 691 ], "spans": [ { "bbox": [ 105, 677, 505, 691 ], "score": 1.0, "content": "ensure agents act in alignment with our intentions. We take the prompt templates from the AI Society", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 689, 506, 702 ], "spans": [ { "bbox": [ 106, 689, 506, 702 ], "score": 1.0, "content": "in Figure 2 as an example to explain our key design choices. The prompts used for the Code scenario", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 699, 506, 713 ], "spans": [ { "bbox": [ 105, 699, 506, 713 ], "score": 1.0, "content": "follow a similar sprint as the AI society scenario, but with some additional engineering related to", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 712, 330, 723 ], "spans": [ { "bbox": [ 105, 712, 330, 723 ], "score": 1.0, "content": "programming languages. More details in the Appendix.", "type": "text" } ], "index": 50 } ], "index": 41.5 } ], "page_idx": 4, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 741, 309, 750 ], "lines": [ { "bbox": [ 301, 740, 310, 753 ], "spans": [ { "bbox": [ 301, 740, 310, 753 ], "score": 1.0, "content": "5", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 106, 70, 506, 138 ], "lines": [], "index": 2.5, "bbox_fs": [ 104, 69, 508, 140 ], "lines_deleted": true }, { "type": "text", "bbox": [ 106, 143, 505, 210 ], "lines": [ { "bbox": [ 106, 144, 504, 155 ], "spans": [ { "bbox": [ 106, 144, 495, 155 ], "score": 1.0, "content": "Conversation Towards Task-Solving. After the role assignment is completed, the AI assistant", "type": "text" }, { "bbox": [ 495, 144, 504, 154 ], "score": 0.68, "content": "\\mathcal { A }", "type": "inline_equation" } ], "index": 6 }, { "bbox": [ 105, 154, 506, 167 ], "spans": [ { "bbox": [ 105, 154, 157, 167 ], "score": 1.0, "content": "and AI user", "type": "text" }, { "bbox": [ 157, 155, 167, 165 ], "score": 0.7, "content": "\\mathcal { U }", "type": "inline_equation" }, { "bbox": [ 167, 154, 506, 167 ], "score": 1.0, "content": "will collaborate in an instruction-following manner to accomplish the task. In the", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 164, 506, 178 ], "spans": [ { "bbox": [ 105, 164, 506, 178 ], "score": 1.0, "content": "AI assistant-user scenario, the AI user is responsible for providing instructions, and the assistant", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 176, 506, 189 ], "spans": [ { "bbox": [ 105, 176, 506, 189 ], "score": 1.0, "content": "is expected to respond with a solution that fulfills the instructions. Formally, we denote the user", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 187, 506, 199 ], "spans": [ { "bbox": [ 105, 187, 250, 199 ], "score": 1.0, "content": "instruction message obtained at time", "type": "text" }, { "bbox": [ 250, 188, 255, 197 ], "score": 0.78, "content": "t", "type": "inline_equation" }, { "bbox": [ 256, 187, 267, 199 ], "score": 1.0, "content": "by", "type": "text" }, { "bbox": [ 268, 188, 278, 198 ], "score": 0.87, "content": "\\mathcal { T } _ { t }", "type": "inline_equation" }, { "bbox": [ 278, 187, 390, 199 ], "score": 1.0, "content": "and the assistant solution by", "type": "text" }, { "bbox": [ 390, 187, 401, 198 ], "score": 0.87, "content": "S _ { t }", "type": "inline_equation" }, { "bbox": [ 401, 187, 506, 199 ], "score": 1.0, "content": ". The set of conversational", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 198, 409, 211 ], "spans": [ { "bbox": [ 106, 198, 236, 211 ], "score": 1.0, "content": "messages obtained up until time", "type": "text" }, { "bbox": [ 237, 199, 242, 208 ], "score": 0.78, "content": "t", "type": "inline_equation" }, { "bbox": [ 242, 198, 409, 211 ], "score": 1.0, "content": "is denoted by Equation (1) shown below:", "type": "text" } ], "index": 11 } ], "index": 8.5, "bbox_fs": [ 105, 144, 506, 211 ] }, { "type": "interline_equation", "bbox": [ 210, 214, 401, 229 ], "lines": [ { "bbox": [ 210, 214, 401, 229 ], "spans": [ { "bbox": [ 210, 214, 401, 229 ], "score": 0.89, "content": "\\mathcal { M } _ { t } = \\{ ( \\mathbb { Z } _ { 0 } , S _ { 0 } ) , . . . , ( \\mathbb { Z } _ { t } , S _ { t } ) \\} = \\{ ( \\mathbb { Z } _ { i } , S _ { i } ) \\} | _ { i = 0 } ^ { t }", "type": "interline_equation", "image_path": "584a2523cce87559c9bcc0b164d41a42e1c2db642aaacb258252ec360bcd3b7b.jpg" } ] } ], "index": 12, "virtual_lines": [ { "bbox": [ 210, 214, 401, 229 ], "spans": [], "index": 12 } ] }, { "type": "text", "bbox": [ 106, 239, 505, 284 ], "lines": [ { "bbox": [ 106, 240, 505, 251 ], "spans": [ { "bbox": [ 106, 240, 198, 251 ], "score": 1.0, "content": "At the next time step,", "type": "text" }, { "bbox": [ 198, 240, 220, 250 ], "score": 0.88, "content": "t + 1", "type": "inline_equation" }, { "bbox": [ 221, 240, 273, 251 ], "score": 1.0, "content": ", the AI user", "type": "text" }, { "bbox": [ 273, 240, 282, 249 ], "score": 0.81, "content": "\\mathcal { U }", "type": "inline_equation" }, { "bbox": [ 283, 240, 469, 251 ], "score": 1.0, "content": "takes the historical conversation message set", "type": "text" }, { "bbox": [ 470, 240, 486, 250 ], "score": 0.89, "content": "\\mathcal { M } _ { t }", "type": "inline_equation" }, { "bbox": [ 487, 240, 505, 251 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 250, 506, 263 ], "spans": [ { "bbox": [ 105, 250, 211, 263 ], "score": 1.0, "content": "provides a new instruction", "type": "text" }, { "bbox": [ 212, 251, 232, 262 ], "score": 0.91, "content": "\\mathcal { T } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 232, 250, 474, 263 ], "score": 1.0, "content": ", as shown in Equation (2). The produced instruction message", "type": "text" }, { "bbox": [ 474, 251, 495, 262 ], "score": 0.91, "content": "\\mathcal { T } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 495, 250, 506, 263 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 261, 505, 273 ], "spans": [ { "bbox": [ 105, 261, 254, 273 ], "score": 1.0, "content": "then passed, along with message set", "type": "text" }, { "bbox": [ 255, 262, 271, 272 ], "score": 0.9, "content": "\\mathcal { M } _ { t }", "type": "inline_equation" }, { "bbox": [ 271, 261, 351, 273 ], "score": 1.0, "content": ", to the AI assistant", "type": "text" }, { "bbox": [ 351, 262, 360, 271 ], "score": 0.8, "content": "\\mathcal { A }", "type": "inline_equation" }, { "bbox": [ 360, 261, 505, 273 ], "score": 1.0, "content": ". The AI assistant will then respond", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 272, 304, 285 ], "spans": [ { "bbox": [ 105, 272, 217, 285 ], "score": 1.0, "content": "with a solution, denoted by", "type": "text" }, { "bbox": [ 217, 272, 237, 284 ], "score": 0.92, "content": "\\boldsymbol { S } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 238, 272, 304, 285 ], "score": 1.0, "content": "in Equation (3):", "type": "text" } ], "index": 16 } ], "index": 14.5, "bbox_fs": [ 105, 240, 506, 285 ] }, { "type": "interline_equation", "bbox": [ 172, 290, 461, 304 ], "lines": [ { "bbox": [ 172, 290, 461, 304 ], "spans": [ { "bbox": [ 172, 290, 461, 304 ], "score": 0.57, "content": "\\mathcal { T } _ { t + 1 } = \\mathcal { U } ( \\mathcal { M } t ) \\qquad ( 2 ) \\qquad \\mathcal { S } t + 1 = \\mathcal { A } ( \\mathcal { M } t , \\mathcal { I } t + 1 )", "type": "interline_equation", "image_path": "d8aeb1e12cd4bac821651717bfb612231d32faa6dd413caca233a48f894d24c4.jpg" } ] } ], "index": 17, "virtual_lines": [ { "bbox": [ 172, 290, 461, 304 ], "spans": [], "index": 17 } ] }, { "type": "text", "bbox": [ 107, 307, 505, 330 ], "lines": [ { "bbox": [ 105, 306, 505, 321 ], "spans": [ { "bbox": [ 105, 306, 218, 321 ], "score": 1.0, "content": "After obtaining the solution", "type": "text" }, { "bbox": [ 219, 308, 240, 319 ], "score": 0.91, "content": "\\boldsymbol { S } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 240, 306, 311, 321 ], "score": 1.0, "content": "to the instruction", "type": "text" }, { "bbox": [ 311, 308, 331, 319 ], "score": 0.91, "content": "\\mathcal { T } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 331, 306, 505, 321 ], "score": 1.0, "content": ", the message set is updated using Equation", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 317, 189, 332 ], "spans": [ { "bbox": [ 105, 317, 158, 332 ], "score": 1.0, "content": "(4) to obtain", "type": "text" }, { "bbox": [ 158, 319, 184, 330 ], "score": 0.92, "content": "\\mathcal { M } _ { t + 1 }", "type": "inline_equation" }, { "bbox": [ 185, 317, 189, 332 ], "score": 1.0, "content": ":", "type": "text" } ], "index": 19 } ], "index": 18.5, "bbox_fs": [ 105, 306, 505, 332 ] }, { "type": "interline_equation", "bbox": [ 245, 345, 366, 358 ], "lines": [ { "bbox": [ 245, 345, 366, 358 ], "spans": [ { "bbox": [ 245, 345, 366, 358 ], "score": 0.93, "content": "\\mathcal { M } _ { t + 1 } \\mathcal { M } _ { t } \\cup ( \\mathcal { T } _ { t + 1 } , S _ { t + 1 } )", "type": "interline_equation", "image_path": "28d361b73b4a4a05c7bbcb3c2fc41fdf3966e24da4d895590ad198a0b7cad260.jpg" } ] } ], "index": 20, "virtual_lines": [ { "bbox": [ 245, 345, 366, 358 ], "spans": [], "index": 20 } ] }, { "type": "text", "bbox": [ 106, 366, 505, 443 ], "lines": [ { "bbox": [ 105, 365, 505, 378 ], "spans": [ { "bbox": [ 105, 365, 505, 378 ], "score": 1.0, "content": "Note that the formulation above not only models AI-AI communicative scenarios, but it can also be", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 377, 507, 389 ], "spans": [ { "bbox": [ 106, 377, 507, 389 ], "score": 1.0, "content": "easily extended to model human-AI communication or communication between more than two agents.", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 389, 506, 401 ], "spans": [ { "bbox": [ 106, 389, 506, 401 ], "score": 1.0, "content": "Specifically, we can use message-passing graphs to model communication between an arbitrary", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 398, 506, 412 ], "spans": [ { "bbox": [ 105, 398, 506, 412 ], "score": 1.0, "content": "number of agents. In Figure 1, we observe that the AI user initiates the installation and import of", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 410, 505, 421 ], "spans": [ { "bbox": [ 106, 410, 505, 421 ], "score": 1.0, "content": "essential Python libraries for sentiment analysis and stock trading by instructing the AI assistant", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 420, 506, 434 ], "spans": [ { "bbox": [ 105, 420, 506, 434 ], "score": 1.0, "content": "through conversations. This example is drawn from our experiments, and the entire conversation is", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 432, 213, 444 ], "spans": [ { "bbox": [ 106, 432, 213, 444 ], "score": 1.0, "content": "available in the Appendix.", "type": "text" } ], "index": 27 } ], "index": 24, "bbox_fs": [ 105, 365, 507, 444 ] }, { "type": "text", "bbox": [ 107, 447, 505, 492 ], "lines": [ { "bbox": [ 106, 447, 505, 460 ], "spans": [ { "bbox": [ 106, 447, 505, 460 ], "score": 1.0, "content": "Critic-In-The-Loop. To enhance the controllability of the role-playing framework, we introduce", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 458, 507, 472 ], "spans": [ { "bbox": [ 105, 458, 507, 472 ], "score": 1.0, "content": "a critic agent capable of selecting proposals from or providing feedback to the role-playing agents.", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 469, 505, 483 ], "spans": [ { "bbox": [ 105, 469, 505, 483 ], "score": 1.0, "content": "This enables tree-search-like decision-making for task-solving. In practice, the critic can be either an", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 481, 501, 493 ], "spans": [ { "bbox": [ 106, 481, 501, 493 ], "score": 1.0, "content": "AI agent or a human. The detailed implementation and case studies can be found in the Appendix.", "type": "text" } ], "index": 31 } ], "index": 29.5, "bbox_fs": [ 105, 447, 507, 493 ] }, { "type": "title", "bbox": [ 107, 505, 219, 517 ], "lines": [ { "bbox": [ 105, 502, 221, 520 ], "spans": [ { "bbox": [ 105, 502, 221, 520 ], "score": 1.0, "content": "3.2 Inception Prompting", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 106, 524, 505, 722 ], "lines": [ { "bbox": [ 106, 525, 505, 538 ], "spans": [ { "bbox": [ 106, 525, 505, 538 ], "score": 1.0, "content": "Since prompt engineering is crucial to our role-playing framework, this section delves deeply into", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 536, 505, 549 ], "spans": [ { "bbox": [ 105, 536, 505, 549 ], "score": 1.0, "content": "our prompting techniques. Our prompt engineering occurs solely at the beginning of role-playing, for", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 548, 505, 559 ], "spans": [ { "bbox": [ 106, 548, 505, 559 ], "score": 1.0, "content": "task specification and role assignment. Once the conversation phase commences, the AI assistant", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 559, 505, 570 ], "spans": [ { "bbox": [ 106, 559, 505, 570 ], "score": 1.0, "content": "and AI user prompt each other automatically in a loop until termination. As such, we refer to our", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 569, 506, 582 ], "spans": [ { "bbox": [ 106, 569, 506, 582 ], "score": 1.0, "content": "technique as Inception Prompting. Our Inception prompt consists of three prompts: the task specifier", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 580, 506, 593 ], "spans": [ { "bbox": [ 106, 580, 138, 593 ], "score": 1.0, "content": "prompt", "type": "text" }, { "bbox": [ 138, 580, 154, 591 ], "score": 0.9, "content": "\\mathcal { P } _ { T }", "type": "inline_equation" }, { "bbox": [ 154, 580, 272, 593 ], "score": 1.0, "content": ", the assistant system prompt", "type": "text" }, { "bbox": [ 272, 580, 287, 591 ], "score": 0.92, "content": "\\mathcal { P } _ { A }", "type": "inline_equation" }, { "bbox": [ 288, 580, 406, 593 ], "score": 1.0, "content": ", and the user system prompt", "type": "text" }, { "bbox": [ 406, 580, 420, 591 ], "score": 0.89, "content": "\\mathcal { P } _ { \\mathcal { U } }", "type": "inline_equation" }, { "bbox": [ 421, 580, 506, 593 ], "score": 1.0, "content": ". As an example, we", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 591, 505, 603 ], "spans": [ { "bbox": [ 106, 591, 258, 603 ], "score": 1.0, "content": "consider the inception prompt of the", "type": "text" }, { "bbox": [ 259, 591, 271, 601 ], "score": 0.27, "content": "A I", "type": "inline_equation" }, { "bbox": [ 272, 591, 492, 603 ], "score": 1.0, "content": "Society scenario. The templates for these prompts of", "type": "text" }, { "bbox": [ 492, 591, 505, 601 ], "score": 0.52, "content": "A I", "type": "inline_equation" } ], "index": 39 }, { "bbox": [ 105, 601, 505, 614 ], "spans": [ { "bbox": [ 105, 601, 505, 614 ], "score": 1.0, "content": "Society role-playing are shown in Figure 2. The task specifier prompt contains information about", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 613, 506, 625 ], "spans": [ { "bbox": [ 106, 613, 506, 625 ], "score": 1.0, "content": "the roles of the AI assistant and AI user in the role-playing session. Therefore, the task specifier", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 623, 506, 636 ], "spans": [ { "bbox": [ 105, 623, 506, 636 ], "score": 1.0, "content": "agent can take a preliminary task/idea as input and generate a specific task using imagination. The AI", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 635, 506, 647 ], "spans": [ { "bbox": [ 106, 635, 204, 647 ], "score": 1.0, "content": "assistant system prompt", "type": "text" }, { "bbox": [ 204, 635, 219, 646 ], "score": 0.89, "content": "\\mathcal { P } _ { A }", "type": "inline_equation" }, { "bbox": [ 219, 635, 345, 647 ], "score": 1.0, "content": "and the AI user system prompt", "type": "text" }, { "bbox": [ 345, 635, 360, 646 ], "score": 0.89, "content": "\\mathcal { P } _ { \\mathcal { U } }", "type": "inline_equation" }, { "bbox": [ 360, 635, 506, 647 ], "score": 1.0, "content": "are mostly symmetrical and include", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 646, 506, 657 ], "spans": [ { "bbox": [ 106, 646, 506, 657 ], "score": 1.0, "content": "information about the assigned task and roles, communication protocols, termination conditions, and", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 657, 506, 668 ], "spans": [ { "bbox": [ 106, 657, 506, 668 ], "score": 1.0, "content": "constraints or requirements to avoid unwanted behaviors. The prompt designs for both roles are", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 667, 506, 680 ], "spans": [ { "bbox": [ 106, 667, 506, 680 ], "score": 1.0, "content": "crucial to achieve autonomous cooperation between agents. It is non-trivial to engineer prompts that", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 677, 505, 691 ], "spans": [ { "bbox": [ 105, 677, 505, 691 ], "score": 1.0, "content": "ensure agents act in alignment with our intentions. We take the prompt templates from the AI Society", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 689, 506, 702 ], "spans": [ { "bbox": [ 106, 689, 506, 702 ], "score": 1.0, "content": "in Figure 2 as an example to explain our key design choices. The prompts used for the Code scenario", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 699, 506, 713 ], "spans": [ { "bbox": [ 105, 699, 506, 713 ], "score": 1.0, "content": "follow a similar sprint as the AI society scenario, but with some additional engineering related to", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 712, 330, 723 ], "spans": [ { "bbox": [ 105, 712, 330, 723 ], "score": 1.0, "content": "programming languages. More details in the Appendix.", "type": "text" } ], "index": 50 } ], "index": 41.5, "bbox_fs": [ 105, 525, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 124, 74, 241, 85 ], "lines": [ { "bbox": [ 123, 72, 243, 87 ], "spans": [ { "bbox": [ 123, 72, 243, 87 ], "score": 1.0, "content": "AI Society Inception Prompt", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "title", "bbox": [ 121, 93, 220, 104 ], "lines": [ { "bbox": [ 120, 92, 221, 104 ], "spans": [ { "bbox": [ 120, 92, 221, 104 ], "score": 1.0, "content": "Task Specifier Prompt:", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 120, 107, 465, 131 ], "lines": [ { "bbox": [ 119, 106, 416, 116 ], "spans": [ { "bbox": [ 119, 106, 416, 116 ], "score": 1.0, "content": "Here is a task that will help to complete: .", "type": "text" } ], "index": 2 }, { "bbox": [ 119, 114, 338, 124 ], "spans": [ { "bbox": [ 119, 114, 338, 124 ], "score": 1.0, "content": "Please make it more specific. Be creative and imaginative.", "type": "text" } ], "index": 3 }, { "bbox": [ 119, 122, 467, 132 ], "spans": [ { "bbox": [ 119, 122, 467, 132 ], "score": 1.0, "content": "Please reply with the specified task in words or less. Do not add anything else.", "type": "text" } ], "index": 4 } ], "index": 3 }, { "type": "title", "bbox": [ 121, 157, 231, 167 ], "lines": [ { "bbox": [ 120, 156, 232, 168 ], "spans": [ { "bbox": [ 120, 156, 232, 168 ], "score": 1.0, "content": "Assistant System Prompt:", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "title", "bbox": [ 251, 157, 343, 167 ], "lines": [ { "bbox": [ 250, 155, 345, 168 ], "spans": [ { "bbox": [ 250, 155, 345, 168 ], "score": 1.0, "content": "User System Prompt:", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 250, 171, 488, 235 ], "lines": [ { "bbox": [ 250, 170, 488, 180 ], "spans": [ { "bbox": [ 250, 170, 488, 180 ], "score": 1.0, "content": "Never forget you are a and I am a .", "type": "text" } ], "index": 7 }, { "bbox": [ 250, 178, 424, 188 ], "spans": [ { "bbox": [ 250, 178, 424, 188 ], "score": 1.0, "content": "Never flip roles! You will always instruct me.", "type": "text" } ], "index": 8 }, { "bbox": [ 249, 185, 472, 197 ], "spans": [ { "bbox": [ 249, 185, 472, 197 ], "score": 1.0, "content": "We share a common interest in collaborating to successfully", "type": "text" } ], "index": 9 }, { "bbox": [ 250, 195, 314, 203 ], "spans": [ { "bbox": [ 250, 195, 314, 203 ], "score": 1.0, "content": "complete a task.", "type": "text" } ], "index": 10 }, { "bbox": [ 250, 202, 390, 212 ], "spans": [ { "bbox": [ 250, 202, 390, 212 ], "score": 1.0, "content": "I must help you to complete the task.", "type": "text" } ], "index": 11 }, { "bbox": [ 250, 210, 431, 219 ], "spans": [ { "bbox": [ 250, 210, 431, 219 ], "score": 1.0, "content": "Here is the task: . Never forget our task!", "type": "text" } ], "index": 12 }, { "bbox": [ 250, 218, 476, 228 ], "spans": [ { "bbox": [ 250, 218, 476, 228 ], "score": 1.0, "content": "You must instruct me based on my expertise and your needs to", "type": "text" } ], "index": 13 }, { "bbox": [ 250, 226, 435, 236 ], "spans": [ { "bbox": [ 250, 226, 435, 236 ], "score": 1.0, "content": "complete the task ONLY in the following two ways:", "type": "text" } ], "index": 14 } ], "index": 10.5 }, { "type": "text", "bbox": [ 106, 524, 504, 547 ], "lines": [ { "bbox": [ 105, 523, 505, 538 ], "spans": [ { "bbox": [ 105, 523, 505, 538 ], "score": 1.0, "content": "Prompt Engineering. To delve deeper into the details in Figure 2, we start by chunking the various", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 536, 339, 548 ], "spans": [ { "bbox": [ 105, 536, 264, 548 ], "score": 1.0, "content": "parts of the AI assistant system prompt", "type": "text" }, { "bbox": [ 264, 536, 280, 547 ], "score": 0.88, "content": "\\mathcal { P } _ { A }", "type": "inline_equation" }, { "bbox": [ 280, 536, 339, 548 ], "score": 1.0, "content": "shown below:", "type": "text" } ], "index": 16 } ], "index": 15.5 }, { "type": "text", "bbox": [ 110, 560, 506, 722 ], "lines": [ { "bbox": [ 110, 559, 505, 571 ], "spans": [ { "bbox": [ 110, 559, 505, 571 ], "score": 1.0, "content": "• Never forget you are a and I am a . This assigns", "type": "text" } ], "index": 17 }, { "bbox": [ 119, 570, 485, 582 ], "spans": [ { "bbox": [ 119, 570, 485, 582 ], "score": 1.0, "content": "the chosen role to the assistant agent and provides it with information about the user’s role.", "type": "text" } ], "index": 18 }, { "bbox": [ 110, 588, 505, 600 ], "spans": [ { "bbox": [ 110, 588, 505, 600 ], "score": 1.0, "content": "• Never flip roles! Never instruct me! This prevents agents from flipping roles. In", "type": "text" } ], "index": 19 }, { "bbox": [ 119, 599, 505, 611 ], "spans": [ { "bbox": [ 119, 599, 505, 611 ], "score": 1.0, "content": "some cases, we have observed the assistant and the user switching roles, where the assistant", "type": "text" } ], "index": 20 }, { "bbox": [ 120, 610, 459, 622 ], "spans": [ { "bbox": [ 120, 610, 459, 622 ], "score": 1.0, "content": "suddenly takes control and instructs the user, and the user follows those instructions.", "type": "text" } ], "index": 21 }, { "bbox": [ 110, 627, 469, 641 ], "spans": [ { "bbox": [ 110, 627, 469, 641 ], "score": 1.0, "content": "• You must decline my instruction honestly if you cannot perform the", "type": "text" } ], "index": 22 }, { "bbox": [ 120, 638, 479, 652 ], "spans": [ { "bbox": [ 120, 638, 479, 652 ], "score": 1.0, "content": "instruction due to physical, moral, legal reasons or your capability", "type": "text" } ], "index": 23 }, { "bbox": [ 120, 649, 506, 663 ], "spans": [ { "bbox": [ 120, 649, 506, 663 ], "score": 1.0, "content": "and explain the reasons. This prohibits the agent from producing harmful, false, illegal,", "type": "text" } ], "index": 24 }, { "bbox": [ 119, 659, 236, 673 ], "spans": [ { "bbox": [ 119, 659, 236, 673 ], "score": 1.0, "content": "and misleading information.", "type": "text" } ], "index": 25 }, { "bbox": [ 110, 678, 462, 691 ], "spans": [ { "bbox": [ 110, 678, 462, 691 ], "score": 1.0, "content": "• Unless I say the task is completed, you should always start with:", "type": "text" } ], "index": 26 }, { "bbox": [ 119, 689, 479, 702 ], "spans": [ { "bbox": [ 119, 689, 479, 702 ], "score": 1.0, "content": "Solution: . should be specific, and", "type": "text" } ], "index": 27 }, { "bbox": [ 119, 700, 506, 713 ], "spans": [ { "bbox": [ 119, 700, 463, 713 ], "score": 1.0, "content": "provide preferable implementations and examples for task-solving.", "type": "text" }, { "bbox": [ 483, 700, 506, 712 ], "score": 1.0, "content": "This", "type": "text" } ], "index": 28 }, { "bbox": [ 119, 711, 505, 724 ], "spans": [ { "bbox": [ 119, 711, 505, 724 ], "score": 1.0, "content": "encourages the assistant always responds in a consistent format, avoiding any deviation from the", "type": "text" } ], "index": 29 } ], "index": 23 } ], "page_idx": 5, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 742, 309, 750 ], "lines": [ { "bbox": [ 302, 741, 310, 752 ], "spans": [ { "bbox": [ 302, 741, 310, 752 ], "score": 1.0, "content": "6", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 124, 74, 241, 85 ], "lines": [ { "bbox": [ 123, 72, 243, 87 ], "spans": [ { "bbox": [ 123, 72, 243, 87 ], "score": 1.0, "content": "AI Society Inception Prompt", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "title", "bbox": [ 121, 93, 220, 104 ], "lines": [ { "bbox": [ 120, 92, 221, 104 ], "spans": [ { "bbox": [ 120, 92, 221, 104 ], "score": 1.0, "content": "Task Specifier Prompt:", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "list", "bbox": [ 120, 107, 465, 131 ], "lines": [ { "bbox": [ 119, 106, 416, 116 ], "spans": [ { "bbox": [ 119, 106, 416, 116 ], "score": 1.0, "content": "Here is a task that will help to complete: .", "type": "text" } ], "index": 2, "is_list_end_line": true }, { "bbox": [ 119, 114, 338, 124 ], "spans": [ { "bbox": [ 119, 114, 338, 124 ], "score": 1.0, "content": "Please make it more specific. Be creative and imaginative.", "type": "text" } ], "index": 3, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 119, 122, 467, 132 ], "spans": [ { "bbox": [ 119, 122, 467, 132 ], "score": 1.0, "content": "Please reply with the specified task in words or less. Do not add anything else.", "type": "text" } ], "index": 4, "is_list_start_line": true, "is_list_end_line": true } ], "index": 3, "bbox_fs": [ 119, 106, 467, 132 ] }, { "type": "title", "bbox": [ 121, 157, 231, 167 ], "lines": [ { "bbox": [ 120, 156, 232, 168 ], "spans": [ { "bbox": [ 120, 156, 232, 168 ], "score": 1.0, "content": "Assistant System Prompt:", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "title", "bbox": [ 251, 157, 343, 167 ], "lines": [ { "bbox": [ 250, 155, 345, 168 ], "spans": [ { "bbox": [ 250, 155, 345, 168 ], "score": 1.0, "content": "User System Prompt:", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "list", "bbox": [ 250, 171, 488, 235 ], "lines": [ { "bbox": [ 250, 170, 488, 180 ], "spans": [ { "bbox": [ 250, 170, 488, 180 ], "score": 1.0, "content": "Never forget you are a and I am a .", "type": "text" } ], "index": 7 }, { "bbox": [ 250, 178, 424, 188 ], "spans": [ { "bbox": [ 250, 178, 424, 188 ], "score": 1.0, "content": "Never flip roles! You will always instruct me.", "type": "text" } ], "index": 8, "is_list_end_line": true }, { "bbox": [ 249, 185, 472, 197 ], "spans": [ { "bbox": [ 249, 185, 472, 197 ], "score": 1.0, "content": "We share a common interest in collaborating to successfully", "type": "text" } ], "index": 9, "is_list_start_line": true }, { "bbox": [ 250, 195, 314, 203 ], "spans": [ { "bbox": [ 250, 195, 314, 203 ], "score": 1.0, "content": "complete a task.", "type": "text" } ], "index": 10, "is_list_end_line": true }, { "bbox": [ 250, 202, 390, 212 ], "spans": [ { "bbox": [ 250, 202, 390, 212 ], "score": 1.0, "content": "I must help you to complete the task.", "type": "text" } ], "index": 11, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 250, 210, 431, 219 ], "spans": [ { "bbox": [ 250, 210, 431, 219 ], "score": 1.0, "content": "Here is the task: . Never forget our task!", "type": "text" } ], "index": 12, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 250, 218, 476, 228 ], "spans": [ { "bbox": [ 250, 218, 476, 228 ], "score": 1.0, "content": "You must instruct me based on my expertise and your needs to", "type": "text" } ], "index": 13, "is_list_start_line": true }, { "bbox": [ 250, 226, 435, 236 ], "spans": [ { "bbox": [ 250, 226, 435, 236 ], "score": 1.0, "content": "complete the task ONLY in the following two ways:", "type": "text" } ], "index": 14, "is_list_end_line": true } ], "index": 10.5, "bbox_fs": [ 249, 170, 488, 236 ] }, { "type": "text", "bbox": [ 106, 524, 504, 547 ], "lines": [ { "bbox": [ 105, 523, 505, 538 ], "spans": [ { "bbox": [ 105, 523, 505, 538 ], "score": 1.0, "content": "Prompt Engineering. To delve deeper into the details in Figure 2, we start by chunking the various", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 536, 339, 548 ], "spans": [ { "bbox": [ 105, 536, 264, 548 ], "score": 1.0, "content": "parts of the AI assistant system prompt", "type": "text" }, { "bbox": [ 264, 536, 280, 547 ], "score": 0.88, "content": "\\mathcal { P } _ { A }", "type": "inline_equation" }, { "bbox": [ 280, 536, 339, 548 ], "score": 1.0, "content": "shown below:", "type": "text" } ], "index": 16 } ], "index": 15.5, "bbox_fs": [ 105, 523, 505, 548 ] }, { "type": "text", "bbox": [ 110, 560, 506, 722 ], "lines": [ { "bbox": [ 110, 559, 505, 571 ], "spans": [ { "bbox": [ 110, 559, 505, 571 ], "score": 1.0, "content": "• Never forget you are a and I am a . This assigns", "type": "text" } ], "index": 17 }, { "bbox": [ 119, 570, 485, 582 ], "spans": [ { "bbox": [ 119, 570, 485, 582 ], "score": 1.0, "content": "the chosen role to the assistant agent and provides it with information about the user’s role.", "type": "text" } ], "index": 18 }, { "bbox": [ 110, 588, 505, 600 ], "spans": [ { "bbox": [ 110, 588, 505, 600 ], "score": 1.0, "content": "• Never flip roles! Never instruct me! This prevents agents from flipping roles. In", "type": "text" } ], "index": 19 }, { "bbox": [ 119, 599, 505, 611 ], "spans": [ { "bbox": [ 119, 599, 505, 611 ], "score": 1.0, "content": "some cases, we have observed the assistant and the user switching roles, where the assistant", "type": "text" } ], "index": 20 }, { "bbox": [ 120, 610, 459, 622 ], "spans": [ { "bbox": [ 120, 610, 459, 622 ], "score": 1.0, "content": "suddenly takes control and instructs the user, and the user follows those instructions.", "type": "text" } ], "index": 21 }, { "bbox": [ 110, 627, 469, 641 ], "spans": [ { "bbox": [ 110, 627, 469, 641 ], "score": 1.0, "content": "• You must decline my instruction honestly if you cannot perform the", "type": "text" } ], "index": 22 }, { "bbox": [ 120, 638, 479, 652 ], "spans": [ { "bbox": [ 120, 638, 479, 652 ], "score": 1.0, "content": "instruction due to physical, moral, legal reasons or your capability", "type": "text" } ], "index": 23 }, { "bbox": [ 120, 649, 506, 663 ], "spans": [ { "bbox": [ 120, 649, 506, 663 ], "score": 1.0, "content": "and explain the reasons. This prohibits the agent from producing harmful, false, illegal,", "type": "text" } ], "index": 24 }, { "bbox": [ 119, 659, 236, 673 ], "spans": [ { "bbox": [ 119, 659, 236, 673 ], "score": 1.0, "content": "and misleading information.", "type": "text" } ], "index": 25 }, { "bbox": [ 110, 678, 462, 691 ], "spans": [ { "bbox": [ 110, 678, 462, 691 ], "score": 1.0, "content": "• Unless I say the task is completed, you should always start with:", "type": "text" } ], "index": 26 }, { "bbox": [ 119, 689, 479, 702 ], "spans": [ { "bbox": [ 119, 689, 479, 702 ], "score": 1.0, "content": "Solution: . should be specific, and", "type": "text" } ], "index": 27 }, { "bbox": [ 119, 700, 506, 713 ], "spans": [ { "bbox": [ 119, 700, 463, 713 ], "score": 1.0, "content": "provide preferable implementations and examples for task-solving.", "type": "text" }, { "bbox": [ 483, 700, 506, 712 ], "score": 1.0, "content": "This", "type": "text" } ], "index": 28 }, { "bbox": [ 119, 711, 505, 724 ], "spans": [ { "bbox": [ 119, 711, 505, 724 ], "score": 1.0, "content": "encourages the assistant always responds in a consistent format, avoiding any deviation from the", "type": "text" } ], "index": 29 }, { "bbox": [ 120, 72, 505, 85 ], "spans": [ { "bbox": [ 120, 72, 505, 85 ], "score": 1.0, "content": "structure of the conversation, and preventing vague or incomplete responses, which we refer to as", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 119, 82, 309, 96 ], "spans": [ { "bbox": [ 119, 82, 217, 96 ], "score": 1.0, "content": "flake responses, such as", "type": "text", "cross_page": true }, { "bbox": [ 218, 84, 227, 94 ], "score": 0.25, "content": "{ } \" \\mathrm { I }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 227, 82, 309, 96 ], "score": 1.0, "content": "will do something\".", "type": "text", "cross_page": true } ], "index": 1 } ], "index": 23, "bbox_fs": [ 110, 559, 506, 724 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 114, 73, 504, 95 ], "lines": [ { "bbox": [ 120, 72, 505, 85 ], "spans": [ { "bbox": [ 120, 72, 505, 85 ], "score": 1.0, "content": "structure of the conversation, and preventing vague or incomplete responses, which we refer to as", "type": "text" } ], "index": 0 }, { "bbox": [ 119, 82, 309, 96 ], "spans": [ { "bbox": [ 119, 82, 217, 96 ], "score": 1.0, "content": "flake responses, such as", "type": "text" }, { "bbox": [ 218, 84, 227, 94 ], "score": 0.25, "content": "{ } \" \\mathrm { I }", "type": "inline_equation" }, { "bbox": [ 227, 82, 309, 96 ], "score": 1.0, "content": "will do something\".", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 110, 99, 504, 120 ], "lines": [ { "bbox": [ 110, 97, 505, 111 ], "spans": [ { "bbox": [ 110, 97, 505, 111 ], "score": 1.0, "content": "• Always end your solution with: Next request. This ensures that the assistant keeps", "type": "text" } ], "index": 2 }, { "bbox": [ 120, 109, 376, 122 ], "spans": [ { "bbox": [ 120, 109, 376, 122 ], "score": 1.0, "content": "the conversation going by requesting a new instruction to solve.", "type": "text" } ], "index": 3 } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 130, 504, 164 ], "lines": [ { "bbox": [ 105, 130, 506, 144 ], "spans": [ { "bbox": [ 105, 130, 229, 144 ], "score": 1.0, "content": "For the AI user system prompt", "type": "text" }, { "bbox": [ 229, 131, 244, 142 ], "score": 0.89, "content": "\\mathcal { P } _ { \\mathcal { U } }", "type": "inline_equation" }, { "bbox": [ 244, 130, 506, 144 ], "score": 1.0, "content": ", we strive to maintain as much symmetry as possible with respect", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 141, 505, 154 ], "spans": [ { "bbox": [ 105, 141, 505, 154 ], "score": 1.0, "content": "to the AI assistant system prompt. Apart from the opposite role assignment, the user system prompt", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 151, 330, 165 ], "spans": [ { "bbox": [ 105, 151, 330, 165 ], "score": 1.0, "content": "differs from the assistant prompt in the following ways:", "type": "text" } ], "index": 6 } ], "index": 5 }, { "type": "text", "bbox": [ 110, 174, 506, 287 ], "lines": [ { "bbox": [ 109, 173, 500, 186 ], "spans": [ { "bbox": [ 109, 173, 500, 186 ], "score": 1.0, "content": "• You must instruct me ... to complete the task ONLY in the following two", "type": "text" } ], "index": 7 }, { "bbox": [ 119, 185, 489, 198 ], "spans": [ { "bbox": [ 119, 185, 489, 198 ], "score": 1.0, "content": "ways: 1. Instruct with a necessary input: ...; 2. Instruct without", "type": "text" } ], "index": 8 }, { "bbox": [ 120, 196, 505, 208 ], "spans": [ { "bbox": [ 120, 196, 505, 208 ], "score": 1.0, "content": "any input: ... This follows the typical data structure of instruction-following, which allows", "type": "text" } ], "index": 9 }, { "bbox": [ 119, 206, 437, 219 ], "spans": [ { "bbox": [ 119, 206, 437, 219 ], "score": 1.0, "content": "the generated instruction-solution pairs to be easily used for fine-tuning LLMs.", "type": "text" } ], "index": 10 }, { "bbox": [ 112, 221, 479, 234 ], "spans": [ { "bbox": [ 112, 221, 479, 234 ], "score": 1.0, "content": "• Keep giving me instructions and necessary inputs until you think the", "type": "text" } ], "index": 11 }, { "bbox": [ 119, 232, 500, 245 ], "spans": [ { "bbox": [ 119, 232, 500, 245 ], "score": 1.0, "content": "task is completed. When the task is completed, you must only reply with", "type": "text" } ], "index": 12 }, { "bbox": [ 119, 242, 506, 255 ], "spans": [ { "bbox": [ 119, 243, 291, 254 ], "score": 1.0, "content": "a single word .", "type": "text" }, { "bbox": [ 300, 242, 506, 255 ], "score": 1.0, "content": "We introduce an end-of-task token, namely,", "type": "text" } ], "index": 13 }, { "bbox": [ 119, 254, 506, 266 ], "spans": [ { "bbox": [ 119, 254, 506, 266 ], "score": 1.0, "content": ". This token is used once the user believes the task is done. This ensures", "type": "text" } ], "index": 14 }, { "bbox": [ 120, 264, 505, 277 ], "spans": [ { "bbox": [ 120, 264, 505, 277 ], "score": 1.0, "content": "that the chat is terminated when the user is satisfied. Without doing so, the agents might fall into", "type": "text" } ], "index": 15 }, { "bbox": [ 119, 275, 502, 289 ], "spans": [ { "bbox": [ 119, 275, 502, 289 ], "score": 1.0, "content": "a chatting loop where they keep on saying “thank you” to each other or “goodbye” indefinitely.", "type": "text" } ], "index": 16 } ], "index": 11.5 }, { "type": "title", "bbox": [ 107, 302, 191, 316 ], "lines": [ { "bbox": [ 104, 300, 193, 318 ], "spans": [ { "bbox": [ 104, 300, 193, 318 ], "score": 1.0, "content": "4 Experiments", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 106, 327, 505, 437 ], "lines": [ { "bbox": [ 105, 326, 505, 340 ], "spans": [ { "bbox": [ 105, 326, 505, 340 ], "score": 1.0, "content": "In this section, we will discuss the various experiments that we conducted to arrive at our final design", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 338, 505, 350 ], "spans": [ { "bbox": [ 106, 338, 505, 350 ], "score": 1.0, "content": "choices. Specifically, we will examine the interesting observations, challenging issues, and several", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 348, 505, 361 ], "spans": [ { "bbox": [ 106, 348, 505, 361 ], "score": 1.0, "content": "examples we have encountered while enabling agents to communicate with each other under different", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 360, 506, 372 ], "spans": [ { "bbox": [ 104, 360, 506, 372 ], "score": 1.0, "content": "prompt design choices to achieve autonomous cooperation. In our experiments, we employed two", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 370, 506, 383 ], "spans": [ { "bbox": [ 105, 370, 506, 383 ], "score": 1.0, "content": "gpt-3.5-turbo agents, referred to as LLM agents for simplicity, with Inception Prompts, as described", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 381, 506, 394 ], "spans": [ { "bbox": [ 105, 381, 506, 394 ], "score": 1.0, "content": "in Section 3.2, to simulate assistant-user cooperation. For our analysis, we set our attention on AI", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 392, 505, 405 ], "spans": [ { "bbox": [ 106, 392, 505, 405 ], "score": 1.0, "content": "Society setting. We also gathered conversational data, named CAMEL AI Society and CAMEL Code", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 403, 505, 415 ], "spans": [ { "bbox": [ 106, 403, 505, 415 ], "score": 1.0, "content": "datasets and problem-solution pairs data named CAMEL Math and CAMEL Science and analyzed", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 414, 506, 427 ], "spans": [ { "bbox": [ 106, 414, 506, 427 ], "score": 1.0, "content": "and evaluated their quality. Moreover, we will discuss potential extensions of our framework and", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 425, 420, 438 ], "spans": [ { "bbox": [ 106, 425, 420, 438 ], "score": 1.0, "content": "highlight both the risks and opportunities that future AI society might present.", "type": "text" } ], "index": 27 } ], "index": 22.5 }, { "type": "title", "bbox": [ 124, 450, 284, 461 ], "lines": [ { "bbox": [ 123, 447, 284, 463 ], "spans": [ { "bbox": [ 123, 447, 284, 463 ], "score": 1.0, "content": "Data Generation Prompts of AI Society", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "title", "bbox": [ 121, 469, 166, 480 ], "lines": [ { "bbox": [ 119, 467, 168, 482 ], "spans": [ { "bbox": [ 119, 467, 168, 482 ], "score": 1.0, "content": "AI Society", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "title", "bbox": [ 121, 488, 256, 498 ], "lines": [ { "bbox": [ 120, 487, 257, 498 ], "spans": [ { "bbox": [ 120, 487, 257, 498 ], "score": 1.0, "content": "Assistant Role Generation Prompt:", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 121, 502, 291, 542 ], "lines": [ { "bbox": [ 119, 501, 293, 511 ], "spans": [ { "bbox": [ 119, 501, 293, 511 ], "score": 1.0, "content": "You are a helpful assistant that can play many", "type": "text" } ], "index": 31 }, { "bbox": [ 120, 509, 286, 519 ], "spans": [ { "bbox": [ 120, 509, 286, 519 ], "score": 1.0, "content": "different roles. Now please list ", "type": "text" } ], "index": 32 }, { "bbox": [ 119, 517, 282, 527 ], "spans": [ { "bbox": [ 119, 517, 282, 527 ], "score": 1.0, "content": "different roles that you can play with your", "type": "text" } ], "index": 33 }, { "bbox": [ 120, 525, 275, 535 ], "spans": [ { "bbox": [ 120, 525, 275, 535 ], "score": 1.0, "content": "expertise in diverse fields. Sort them by", "type": "text" } ], "index": 34 }, { "bbox": [ 120, 533, 285, 542 ], "spans": [ { "bbox": [ 120, 533, 285, 542 ], "score": 1.0, "content": "alphabetical order. No explanation required.", "type": "text" } ], "index": 35 } ], "index": 33 }, { "type": "title", "bbox": [ 305, 488, 423, 498 ], "lines": [ { "bbox": [ 304, 486, 424, 498 ], "spans": [ { "bbox": [ 304, 486, 424, 498 ], "score": 1.0, "content": "User Role Generation Prompt:", "type": "text" } ], "index": 36 } ], "index": 36 }, { "type": "text", "bbox": [ 304, 502, 481, 541 ], "lines": [ { "bbox": [ 303, 501, 482, 510 ], "spans": [ { "bbox": [ 303, 501, 482, 510 ], "score": 1.0, "content": "Please list most common and diverse", "type": "text" } ], "index": 37 }, { "bbox": [ 303, 509, 455, 519 ], "spans": [ { "bbox": [ 303, 509, 455, 519 ], "score": 1.0, "content": "groups of internet users or occupations.", "type": "text" } ], "index": 38 }, { "bbox": [ 303, 517, 432, 527 ], "spans": [ { "bbox": [ 303, 517, 432, 527 ], "score": 1.0, "content": "Use singular form. No explanation.", "type": "text" } ], "index": 39 }, { "bbox": [ 303, 524, 482, 535 ], "spans": [ { "bbox": [ 303, 524, 482, 535 ], "score": 1.0, "content": "Sort them by alphabetical order. No explanation", "type": "text" } ], "index": 40 }, { "bbox": [ 303, 533, 340, 543 ], "spans": [ { "bbox": [ 303, 533, 340, 543 ], "score": 1.0, "content": "required.", "type": "text" } ], "index": 41 } ], "index": 39 }, { "type": "title", "bbox": [ 121, 545, 220, 554 ], "lines": [ { "bbox": [ 120, 543, 221, 555 ], "spans": [ { "bbox": [ 120, 543, 221, 555 ], "score": 1.0, "content": "Task Generation Prompt:", "type": "text" } ], "index": 42 } ], "index": 42 }, { "type": "text", "bbox": [ 119, 558, 462, 574 ], "lines": [ { "bbox": [ 118, 556, 464, 568 ], "spans": [ { "bbox": [ 118, 556, 464, 568 ], "score": 1.0, "content": "List diverse tasks that can assist cooperatively to", "type": "text" } ], "index": 43 }, { "bbox": [ 119, 566, 277, 575 ], "spans": [ { "bbox": [ 119, 566, 277, 575 ], "score": 1.0, "content": "achieve together. Be concise. Be creative.", "type": "text" } ], "index": 44 } ], "index": 43.5 }, { "type": "text", "bbox": [ 107, 593, 504, 627 ], "lines": [ { "bbox": [ 105, 592, 505, 606 ], "spans": [ { "bbox": [ 105, 592, 505, 606 ], "score": 1.0, "content": "Figure 3: Data Generation Prompts. In order to maintain a scalable approach our data parameters", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 604, 506, 618 ], "spans": [ { "bbox": [ 105, 604, 506, 618 ], "score": 1.0, "content": "are generated using an LLM model to reduce human involvement in the generation process. The", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 615, 417, 628 ], "spans": [ { "bbox": [ 105, 615, 417, 628 ], "score": 1.0, "content": "generation prompts for both AI Society dataset are summarized in this figure.", "type": "text" } ], "index": 47 } ], "index": 46 }, { "type": "title", "bbox": [ 108, 646, 246, 658 ], "lines": [ { "bbox": [ 105, 645, 248, 662 ], "spans": [ { "bbox": [ 105, 645, 248, 662 ], "score": 1.0, "content": "4.1 Role-Playing for AI Society", "type": "text" } ], "index": 48 } ], "index": 48 }, { "type": "text", "bbox": [ 107, 667, 505, 722 ], "lines": [ { "bbox": [ 106, 667, 506, 679 ], "spans": [ { "bbox": [ 106, 667, 506, 679 ], "score": 1.0, "content": "To create our AI Society dataset, we have developed a scalable approach that follows a series of steps.", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 677, 506, 691 ], "spans": [ { "bbox": [ 105, 677, 506, 691 ], "score": 1.0, "content": "Firstly, we prompt the LLM agent to generate possible roles for the assistant and the user. We achieve", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 688, 506, 703 ], "spans": [ { "bbox": [ 105, 688, 506, 703 ], "score": 1.0, "content": "this by providing the LLM agent with specific prompts designed to elicit these roles. Next, we ask the", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 699, 505, 713 ], "spans": [ { "bbox": [ 105, 699, 505, 713 ], "score": 1.0, "content": "LLM agent to generate a range of possible tasks that can be solved through collaboration between the", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 710, 505, 723 ], "spans": [ { "bbox": [ 105, 710, 505, 723 ], "score": 1.0, "content": "assistant and user roles generated previously. After generating a range of possible tasks as described", "type": "text" } ], "index": 53 } ], "index": 51 } ], "page_idx": 6, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 741, 309, 750 ], "lines": [ { "bbox": [ 302, 741, 309, 752 ], "spans": [ { "bbox": [ 302, 741, 309, 752 ], "score": 1.0, "content": "7", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 114, 73, 504, 95 ], "lines": [], "index": 0.5, "bbox_fs": [ 119, 72, 505, 96 ], "lines_deleted": true }, { "type": "text", "bbox": [ 110, 99, 504, 120 ], "lines": [ { "bbox": [ 110, 97, 505, 111 ], "spans": [ { "bbox": [ 110, 97, 505, 111 ], "score": 1.0, "content": "• Always end your solution with: Next request. This ensures that the assistant keeps", "type": "text" } ], "index": 2 }, { "bbox": [ 120, 109, 376, 122 ], "spans": [ { "bbox": [ 120, 109, 376, 122 ], "score": 1.0, "content": "the conversation going by requesting a new instruction to solve.", "type": "text" } ], "index": 3 } ], "index": 2.5, "bbox_fs": [ 110, 97, 505, 122 ] }, { "type": "text", "bbox": [ 107, 130, 504, 164 ], "lines": [ { "bbox": [ 105, 130, 506, 144 ], "spans": [ { "bbox": [ 105, 130, 229, 144 ], "score": 1.0, "content": "For the AI user system prompt", "type": "text" }, { "bbox": [ 229, 131, 244, 142 ], "score": 0.89, "content": "\\mathcal { P } _ { \\mathcal { U } }", "type": "inline_equation" }, { "bbox": [ 244, 130, 506, 144 ], "score": 1.0, "content": ", we strive to maintain as much symmetry as possible with respect", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 141, 505, 154 ], "spans": [ { "bbox": [ 105, 141, 505, 154 ], "score": 1.0, "content": "to the AI assistant system prompt. Apart from the opposite role assignment, the user system prompt", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 151, 330, 165 ], "spans": [ { "bbox": [ 105, 151, 330, 165 ], "score": 1.0, "content": "differs from the assistant prompt in the following ways:", "type": "text" } ], "index": 6 } ], "index": 5, "bbox_fs": [ 105, 130, 506, 165 ] }, { "type": "text", "bbox": [ 110, 174, 506, 287 ], "lines": [ { "bbox": [ 109, 173, 500, 186 ], "spans": [ { "bbox": [ 109, 173, 500, 186 ], "score": 1.0, "content": "• You must instruct me ... to complete the task ONLY in the following two", "type": "text" } ], "index": 7 }, { "bbox": [ 119, 185, 489, 198 ], "spans": [ { "bbox": [ 119, 185, 489, 198 ], "score": 1.0, "content": "ways: 1. Instruct with a necessary input: ...; 2. Instruct without", "type": "text" } ], "index": 8 }, { "bbox": [ 120, 196, 505, 208 ], "spans": [ { "bbox": [ 120, 196, 505, 208 ], "score": 1.0, "content": "any input: ... This follows the typical data structure of instruction-following, which allows", "type": "text" } ], "index": 9 }, { "bbox": [ 119, 206, 437, 219 ], "spans": [ { "bbox": [ 119, 206, 437, 219 ], "score": 1.0, "content": "the generated instruction-solution pairs to be easily used for fine-tuning LLMs.", "type": "text" } ], "index": 10 }, { "bbox": [ 112, 221, 479, 234 ], "spans": [ { "bbox": [ 112, 221, 479, 234 ], "score": 1.0, "content": "• Keep giving me instructions and necessary inputs until you think the", "type": "text" } ], "index": 11 }, { "bbox": [ 119, 232, 500, 245 ], "spans": [ { "bbox": [ 119, 232, 500, 245 ], "score": 1.0, "content": "task is completed. When the task is completed, you must only reply with", "type": "text" } ], "index": 12 }, { "bbox": [ 119, 242, 506, 255 ], "spans": [ { "bbox": [ 119, 243, 291, 254 ], "score": 1.0, "content": "a single word .", "type": "text" }, { "bbox": [ 300, 242, 506, 255 ], "score": 1.0, "content": "We introduce an end-of-task token, namely,", "type": "text" } ], "index": 13 }, { "bbox": [ 119, 254, 506, 266 ], "spans": [ { "bbox": [ 119, 254, 506, 266 ], "score": 1.0, "content": ". This token is used once the user believes the task is done. This ensures", "type": "text" } ], "index": 14 }, { "bbox": [ 120, 264, 505, 277 ], "spans": [ { "bbox": [ 120, 264, 505, 277 ], "score": 1.0, "content": "that the chat is terminated when the user is satisfied. Without doing so, the agents might fall into", "type": "text" } ], "index": 15 }, { "bbox": [ 119, 275, 502, 289 ], "spans": [ { "bbox": [ 119, 275, 502, 289 ], "score": 1.0, "content": "a chatting loop where they keep on saying “thank you” to each other or “goodbye” indefinitely.", "type": "text" } ], "index": 16 } ], "index": 11.5, "bbox_fs": [ 109, 173, 506, 289 ] }, { "type": "title", "bbox": [ 107, 302, 191, 316 ], "lines": [ { "bbox": [ 104, 300, 193, 318 ], "spans": [ { "bbox": [ 104, 300, 193, 318 ], "score": 1.0, "content": "4 Experiments", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 106, 327, 505, 437 ], "lines": [ { "bbox": [ 105, 326, 505, 340 ], "spans": [ { "bbox": [ 105, 326, 505, 340 ], "score": 1.0, "content": "In this section, we will discuss the various experiments that we conducted to arrive at our final design", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 338, 505, 350 ], "spans": [ { "bbox": [ 106, 338, 505, 350 ], "score": 1.0, "content": "choices. Specifically, we will examine the interesting observations, challenging issues, and several", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 348, 505, 361 ], "spans": [ { "bbox": [ 106, 348, 505, 361 ], "score": 1.0, "content": "examples we have encountered while enabling agents to communicate with each other under different", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 360, 506, 372 ], "spans": [ { "bbox": [ 104, 360, 506, 372 ], "score": 1.0, "content": "prompt design choices to achieve autonomous cooperation. In our experiments, we employed two", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 370, 506, 383 ], "spans": [ { "bbox": [ 105, 370, 506, 383 ], "score": 1.0, "content": "gpt-3.5-turbo agents, referred to as LLM agents for simplicity, with Inception Prompts, as described", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 381, 506, 394 ], "spans": [ { "bbox": [ 105, 381, 506, 394 ], "score": 1.0, "content": "in Section 3.2, to simulate assistant-user cooperation. For our analysis, we set our attention on AI", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 392, 505, 405 ], "spans": [ { "bbox": [ 106, 392, 505, 405 ], "score": 1.0, "content": "Society setting. We also gathered conversational data, named CAMEL AI Society and CAMEL Code", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 403, 505, 415 ], "spans": [ { "bbox": [ 106, 403, 505, 415 ], "score": 1.0, "content": "datasets and problem-solution pairs data named CAMEL Math and CAMEL Science and analyzed", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 414, 506, 427 ], "spans": [ { "bbox": [ 106, 414, 506, 427 ], "score": 1.0, "content": "and evaluated their quality. Moreover, we will discuss potential extensions of our framework and", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 425, 420, 438 ], "spans": [ { "bbox": [ 106, 425, 420, 438 ], "score": 1.0, "content": "highlight both the risks and opportunities that future AI society might present.", "type": "text" } ], "index": 27 } ], "index": 22.5, "bbox_fs": [ 104, 326, 506, 438 ] }, { "type": "title", "bbox": [ 124, 450, 284, 461 ], "lines": [ { "bbox": [ 123, 447, 284, 463 ], "spans": [ { "bbox": [ 123, 447, 284, 463 ], "score": 1.0, "content": "Data Generation Prompts of AI Society", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "title", "bbox": [ 121, 469, 166, 480 ], "lines": [ { "bbox": [ 119, 467, 168, 482 ], "spans": [ { "bbox": [ 119, 467, 168, 482 ], "score": 1.0, "content": "AI Society", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "title", "bbox": [ 121, 488, 256, 498 ], "lines": [ { "bbox": [ 120, 487, 257, 498 ], "spans": [ { "bbox": [ 120, 487, 257, 498 ], "score": 1.0, "content": "Assistant Role Generation Prompt:", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 121, 502, 291, 542 ], "lines": [ { "bbox": [ 119, 501, 293, 511 ], "spans": [ { "bbox": [ 119, 501, 293, 511 ], "score": 1.0, "content": "You are a helpful assistant that can play many", "type": "text" } ], "index": 31 }, { "bbox": [ 120, 509, 286, 519 ], "spans": [ { "bbox": [ 120, 509, 286, 519 ], "score": 1.0, "content": "different roles. Now please list ", "type": "text" } ], "index": 32 }, { "bbox": [ 119, 517, 282, 527 ], "spans": [ { "bbox": [ 119, 517, 282, 527 ], "score": 1.0, "content": "different roles that you can play with your", "type": "text" } ], "index": 33 }, { "bbox": [ 120, 525, 275, 535 ], "spans": [ { "bbox": [ 120, 525, 275, 535 ], "score": 1.0, "content": "expertise in diverse fields. Sort them by", "type": "text" } ], "index": 34 }, { "bbox": [ 120, 533, 285, 542 ], "spans": [ { "bbox": [ 120, 533, 285, 542 ], "score": 1.0, "content": "alphabetical order. No explanation required.", "type": "text" } ], "index": 35 } ], "index": 33, "bbox_fs": [ 119, 501, 293, 542 ] }, { "type": "title", "bbox": [ 305, 488, 423, 498 ], "lines": [ { "bbox": [ 304, 486, 424, 498 ], "spans": [ { "bbox": [ 304, 486, 424, 498 ], "score": 1.0, "content": "User Role Generation Prompt:", "type": "text" } ], "index": 36 } ], "index": 36 }, { "type": "text", "bbox": [ 304, 502, 481, 541 ], "lines": [ { "bbox": [ 303, 501, 482, 510 ], "spans": [ { "bbox": [ 303, 501, 482, 510 ], "score": 1.0, "content": "Please list most common and diverse", "type": "text" } ], "index": 37 }, { "bbox": [ 303, 509, 455, 519 ], "spans": [ { "bbox": [ 303, 509, 455, 519 ], "score": 1.0, "content": "groups of internet users or occupations.", "type": "text" } ], "index": 38 }, { "bbox": [ 303, 517, 432, 527 ], "spans": [ { "bbox": [ 303, 517, 432, 527 ], "score": 1.0, "content": "Use singular form. No explanation.", "type": "text" } ], "index": 39 }, { "bbox": [ 303, 524, 482, 535 ], "spans": [ { "bbox": [ 303, 524, 482, 535 ], "score": 1.0, "content": "Sort them by alphabetical order. No explanation", "type": "text" } ], "index": 40 }, { "bbox": [ 303, 533, 340, 543 ], "spans": [ { "bbox": [ 303, 533, 340, 543 ], "score": 1.0, "content": "required.", "type": "text" } ], "index": 41 } ], "index": 39, "bbox_fs": [ 303, 501, 482, 543 ] }, { "type": "title", "bbox": [ 121, 545, 220, 554 ], "lines": [ { "bbox": [ 120, 543, 221, 555 ], "spans": [ { "bbox": [ 120, 543, 221, 555 ], "score": 1.0, "content": "Task Generation Prompt:", "type": "text" } ], "index": 42 } ], "index": 42 }, { "type": "text", "bbox": [ 119, 558, 462, 574 ], "lines": [ { "bbox": [ 118, 556, 464, 568 ], "spans": [ { "bbox": [ 118, 556, 464, 568 ], "score": 1.0, "content": "List diverse tasks that can assist cooperatively to", "type": "text" } ], "index": 43 }, { "bbox": [ 119, 566, 277, 575 ], "spans": [ { "bbox": [ 119, 566, 277, 575 ], "score": 1.0, "content": "achieve together. Be concise. Be creative.", "type": "text" } ], "index": 44 } ], "index": 43.5, "bbox_fs": [ 118, 556, 464, 575 ] }, { "type": "text", "bbox": [ 107, 593, 504, 627 ], "lines": [ { "bbox": [ 105, 592, 505, 606 ], "spans": [ { "bbox": [ 105, 592, 505, 606 ], "score": 1.0, "content": "Figure 3: Data Generation Prompts. In order to maintain a scalable approach our data parameters", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 604, 506, 618 ], "spans": [ { "bbox": [ 105, 604, 506, 618 ], "score": 1.0, "content": "are generated using an LLM model to reduce human involvement in the generation process. The", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 615, 417, 628 ], "spans": [ { "bbox": [ 105, 615, 417, 628 ], "score": 1.0, "content": "generation prompts for both AI Society dataset are summarized in this figure.", "type": "text" } ], "index": 47 } ], "index": 46, "bbox_fs": [ 105, 592, 506, 628 ] }, { "type": "title", "bbox": [ 108, 646, 246, 658 ], "lines": [ { "bbox": [ 105, 645, 248, 662 ], "spans": [ { "bbox": [ 105, 645, 248, 662 ], "score": 1.0, "content": "4.1 Role-Playing for AI Society", "type": "text" } ], "index": 48 } ], "index": 48 }, { "type": "text", "bbox": [ 107, 667, 505, 722 ], "lines": [ { "bbox": [ 106, 667, 506, 679 ], "spans": [ { "bbox": [ 106, 667, 506, 679 ], "score": 1.0, "content": "To create our AI Society dataset, we have developed a scalable approach that follows a series of steps.", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 677, 506, 691 ], "spans": [ { "bbox": [ 105, 677, 506, 691 ], "score": 1.0, "content": "Firstly, we prompt the LLM agent to generate possible roles for the assistant and the user. We achieve", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 688, 506, 703 ], "spans": [ { "bbox": [ 105, 688, 506, 703 ], "score": 1.0, "content": "this by providing the LLM agent with specific prompts designed to elicit these roles. Next, we ask the", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 699, 505, 713 ], "spans": [ { "bbox": [ 105, 699, 505, 713 ], "score": 1.0, "content": "LLM agent to generate a range of possible tasks that can be solved through collaboration between the", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 710, 505, 723 ], "spans": [ { "bbox": [ 105, 710, 505, 723 ], "score": 1.0, "content": "assistant and user roles generated previously. After generating a range of possible tasks as described", "type": "text" } ], "index": 53 }, { "bbox": [ 106, 73, 506, 85 ], "spans": [ { "bbox": [ 106, 73, 506, 85 ], "score": 1.0, "content": "in the previous step, we then use the task specifier prompt passed to the LLM agent to make the task", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 105, 83, 506, 96 ], "spans": [ { "bbox": [ 105, 83, 506, 96 ], "score": 1.0, "content": "more specific. The prompts for assistant role generation, user role generation, and task generation", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 105, 94, 505, 106 ], "spans": [ { "bbox": [ 105, 94, 505, 106 ], "score": 1.0, "content": "are shown in Figure 5 (AI Society). For our AI society dataset, we generated 50 assistant roles, 50", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 106, 106, 505, 117 ], "spans": [ { "bbox": [ 106, 106, 505, 117 ], "score": 1.0, "content": "user roles, and 10 tasks for each combination of roles yielding a total of 25,000 conversations. The", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 105, 116, 507, 129 ], "spans": [ { "bbox": [ 105, 116, 507, 129 ], "score": 1.0, "content": "generated assistant roles and user roles for AI Society as well as details about the generation of Code,", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 105, 127, 337, 139 ], "spans": [ { "bbox": [ 105, 127, 337, 139 ], "score": 1.0, "content": "Math and Science datasets can be found in the Appendix.", "type": "text", "cross_page": true } ], "index": 5 } ], "index": 51, "bbox_fs": [ 105, 667, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 72, 505, 138 ], "lines": [ { "bbox": [ 106, 73, 506, 85 ], "spans": [ { "bbox": [ 106, 73, 506, 85 ], "score": 1.0, "content": "in the previous step, we then use the task specifier prompt passed to the LLM agent to make the task", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 83, 506, 96 ], "spans": [ { "bbox": [ 105, 83, 506, 96 ], "score": 1.0, "content": "more specific. The prompts for assistant role generation, user role generation, and task generation", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 94, 505, 106 ], "spans": [ { "bbox": [ 105, 94, 505, 106 ], "score": 1.0, "content": "are shown in Figure 5 (AI Society). For our AI society dataset, we generated 50 assistant roles, 50", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 106, 505, 117 ], "spans": [ { "bbox": [ 106, 106, 505, 117 ], "score": 1.0, "content": "user roles, and 10 tasks for each combination of roles yielding a total of 25,000 conversations. The", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 116, 507, 129 ], "spans": [ { "bbox": [ 105, 116, 507, 129 ], "score": 1.0, "content": "generated assistant roles and user roles for AI Society as well as details about the generation of Code,", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 127, 337, 139 ], "spans": [ { "bbox": [ 105, 127, 337, 139 ], "score": 1.0, "content": "Math and Science datasets can be found in the Appendix.", "type": "text" } ], "index": 5 } ], "index": 2.5 }, { "type": "text", "bbox": [ 108, 144, 503, 176 ], "lines": [ { "bbox": [ 106, 144, 505, 156 ], "spans": [ { "bbox": [ 106, 144, 505, 156 ], "score": 1.0, "content": "Challenges and Observations. In this section, we explore the four main challenges that we identified", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 155, 505, 167 ], "spans": [ { "bbox": [ 106, 155, 505, 167 ], "score": 1.0, "content": "during our analysis of the generated datasets. Our observations shed light on some interesting aspects", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 165, 372, 177 ], "spans": [ { "bbox": [ 106, 165, 372, 177 ], "score": 1.0, "content": "of cooperative AI and the difficulties that arise in its development.", "type": "text" } ], "index": 8 } ], "index": 7 }, { "type": "text", "bbox": [ 111, 187, 506, 354 ], "lines": [ { "bbox": [ 110, 187, 506, 200 ], "spans": [ { "bbox": [ 110, 187, 506, 200 ], "score": 1.0, "content": "• Role Flipping: One challenge we encountered was role flipping, where the assistant and user", "type": "text" } ], "index": 9 }, { "bbox": [ 119, 198, 505, 211 ], "spans": [ { "bbox": [ 119, 198, 505, 211 ], "score": 1.0, "content": "switch roles during the conversation. This issue typically arises when the assistant starts providing", "type": "text" } ], "index": 10 }, { "bbox": [ 120, 209, 505, 222 ], "spans": [ { "bbox": [ 120, 209, 505, 222 ], "score": 1.0, "content": "instructions or commands instead of following the user’s prompts, which can lead to confusion", "type": "text" } ], "index": 11 }, { "bbox": [ 119, 219, 505, 233 ], "spans": [ { "bbox": [ 119, 219, 505, 233 ], "score": 1.0, "content": "and a reversal of roles. To avoid role flipping, it is crucial for the assistant not to ask questions, as", "type": "text" } ], "index": 12 }, { "bbox": [ 120, 231, 278, 243 ], "spans": [ { "bbox": [ 120, 231, 278, 243 ], "score": 1.0, "content": "this can also contribute to the problem.", "type": "text" } ], "index": 13 }, { "bbox": [ 112, 247, 505, 258 ], "spans": [ { "bbox": [ 112, 247, 505, 258 ], "score": 1.0, "content": "• Assistant Repeats Instruction: Another challenge that we observed was the assistant", "type": "text" } ], "index": 14 }, { "bbox": [ 119, 257, 424, 271 ], "spans": [ { "bbox": [ 119, 257, 424, 271 ], "score": 1.0, "content": "simply repeating the user’s instructions without any role flipping occurring.", "type": "text" } ], "index": 15 }, { "bbox": [ 111, 272, 505, 285 ], "spans": [ { "bbox": [ 111, 272, 505, 285 ], "score": 1.0, "content": "• Flake Replies: We also observed instances where the assistant agent responds with a flake", "type": "text" } ], "index": 16 }, { "bbox": [ 120, 284, 505, 296 ], "spans": [ { "bbox": [ 120, 284, 505, 296 ], "score": 1.0, "content": "reply, often taking the form of \"I will...\". These messages do not contribute to the task at hand, as", "type": "text" } ], "index": 17 }, { "bbox": [ 120, 294, 417, 307 ], "spans": [ { "bbox": [ 120, 294, 417, 307 ], "score": 1.0, "content": "the assistant promises to take action but ultimately fails to follow through.", "type": "text" } ], "index": 18 }, { "bbox": [ 113, 310, 505, 322 ], "spans": [ { "bbox": [ 113, 310, 505, 322 ], "score": 1.0, "content": "• Infinite Loop of Messages: An interesting challenge that we encountered was when the", "type": "text" } ], "index": 19 }, { "bbox": [ 120, 321, 505, 333 ], "spans": [ { "bbox": [ 120, 321, 505, 333 ], "score": 1.0, "content": "assistant and user engage in an infinite loop of meaningless conversation, such as repeatedly", "type": "text" } ], "index": 20 }, { "bbox": [ 120, 331, 507, 344 ], "spans": [ { "bbox": [ 120, 331, 507, 344 ], "score": 1.0, "content": "thanking each other or saying goodbye without progressing the task. Interestingly, in some cases,", "type": "text" } ], "index": 21 }, { "bbox": [ 120, 342, 493, 355 ], "spans": [ { "bbox": [ 120, 342, 493, 355 ], "score": 1.0, "content": "the assistant and user are aware that they are stuck in a loop, but are unable to break out of it.", "type": "text" } ], "index": 22 } ], "index": 15.5 }, { "type": "text", "bbox": [ 107, 364, 505, 408 ], "lines": [ { "bbox": [ 105, 363, 507, 377 ], "spans": [ { "bbox": [ 105, 363, 507, 377 ], "score": 1.0, "content": "The Appendix shows examples of each of the four challenges discussed above. Overall, our observa-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 375, 505, 388 ], "spans": [ { "bbox": [ 105, 375, 505, 388 ], "score": 1.0, "content": "tions highlight the complexity of cooperative AI development and the need for continued exploration", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 386, 505, 398 ], "spans": [ { "bbox": [ 105, 386, 505, 398 ], "score": 1.0, "content": "and innovation to overcome the challenges we face. By identifying these issues, we hope to contribute", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 397, 410, 410 ], "spans": [ { "bbox": [ 105, 397, 410, 410 ], "score": 1.0, "content": "to the development of more effective and engaging cooperative AI systems.", "type": "text" } ], "index": 26 } ], "index": 24.5 }, { "type": "text", "bbox": [ 107, 413, 505, 457 ], "lines": [ { "bbox": [ 105, 413, 505, 425 ], "spans": [ { "bbox": [ 105, 413, 505, 425 ], "score": 1.0, "content": "Termination Conditions. The conversation between the assistant and user agents is designed to", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 424, 505, 436 ], "spans": [ { "bbox": [ 105, 424, 505, 436 ], "score": 1.0, "content": "follow a specific format to ensure consistent and accurate data generation. To ensure that both the", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 435, 505, 447 ], "spans": [ { "bbox": [ 105, 435, 505, 447 ], "score": 1.0, "content": "user and assistant adhere to their respective roles and responsibilities, certain conditions have been", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 446, 440, 458 ], "spans": [ { "bbox": [ 105, 446, 440, 458 ], "score": 1.0, "content": "set in place to terminate the chat if necessary. These conditions are outlined below:", "type": "text" } ], "index": 30 } ], "index": 28.5 }, { "type": "text", "bbox": [ 110, 466, 506, 617 ], "lines": [ { "bbox": [ 111, 469, 505, 479 ], "spans": [ { "bbox": [ 111, 469, 505, 479 ], "score": 1.0, "content": "User No Instruct: If the user does not instruct the assistant for 3 rounds, conversation is ended.", "type": "text" } ], "index": 31 }, { "bbox": [ 110, 483, 505, 495 ], "spans": [ { "bbox": [ 110, 483, 505, 495 ], "score": 1.0, "content": "• Assistant Instruct: If the assistant provides an instruction to the user, it indicates a role", "type": "text" } ], "index": 32 }, { "bbox": [ 120, 495, 298, 506 ], "spans": [ { "bbox": [ 120, 495, 298, 506 ], "score": 1.0, "content": "reversal, and the conversation is terminated.", "type": "text" } ], "index": 33 }, { "bbox": [ 111, 510, 505, 521 ], "spans": [ { "bbox": [ 111, 510, 505, 521 ], "score": 1.0, "content": "• End of Task Token: If the user believes that the task has been solved, they are expected to", "type": "text" } ], "index": 34 }, { "bbox": [ 119, 520, 506, 533 ], "spans": [ { "bbox": [ 119, 520, 506, 533 ], "score": 1.0, "content": "say to signify the completion of the task. Once this message is received,", "type": "text" } ], "index": 35 }, { "bbox": [ 120, 532, 244, 543 ], "spans": [ { "bbox": [ 120, 532, 244, 543 ], "score": 1.0, "content": "the conversation is terminated.", "type": "text" } ], "index": 36 }, { "bbox": [ 110, 546, 505, 558 ], "spans": [ { "bbox": [ 110, 546, 505, 558 ], "score": 1.0, "content": "Assistant&User Token Limit: Given that gpt-3.5-turbo has a limitation on the number", "type": "text" } ], "index": 37 }, { "bbox": [ 120, 558, 500, 569 ], "spans": [ { "bbox": [ 120, 558, 500, 569 ], "score": 1.0, "content": "of tokens, the conversation is terminated if either the assistant or the user reach the token limit.", "type": "text" } ], "index": 38 }, { "bbox": [ 110, 573, 505, 585 ], "spans": [ { "bbox": [ 110, 573, 505, 585 ], "score": 1.0, "content": "• Maximum Number of Messages: To keep the cost of generated chats in check, we have set", "type": "text" } ], "index": 39 }, { "bbox": [ 118, 583, 505, 596 ], "spans": [ { "bbox": [ 118, 583, 505, 596 ], "score": 1.0, "content": "a maximum limit of 40 messages. This limit guarantees a long enough conversation between", "type": "text" } ], "index": 40 }, { "bbox": [ 120, 595, 506, 607 ], "spans": [ { "bbox": [ 120, 595, 506, 607 ], "score": 1.0, "content": "the user and assistant while also ensuring that the data generated is not too costly to produce.", "type": "text" } ], "index": 41 }, { "bbox": [ 120, 606, 506, 617 ], "spans": [ { "bbox": [ 120, 606, 506, 617 ], "score": 1.0, "content": "The cost grows quadratically with the length of the conversation, making it essential to set a limit.", "type": "text" } ], "index": 42 } ], "index": 36.5 }, { "type": "title", "bbox": [ 107, 632, 181, 646 ], "lines": [ { "bbox": [ 105, 632, 182, 648 ], "spans": [ { "bbox": [ 105, 632, 182, 648 ], "score": 1.0, "content": "5 Evaluation", "type": "text" } ], "index": 43 } ], "index": 43 }, { "type": "title", "bbox": [ 107, 658, 205, 669 ], "lines": [ { "bbox": [ 105, 656, 206, 672 ], "spans": [ { "bbox": [ 105, 656, 206, 672 ], "score": 1.0, "content": "5.1 Agent Evaluation", "type": "text" } ], "index": 44 } ], "index": 44 }, { "type": "text", "bbox": [ 107, 678, 505, 722 ], "lines": [ { "bbox": [ 105, 677, 506, 691 ], "spans": [ { "bbox": [ 105, 677, 506, 691 ], "score": 1.0, "content": "In order to assess the performance of CAMEL (Cooperative Role-playing Communication), we", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 689, 505, 701 ], "spans": [ { "bbox": [ 105, 689, 505, 701 ], "score": 1.0, "content": "conduct two types of evaluations: (1) Human evaluation, and (2) GPT4 evaluation. We randomly", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 700, 507, 712 ], "spans": [ { "bbox": [ 105, 700, 507, 712 ], "score": 1.0, "content": "select 100 tasks from our AI Society dataset for evaluation and 100 tasks from our Code dataset.", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 711, 505, 723 ], "spans": [ { "bbox": [ 105, 711, 505, 723 ], "score": 1.0, "content": "Then, we employ the GPT4 model to summarize the content of the CAMEL conversation-based", "type": "text" } ], "index": 48 } ], "index": 46.5 } ], "page_idx": 7, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 742, 308, 750 ], "lines": [ { "bbox": [ 302, 741, 309, 752 ], "spans": [ { "bbox": [ 302, 741, 309, 752 ], "score": 1.0, "content": "8", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 72, 505, 138 ], "lines": [], "index": 2.5, "bbox_fs": [ 105, 73, 507, 139 ], "lines_deleted": true }, { "type": "text", "bbox": [ 108, 144, 503, 176 ], "lines": [ { "bbox": [ 106, 144, 505, 156 ], "spans": [ { "bbox": [ 106, 144, 505, 156 ], "score": 1.0, "content": "Challenges and Observations. In this section, we explore the four main challenges that we identified", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 155, 505, 167 ], "spans": [ { "bbox": [ 106, 155, 505, 167 ], "score": 1.0, "content": "during our analysis of the generated datasets. Our observations shed light on some interesting aspects", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 165, 372, 177 ], "spans": [ { "bbox": [ 106, 165, 372, 177 ], "score": 1.0, "content": "of cooperative AI and the difficulties that arise in its development.", "type": "text" } ], "index": 8 } ], "index": 7, "bbox_fs": [ 106, 144, 505, 177 ] }, { "type": "text", "bbox": [ 111, 187, 506, 354 ], "lines": [ { "bbox": [ 110, 187, 506, 200 ], "spans": [ { "bbox": [ 110, 187, 506, 200 ], "score": 1.0, "content": "• Role Flipping: One challenge we encountered was role flipping, where the assistant and user", "type": "text" } ], "index": 9 }, { "bbox": [ 119, 198, 505, 211 ], "spans": [ { "bbox": [ 119, 198, 505, 211 ], "score": 1.0, "content": "switch roles during the conversation. This issue typically arises when the assistant starts providing", "type": "text" } ], "index": 10 }, { "bbox": [ 120, 209, 505, 222 ], "spans": [ { "bbox": [ 120, 209, 505, 222 ], "score": 1.0, "content": "instructions or commands instead of following the user’s prompts, which can lead to confusion", "type": "text" } ], "index": 11 }, { "bbox": [ 119, 219, 505, 233 ], "spans": [ { "bbox": [ 119, 219, 505, 233 ], "score": 1.0, "content": "and a reversal of roles. To avoid role flipping, it is crucial for the assistant not to ask questions, as", "type": "text" } ], "index": 12 }, { "bbox": [ 120, 231, 278, 243 ], "spans": [ { "bbox": [ 120, 231, 278, 243 ], "score": 1.0, "content": "this can also contribute to the problem.", "type": "text" } ], "index": 13 }, { "bbox": [ 112, 247, 505, 258 ], "spans": [ { "bbox": [ 112, 247, 505, 258 ], "score": 1.0, "content": "• Assistant Repeats Instruction: Another challenge that we observed was the assistant", "type": "text" } ], "index": 14 }, { "bbox": [ 119, 257, 424, 271 ], "spans": [ { "bbox": [ 119, 257, 424, 271 ], "score": 1.0, "content": "simply repeating the user’s instructions without any role flipping occurring.", "type": "text" } ], "index": 15 }, { "bbox": [ 111, 272, 505, 285 ], "spans": [ { "bbox": [ 111, 272, 505, 285 ], "score": 1.0, "content": "• Flake Replies: We also observed instances where the assistant agent responds with a flake", "type": "text" } ], "index": 16 }, { "bbox": [ 120, 284, 505, 296 ], "spans": [ { "bbox": [ 120, 284, 505, 296 ], "score": 1.0, "content": "reply, often taking the form of \"I will...\". These messages do not contribute to the task at hand, as", "type": "text" } ], "index": 17 }, { "bbox": [ 120, 294, 417, 307 ], "spans": [ { "bbox": [ 120, 294, 417, 307 ], "score": 1.0, "content": "the assistant promises to take action but ultimately fails to follow through.", "type": "text" } ], "index": 18 }, { "bbox": [ 113, 310, 505, 322 ], "spans": [ { "bbox": [ 113, 310, 505, 322 ], "score": 1.0, "content": "• Infinite Loop of Messages: An interesting challenge that we encountered was when the", "type": "text" } ], "index": 19 }, { "bbox": [ 120, 321, 505, 333 ], "spans": [ { "bbox": [ 120, 321, 505, 333 ], "score": 1.0, "content": "assistant and user engage in an infinite loop of meaningless conversation, such as repeatedly", "type": "text" } ], "index": 20 }, { "bbox": [ 120, 331, 507, 344 ], "spans": [ { "bbox": [ 120, 331, 507, 344 ], "score": 1.0, "content": "thanking each other or saying goodbye without progressing the task. Interestingly, in some cases,", "type": "text" } ], "index": 21 }, { "bbox": [ 120, 342, 493, 355 ], "spans": [ { "bbox": [ 120, 342, 493, 355 ], "score": 1.0, "content": "the assistant and user are aware that they are stuck in a loop, but are unable to break out of it.", "type": "text" } ], "index": 22 } ], "index": 15.5, "bbox_fs": [ 110, 187, 507, 355 ] }, { "type": "text", "bbox": [ 107, 364, 505, 408 ], "lines": [ { "bbox": [ 105, 363, 507, 377 ], "spans": [ { "bbox": [ 105, 363, 507, 377 ], "score": 1.0, "content": "The Appendix shows examples of each of the four challenges discussed above. Overall, our observa-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 375, 505, 388 ], "spans": [ { "bbox": [ 105, 375, 505, 388 ], "score": 1.0, "content": "tions highlight the complexity of cooperative AI development and the need for continued exploration", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 386, 505, 398 ], "spans": [ { "bbox": [ 105, 386, 505, 398 ], "score": 1.0, "content": "and innovation to overcome the challenges we face. By identifying these issues, we hope to contribute", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 397, 410, 410 ], "spans": [ { "bbox": [ 105, 397, 410, 410 ], "score": 1.0, "content": "to the development of more effective and engaging cooperative AI systems.", "type": "text" } ], "index": 26 } ], "index": 24.5, "bbox_fs": [ 105, 363, 507, 410 ] }, { "type": "text", "bbox": [ 107, 413, 505, 457 ], "lines": [ { "bbox": [ 105, 413, 505, 425 ], "spans": [ { "bbox": [ 105, 413, 505, 425 ], "score": 1.0, "content": "Termination Conditions. The conversation between the assistant and user agents is designed to", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 424, 505, 436 ], "spans": [ { "bbox": [ 105, 424, 505, 436 ], "score": 1.0, "content": "follow a specific format to ensure consistent and accurate data generation. To ensure that both the", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 435, 505, 447 ], "spans": [ { "bbox": [ 105, 435, 505, 447 ], "score": 1.0, "content": "user and assistant adhere to their respective roles and responsibilities, certain conditions have been", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 446, 440, 458 ], "spans": [ { "bbox": [ 105, 446, 440, 458 ], "score": 1.0, "content": "set in place to terminate the chat if necessary. These conditions are outlined below:", "type": "text" } ], "index": 30 } ], "index": 28.5, "bbox_fs": [ 105, 413, 505, 458 ] }, { "type": "list", "bbox": [ 110, 466, 506, 617 ], "lines": [ { "bbox": [ 111, 469, 505, 479 ], "spans": [ { "bbox": [ 111, 469, 505, 479 ], "score": 1.0, "content": "User No Instruct: If the user does not instruct the assistant for 3 rounds, conversation is ended.", "type": "text" } ], "index": 31, "is_list_start_line": true }, { "bbox": [ 110, 483, 505, 495 ], "spans": [ { "bbox": [ 110, 483, 505, 495 ], "score": 1.0, "content": "• Assistant Instruct: If the assistant provides an instruction to the user, it indicates a role", "type": "text" } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 120, 495, 298, 506 ], "spans": [ { "bbox": [ 120, 495, 298, 506 ], "score": 1.0, "content": "reversal, and the conversation is terminated.", "type": "text" } ], "index": 33, "is_list_end_line": true }, { "bbox": [ 111, 510, 505, 521 ], "spans": [ { "bbox": [ 111, 510, 505, 521 ], "score": 1.0, "content": "• End of Task Token: If the user believes that the task has been solved, they are expected to", "type": "text" } ], "index": 34, "is_list_start_line": true }, { "bbox": [ 119, 520, 506, 533 ], "spans": [ { "bbox": [ 119, 520, 506, 533 ], "score": 1.0, "content": "say to signify the completion of the task. Once this message is received,", "type": "text" } ], "index": 35 }, { "bbox": [ 120, 532, 244, 543 ], "spans": [ { "bbox": [ 120, 532, 244, 543 ], "score": 1.0, "content": "the conversation is terminated.", "type": "text" } ], "index": 36, "is_list_end_line": true }, { "bbox": [ 110, 546, 505, 558 ], "spans": [ { "bbox": [ 110, 546, 505, 558 ], "score": 1.0, "content": "Assistant&User Token Limit: Given that gpt-3.5-turbo has a limitation on the number", "type": "text" } ], "index": 37, "is_list_start_line": true }, { "bbox": [ 120, 558, 500, 569 ], "spans": [ { "bbox": [ 120, 558, 500, 569 ], "score": 1.0, "content": "of tokens, the conversation is terminated if either the assistant or the user reach the token limit.", "type": "text" } ], "index": 38 }, { "bbox": [ 110, 573, 505, 585 ], "spans": [ { "bbox": [ 110, 573, 505, 585 ], "score": 1.0, "content": "• Maximum Number of Messages: To keep the cost of generated chats in check, we have set", "type": "text" } ], "index": 39, "is_list_start_line": true }, { "bbox": [ 118, 583, 505, 596 ], "spans": [ { "bbox": [ 118, 583, 505, 596 ], "score": 1.0, "content": "a maximum limit of 40 messages. This limit guarantees a long enough conversation between", "type": "text" } ], "index": 40 }, { "bbox": [ 120, 595, 506, 607 ], "spans": [ { "bbox": [ 120, 595, 506, 607 ], "score": 1.0, "content": "the user and assistant while also ensuring that the data generated is not too costly to produce.", "type": "text" } ], "index": 41 }, { "bbox": [ 120, 606, 506, 617 ], "spans": [ { "bbox": [ 120, 606, 506, 617 ], "score": 1.0, "content": "The cost grows quadratically with the length of the conversation, making it essential to set a limit.", "type": "text" } ], "index": 42 } ], "index": 36.5, "bbox_fs": [ 110, 469, 506, 617 ] }, { "type": "title", "bbox": [ 107, 632, 181, 646 ], "lines": [ { "bbox": [ 105, 632, 182, 648 ], "spans": [ { "bbox": [ 105, 632, 182, 648 ], "score": 1.0, "content": "5 Evaluation", "type": "text" } ], "index": 43 } ], "index": 43 }, { "type": "title", "bbox": [ 107, 658, 205, 669 ], "lines": [ { "bbox": [ 105, 656, 206, 672 ], "spans": [ { "bbox": [ 105, 656, 206, 672 ], "score": 1.0, "content": "5.1 Agent Evaluation", "type": "text" } ], "index": 44 } ], "index": 44 }, { "type": "text", "bbox": [ 107, 678, 505, 722 ], "lines": [ { "bbox": [ 105, 677, 506, 691 ], "spans": [ { "bbox": [ 105, 677, 506, 691 ], "score": 1.0, "content": "In order to assess the performance of CAMEL (Cooperative Role-playing Communication), we", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 689, 505, 701 ], "spans": [ { "bbox": [ 105, 689, 505, 701 ], "score": 1.0, "content": "conduct two types of evaluations: (1) Human evaluation, and (2) GPT4 evaluation. We randomly", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 700, 507, 712 ], "spans": [ { "bbox": [ 105, 700, 507, 712 ], "score": 1.0, "content": "select 100 tasks from our AI Society dataset for evaluation and 100 tasks from our Code dataset.", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 711, 505, 723 ], "spans": [ { "bbox": [ 105, 711, 505, 723 ], "score": 1.0, "content": "Then, we employ the GPT4 model to summarize the content of the CAMEL conversation-based", "type": "text" } ], "index": 48 }, { "bbox": [ 106, 73, 506, 85 ], "spans": [ { "bbox": [ 106, 73, 506, 85 ], "score": 1.0, "content": "solution, presenting a consolidated final solution. Particularly, a GPT4 is used since it possesses a", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 106, 84, 505, 95 ], "spans": [ { "bbox": [ 106, 84, 505, 95 ], "score": 1.0, "content": "larger token limit which is suitable for summarization. Summarization also makes CAMEL agents’", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 106, 95, 505, 107 ], "spans": [ { "bbox": [ 106, 95, 505, 107 ], "score": 1.0, "content": "solution undetectable by its format, allowing for a more fair comparison. Subsequently, this solution", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 106, 105, 506, 119 ], "spans": [ { "bbox": [ 106, 105, 506, 119 ], "score": 1.0, "content": "is compared with a single-shot solution generated by the gpt-3.5-turbo model for the same task.", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 106, 116, 282, 129 ], "spans": [ { "bbox": [ 106, 116, 282, 129 ], "score": 1.0, "content": "Sample tasks are provided in the Appendix.", "type": "text", "cross_page": true } ], "index": 4 } ], "index": 46.5, "bbox_fs": [ 105, 677, 507, 723 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 72, 505, 128 ], "lines": [ { "bbox": [ 106, 73, 506, 85 ], "spans": [ { "bbox": [ 106, 73, 506, 85 ], "score": 1.0, "content": "solution, presenting a consolidated final solution. Particularly, a GPT4 is used since it possesses a", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 84, 505, 95 ], "spans": [ { "bbox": [ 106, 84, 505, 95 ], "score": 1.0, "content": "larger token limit which is suitable for summarization. Summarization also makes CAMEL agents’", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 95, 505, 107 ], "spans": [ { "bbox": [ 106, 95, 505, 107 ], "score": 1.0, "content": "solution undetectable by its format, allowing for a more fair comparison. Subsequently, this solution", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 105, 506, 119 ], "spans": [ { "bbox": [ 106, 105, 506, 119 ], "score": 1.0, "content": "is compared with a single-shot solution generated by the gpt-3.5-turbo model for the same task.", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 116, 282, 129 ], "spans": [ { "bbox": [ 106, 116, 282, 129 ], "score": 1.0, "content": "Sample tasks are provided in the Appendix.", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "text", "bbox": [ 106, 132, 505, 199 ], "lines": [ { "bbox": [ 106, 132, 505, 145 ], "spans": [ { "bbox": [ 106, 132, 505, 145 ], "score": 1.0, "content": "Human Evaluation. For this evaluation, we present both the CAMEL summarized agent solution", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 144, 505, 155 ], "spans": [ { "bbox": [ 106, 144, 505, 155 ], "score": 1.0, "content": "and the gpt-3.5-turbo single-shot solution side-by-side to human participants. The identity behind", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 154, 506, 167 ], "spans": [ { "bbox": [ 105, 154, 506, 167 ], "score": 1.0, "content": "each solution is not revealed. Participants are then asked to vote on whether one solution is superior", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 164, 506, 178 ], "spans": [ { "bbox": [ 105, 164, 506, 178 ], "score": 1.0, "content": "to the other or if they are equally good. A total of 453 responses were collected during this evaluation.", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 176, 506, 189 ], "spans": [ { "bbox": [ 105, 176, 506, 189 ], "score": 1.0, "content": "Note that, human evaluation is only done for AI Society, as assessing code is generally harder for", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 188, 251, 199 ], "spans": [ { "bbox": [ 106, 188, 251, 199 ], "score": 1.0, "content": "humans (without running the code).", "type": "text" } ], "index": 10 } ], "index": 7.5 }, { "type": "text", "bbox": [ 107, 203, 505, 237 ], "lines": [ { "bbox": [ 105, 202, 505, 217 ], "spans": [ { "bbox": [ 105, 202, 505, 217 ], "score": 1.0, "content": "GPT4 Evaluation. We engage a GPT4 agent to evaluate the effectiveness of Model 1 (CAMEL Agent", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 215, 505, 227 ], "spans": [ { "bbox": [ 106, 215, 505, 227 ], "score": 1.0, "content": "solution) versus Model 2 (gpt-3.5-turbo single-shot solution) for each task. More specifically, we", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 225, 423, 238 ], "spans": [ { "bbox": [ 105, 225, 423, 238 ], "score": 1.0, "content": "prompt GPT4 to score and decide which solution of the two solutions is better.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "text", "bbox": [ 107, 241, 505, 286 ], "lines": [ { "bbox": [ 105, 241, 505, 254 ], "spans": [ { "bbox": [ 105, 241, 505, 254 ], "score": 1.0, "content": "Results. The summarized results of each evaluation are outlined in Table 1 which showcases that the", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 252, 506, 266 ], "spans": [ { "bbox": [ 105, 252, 506, 266 ], "score": 1.0, "content": "CAMEL solution outperforms gpt-3.5-turbo single-shot solution in both the human evaluation", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 263, 506, 277 ], "spans": [ { "bbox": [ 105, 263, 506, 277 ], "score": 1.0, "content": "and the GPT4 evaluation by a big margin. It is also worth noting that both human evaluation and", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 274, 253, 287 ], "spans": [ { "bbox": [ 105, 274, 253, 287 ], "score": 1.0, "content": "GPT4 evaluation are highly aligned.", "type": "text" } ], "index": 17 } ], "index": 15.5 }, { "type": "table", "bbox": [ 154, 341, 455, 390 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 307, 504, 340 ], "group_id": 0, "lines": [ { "bbox": [ 105, 305, 506, 320 ], "spans": [ { "bbox": [ 105, 305, 506, 320 ], "score": 1.0, "content": "Table 1: Agent Evaluation Results: Results of the evaluations of the CAMEL agent against", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 317, 506, 331 ], "spans": [ { "bbox": [ 105, 317, 506, 331 ], "score": 1.0, "content": "gpt-3.5-turbo using both human evaluators and GPT4 consistently show that utilizing a multi-", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 329, 466, 342 ], "spans": [ { "bbox": [ 106, 329, 323, 342 ], "score": 1.0, "content": "agent cooperative approach is more effective than gpt", "type": "text" }, { "bbox": [ 323, 330, 345, 339 ], "score": 0.26, "content": "- 3 . 5", "type": "inline_equation" }, { "bbox": [ 345, 329, 466, 342 ], "score": 1.0, "content": "-turbo’s single shot solution.", "type": "text" } ], "index": 20 } ], "index": 19 }, { "type": "table_body", "bbox": [ 154, 341, 455, 390 ], "group_id": 0, "lines": [ { "bbox": [ 154, 341, 455, 390 ], "spans": [ { "bbox": [ 154, 341, 455, 390 ], "score": 0.974, "html": "
DatasetEvaluation TypeDrawgpt-3.5-turbo WinsCAMEL Agents Win
AI SocietyHuman Evaluation13.3%10.4%76.3%
GPT4 Evaluation4.0%23.0%73.0%
CodeGPT4 Evaluation0.0%24.0%76.0%
", "type": "table", "image_path": "cb70d9df36c9f2d3f6a1f1f0485d9712ce61a16e184b4ac22c237e614eb23cbb.jpg" } ] } ], "index": 22, "virtual_lines": [ { "bbox": [ 154, 341, 455, 357.3333333333333 ], "spans": [], "index": 21 }, { "bbox": [ 154, 357.3333333333333, 455, 373.66666666666663 ], "spans": [], "index": 22 }, { "bbox": [ 154, 373.66666666666663, 455, 389.99999999999994 ], "spans": [], "index": 23 } ] } ], "index": 20.5 }, { "type": "title", "bbox": [ 107, 415, 258, 428 ], "lines": [ { "bbox": [ 106, 416, 259, 429 ], "spans": [ { "bbox": [ 106, 416, 259, 429 ], "score": 1.0, "content": "5.2 GPT4 for ChatBot Evaluation", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 437, 505, 482 ], "lines": [ { "bbox": [ 105, 437, 505, 451 ], "spans": [ { "bbox": [ 105, 437, 505, 451 ], "score": 1.0, "content": "In this section, we progressively fine-tune a LLaMA 7B model on our generated datasets. By", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 449, 506, 461 ], "spans": [ { "bbox": [ 105, 449, 506, 461 ], "score": 1.0, "content": "progressively incorporating diverse datasets like AI society, code, math, and science, we expect", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 459, 506, 473 ], "spans": [ { "bbox": [ 105, 459, 506, 473 ], "score": 1.0, "content": "fine-tuned model to demonstrate the ability to develop an increasingly sophisticated understanding of", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 470, 168, 482 ], "spans": [ { "bbox": [ 106, 470, 168, 482 ], "score": 1.0, "content": "these domains.", "type": "text" } ], "index": 28 } ], "index": 26.5 }, { "type": "text", "bbox": [ 106, 487, 505, 596 ], "lines": [ { "bbox": [ 105, 486, 506, 500 ], "spans": [ { "bbox": [ 105, 486, 506, 500 ], "score": 1.0, "content": "We initially start by training on AI society dataset, which aims to let the model learn about human", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 498, 505, 510 ], "spans": [ { "bbox": [ 106, 498, 505, 510 ], "score": 1.0, "content": "interactions and societal dynamics. As additional datasets were introduced, such as code, the model", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 508, 506, 522 ], "spans": [ { "bbox": [ 105, 508, 506, 522 ], "score": 1.0, "content": "gained knowledge of programming logic and syntax, enabling it to generate coherent and executable", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 518, 506, 534 ], "spans": [ { "bbox": [ 105, 518, 506, 534 ], "score": 1.0, "content": "code snippets. The inclusion of the math dataset further expanded the model’s capabilities, allowing", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 530, 506, 543 ], "spans": [ { "bbox": [ 105, 530, 506, 543 ], "score": 1.0, "content": "it to solve complex equations, reason about abstract concepts, and perform precise calculations.", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 540, 507, 555 ], "spans": [ { "bbox": [ 105, 540, 507, 555 ], "score": 1.0, "content": "Finally, exposure to the science dataset broadened the model’s understanding of scientific theories,", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 552, 506, 565 ], "spans": [ { "bbox": [ 105, 552, 506, 565 ], "score": 1.0, "content": "empirical observations, and experimental methods. The emergence of model capabilities is measured", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 563, 506, 576 ], "spans": [ { "bbox": [ 105, 563, 506, 576 ], "score": 1.0, "content": "by evaluating the quality of the model responses, before and after training on the new domain, on a", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 574, 506, 586 ], "spans": [ { "bbox": [ 105, 574, 506, 586 ], "score": 1.0, "content": "set of questions of varying difficulties from each domain. More precisely, the model is tested on 20", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 586, 416, 597 ], "spans": [ { "bbox": [ 106, 586, 416, 597 ], "score": 1.0, "content": "AI Society related tasks, 20 coding tasks, 20 math tasks and 60 science tasks.", "type": "text" } ], "index": 38 } ], "index": 33.5 }, { "type": "text", "bbox": [ 107, 601, 505, 722 ], "lines": [ { "bbox": [ 106, 601, 505, 614 ], "spans": [ { "bbox": [ 106, 601, 505, 614 ], "score": 1.0, "content": "Those results are highlighted in Table 2 where we see that each time we add a dataset, the model", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 612, 507, 626 ], "spans": [ { "bbox": [ 105, 612, 507, 626 ], "score": 1.0, "content": "performs better on the incorporated domain. Note that to measure the quality of the models’ responses,", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 623, 505, 636 ], "spans": [ { "bbox": [ 106, 623, 505, 636 ], "score": 1.0, "content": "we follow the evaluation from Section T, which involves prompting a GPT4 agent to score and decide", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 635, 505, 646 ], "spans": [ { "bbox": [ 106, 635, 505, 646 ], "score": 1.0, "content": "which solution is better. It is worth noting that an improvement on other domains is also observed in", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 645, 506, 657 ], "spans": [ { "bbox": [ 105, 645, 506, 657 ], "score": 1.0, "content": "some cases such as when we train on Code we improve on Science. This is because our Code dataset", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 655, 506, 668 ], "spans": [ { "bbox": [ 106, 655, 506, 668 ], "score": 1.0, "content": "contains problems that solve tasks in particular domains which include scientific domain. Similarly,", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 667, 506, 680 ], "spans": [ { "bbox": [ 106, 667, 506, 680 ], "score": 1.0, "content": "training on AI Society improves code as AI Society contains the role of a \"programmer\" and hence", "type": "text" } ], "index": 45 }, { "bbox": [ 104, 677, 506, 691 ], "spans": [ { "bbox": [ 104, 677, 506, 691 ], "score": 1.0, "content": "coding related conversations. Finally, note that the draws observed in LLaMA-7B vs AI Society in", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 688, 505, 701 ], "spans": [ { "bbox": [ 105, 688, 430, 701 ], "score": 1.0, "content": "Math reflects equally bad solutions compared to the draws observed in AI Society", "type": "text" }, { "bbox": [ 431, 689, 488, 700 ], "score": 0.53, "content": "\\mathbf { + C o d e + M a t } ]", "type": "inline_equation" }, { "bbox": [ 489, 688, 505, 701 ], "score": 1.0, "content": "h vs", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 700, 505, 713 ], "spans": [ { "bbox": [ 105, 700, 151, 713 ], "score": 1.0, "content": "AI Society", "type": "text" }, { "bbox": [ 151, 700, 228, 711 ], "score": 0.61, "content": "+ \\mathrm { C o d e } + \\mathrm { M a t h } + \\Omega", "type": "inline_equation" }, { "bbox": [ 228, 700, 505, 713 ], "score": 1.0, "content": "cience where the draws are equally good solutions. This progression", "type": "text" } ], "index": 48 }, { "bbox": [ 106, 712, 505, 723 ], "spans": [ { "bbox": [ 106, 712, 505, 723 ], "score": 1.0, "content": "from AI society to code to math to science highlights the potential of AI models to acquire a versatile", "type": "text" } ], "index": 49 } ], "index": 44 } ], "page_idx": 8, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 741, 309, 750 ], "lines": [ { "bbox": [ 302, 741, 309, 752 ], "spans": [ { "bbox": [ 302, 741, 309, 752 ], "score": 1.0, "content": "9", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 72, 505, 128 ], "lines": [], "index": 2, "bbox_fs": [ 106, 73, 506, 129 ], "lines_deleted": true }, { "type": "text", "bbox": [ 106, 132, 505, 199 ], "lines": [ { "bbox": [ 106, 132, 505, 145 ], "spans": [ { "bbox": [ 106, 132, 505, 145 ], "score": 1.0, "content": "Human Evaluation. For this evaluation, we present both the CAMEL summarized agent solution", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 144, 505, 155 ], "spans": [ { "bbox": [ 106, 144, 505, 155 ], "score": 1.0, "content": "and the gpt-3.5-turbo single-shot solution side-by-side to human participants. The identity behind", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 154, 506, 167 ], "spans": [ { "bbox": [ 105, 154, 506, 167 ], "score": 1.0, "content": "each solution is not revealed. Participants are then asked to vote on whether one solution is superior", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 164, 506, 178 ], "spans": [ { "bbox": [ 105, 164, 506, 178 ], "score": 1.0, "content": "to the other or if they are equally good. A total of 453 responses were collected during this evaluation.", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 176, 506, 189 ], "spans": [ { "bbox": [ 105, 176, 506, 189 ], "score": 1.0, "content": "Note that, human evaluation is only done for AI Society, as assessing code is generally harder for", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 188, 251, 199 ], "spans": [ { "bbox": [ 106, 188, 251, 199 ], "score": 1.0, "content": "humans (without running the code).", "type": "text" } ], "index": 10 } ], "index": 7.5, "bbox_fs": [ 105, 132, 506, 199 ] }, { "type": "text", "bbox": [ 107, 203, 505, 237 ], "lines": [ { "bbox": [ 105, 202, 505, 217 ], "spans": [ { "bbox": [ 105, 202, 505, 217 ], "score": 1.0, "content": "GPT4 Evaluation. We engage a GPT4 agent to evaluate the effectiveness of Model 1 (CAMEL Agent", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 215, 505, 227 ], "spans": [ { "bbox": [ 106, 215, 505, 227 ], "score": 1.0, "content": "solution) versus Model 2 (gpt-3.5-turbo single-shot solution) for each task. More specifically, we", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 225, 423, 238 ], "spans": [ { "bbox": [ 105, 225, 423, 238 ], "score": 1.0, "content": "prompt GPT4 to score and decide which solution of the two solutions is better.", "type": "text" } ], "index": 13 } ], "index": 12, "bbox_fs": [ 105, 202, 505, 238 ] }, { "type": "text", "bbox": [ 107, 241, 505, 286 ], "lines": [ { "bbox": [ 105, 241, 505, 254 ], "spans": [ { "bbox": [ 105, 241, 505, 254 ], "score": 1.0, "content": "Results. The summarized results of each evaluation are outlined in Table 1 which showcases that the", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 252, 506, 266 ], "spans": [ { "bbox": [ 105, 252, 506, 266 ], "score": 1.0, "content": "CAMEL solution outperforms gpt-3.5-turbo single-shot solution in both the human evaluation", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 263, 506, 277 ], "spans": [ { "bbox": [ 105, 263, 506, 277 ], "score": 1.0, "content": "and the GPT4 evaluation by a big margin. It is also worth noting that both human evaluation and", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 274, 253, 287 ], "spans": [ { "bbox": [ 105, 274, 253, 287 ], "score": 1.0, "content": "GPT4 evaluation are highly aligned.", "type": "text" } ], "index": 17 } ], "index": 15.5, "bbox_fs": [ 105, 241, 506, 287 ] }, { "type": "table", "bbox": [ 154, 341, 455, 390 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 307, 504, 340 ], "group_id": 0, "lines": [ { "bbox": [ 105, 305, 506, 320 ], "spans": [ { "bbox": [ 105, 305, 506, 320 ], "score": 1.0, "content": "Table 1: Agent Evaluation Results: Results of the evaluations of the CAMEL agent against", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 317, 506, 331 ], "spans": [ { "bbox": [ 105, 317, 506, 331 ], "score": 1.0, "content": "gpt-3.5-turbo using both human evaluators and GPT4 consistently show that utilizing a multi-", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 329, 466, 342 ], "spans": [ { "bbox": [ 106, 329, 323, 342 ], "score": 1.0, "content": "agent cooperative approach is more effective than gpt", "type": "text" }, { "bbox": [ 323, 330, 345, 339 ], "score": 0.26, "content": "- 3 . 5", "type": "inline_equation" }, { "bbox": [ 345, 329, 466, 342 ], "score": 1.0, "content": "-turbo’s single shot solution.", "type": "text" } ], "index": 20 } ], "index": 19 }, { "type": "table_body", "bbox": [ 154, 341, 455, 390 ], "group_id": 0, "lines": [ { "bbox": [ 154, 341, 455, 390 ], "spans": [ { "bbox": [ 154, 341, 455, 390 ], "score": 0.974, "html": "
DatasetEvaluation TypeDrawgpt-3.5-turbo WinsCAMEL Agents Win
AI SocietyHuman Evaluation13.3%10.4%76.3%
GPT4 Evaluation4.0%23.0%73.0%
CodeGPT4 Evaluation0.0%24.0%76.0%
", "type": "table", "image_path": "cb70d9df36c9f2d3f6a1f1f0485d9712ce61a16e184b4ac22c237e614eb23cbb.jpg" } ] } ], "index": 22, "virtual_lines": [ { "bbox": [ 154, 341, 455, 357.3333333333333 ], "spans": [], "index": 21 }, { "bbox": [ 154, 357.3333333333333, 455, 373.66666666666663 ], "spans": [], "index": 22 }, { "bbox": [ 154, 373.66666666666663, 455, 389.99999999999994 ], "spans": [], "index": 23 } ] } ], "index": 20.5 }, { "type": "title", "bbox": [ 107, 415, 258, 428 ], "lines": [ { "bbox": [ 106, 416, 259, 429 ], "spans": [ { "bbox": [ 106, 416, 259, 429 ], "score": 1.0, "content": "5.2 GPT4 for ChatBot Evaluation", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 107, 437, 505, 482 ], "lines": [ { "bbox": [ 105, 437, 505, 451 ], "spans": [ { "bbox": [ 105, 437, 505, 451 ], "score": 1.0, "content": "In this section, we progressively fine-tune a LLaMA 7B model on our generated datasets. By", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 449, 506, 461 ], "spans": [ { "bbox": [ 105, 449, 506, 461 ], "score": 1.0, "content": "progressively incorporating diverse datasets like AI society, code, math, and science, we expect", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 459, 506, 473 ], "spans": [ { "bbox": [ 105, 459, 506, 473 ], "score": 1.0, "content": "fine-tuned model to demonstrate the ability to develop an increasingly sophisticated understanding of", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 470, 168, 482 ], "spans": [ { "bbox": [ 106, 470, 168, 482 ], "score": 1.0, "content": "these domains.", "type": "text" } ], "index": 28 } ], "index": 26.5, "bbox_fs": [ 105, 437, 506, 482 ] }, { "type": "text", "bbox": [ 106, 487, 505, 596 ], "lines": [ { "bbox": [ 105, 486, 506, 500 ], "spans": [ { "bbox": [ 105, 486, 506, 500 ], "score": 1.0, "content": "We initially start by training on AI society dataset, which aims to let the model learn about human", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 498, 505, 510 ], "spans": [ { "bbox": [ 106, 498, 505, 510 ], "score": 1.0, "content": "interactions and societal dynamics. As additional datasets were introduced, such as code, the model", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 508, 506, 522 ], "spans": [ { "bbox": [ 105, 508, 506, 522 ], "score": 1.0, "content": "gained knowledge of programming logic and syntax, enabling it to generate coherent and executable", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 518, 506, 534 ], "spans": [ { "bbox": [ 105, 518, 506, 534 ], "score": 1.0, "content": "code snippets. The inclusion of the math dataset further expanded the model’s capabilities, allowing", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 530, 506, 543 ], "spans": [ { "bbox": [ 105, 530, 506, 543 ], "score": 1.0, "content": "it to solve complex equations, reason about abstract concepts, and perform precise calculations.", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 540, 507, 555 ], "spans": [ { "bbox": [ 105, 540, 507, 555 ], "score": 1.0, "content": "Finally, exposure to the science dataset broadened the model’s understanding of scientific theories,", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 552, 506, 565 ], "spans": [ { "bbox": [ 105, 552, 506, 565 ], "score": 1.0, "content": "empirical observations, and experimental methods. The emergence of model capabilities is measured", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 563, 506, 576 ], "spans": [ { "bbox": [ 105, 563, 506, 576 ], "score": 1.0, "content": "by evaluating the quality of the model responses, before and after training on the new domain, on a", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 574, 506, 586 ], "spans": [ { "bbox": [ 105, 574, 506, 586 ], "score": 1.0, "content": "set of questions of varying difficulties from each domain. More precisely, the model is tested on 20", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 586, 416, 597 ], "spans": [ { "bbox": [ 106, 586, 416, 597 ], "score": 1.0, "content": "AI Society related tasks, 20 coding tasks, 20 math tasks and 60 science tasks.", "type": "text" } ], "index": 38 } ], "index": 33.5, "bbox_fs": [ 105, 486, 507, 597 ] }, { "type": "text", "bbox": [ 107, 601, 505, 722 ], "lines": [ { "bbox": [ 106, 601, 505, 614 ], "spans": [ { "bbox": [ 106, 601, 505, 614 ], "score": 1.0, "content": "Those results are highlighted in Table 2 where we see that each time we add a dataset, the model", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 612, 507, 626 ], "spans": [ { "bbox": [ 105, 612, 507, 626 ], "score": 1.0, "content": "performs better on the incorporated domain. Note that to measure the quality of the models’ responses,", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 623, 505, 636 ], "spans": [ { "bbox": [ 106, 623, 505, 636 ], "score": 1.0, "content": "we follow the evaluation from Section T, which involves prompting a GPT4 agent to score and decide", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 635, 505, 646 ], "spans": [ { "bbox": [ 106, 635, 505, 646 ], "score": 1.0, "content": "which solution is better. It is worth noting that an improvement on other domains is also observed in", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 645, 506, 657 ], "spans": [ { "bbox": [ 105, 645, 506, 657 ], "score": 1.0, "content": "some cases such as when we train on Code we improve on Science. This is because our Code dataset", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 655, 506, 668 ], "spans": [ { "bbox": [ 106, 655, 506, 668 ], "score": 1.0, "content": "contains problems that solve tasks in particular domains which include scientific domain. Similarly,", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 667, 506, 680 ], "spans": [ { "bbox": [ 106, 667, 506, 680 ], "score": 1.0, "content": "training on AI Society improves code as AI Society contains the role of a \"programmer\" and hence", "type": "text" } ], "index": 45 }, { "bbox": [ 104, 677, 506, 691 ], "spans": [ { "bbox": [ 104, 677, 506, 691 ], "score": 1.0, "content": "coding related conversations. Finally, note that the draws observed in LLaMA-7B vs AI Society in", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 688, 505, 701 ], "spans": [ { "bbox": [ 105, 688, 430, 701 ], "score": 1.0, "content": "Math reflects equally bad solutions compared to the draws observed in AI Society", "type": "text" }, { "bbox": [ 431, 689, 488, 700 ], "score": 0.53, "content": "\\mathbf { + C o d e + M a t } ]", "type": "inline_equation" }, { "bbox": [ 489, 688, 505, 701 ], "score": 1.0, "content": "h vs", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 700, 505, 713 ], "spans": [ { "bbox": [ 105, 700, 151, 713 ], "score": 1.0, "content": "AI Society", "type": "text" }, { "bbox": [ 151, 700, 228, 711 ], "score": 0.61, "content": "+ \\mathrm { C o d e } + \\mathrm { M a t h } + \\Omega", "type": "inline_equation" }, { "bbox": [ 228, 700, 505, 713 ], "score": 1.0, "content": "cience where the draws are equally good solutions. This progression", "type": "text" } ], "index": 48 }, { "bbox": [ 106, 712, 505, 723 ], "spans": [ { "bbox": [ 106, 712, 505, 723 ], "score": 1.0, "content": "from AI society to code to math to science highlights the potential of AI models to acquire a versatile", "type": "text" } ], "index": 49 }, { "bbox": [ 106, 73, 505, 86 ], "spans": [ { "bbox": [ 106, 73, 505, 86 ], "score": 1.0, "content": "and adaptable knowledge base, paralleling the way humans gain expertise in diverse subjects. Sample", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 106, 84, 250, 96 ], "spans": [ { "bbox": [ 106, 84, 250, 96 ], "score": 1.0, "content": "tasks are provided in the Appendix.", "type": "text", "cross_page": true } ], "index": 1 } ], "index": 44, "bbox_fs": [ 104, 601, 507, 723 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 72, 504, 95 ], "lines": [ { "bbox": [ 106, 73, 505, 86 ], "spans": [ { "bbox": [ 106, 73, 505, 86 ], "score": 1.0, "content": "and adaptable knowledge base, paralleling the way humans gain expertise in diverse subjects. Sample", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 84, 250, 96 ], "spans": [ { "bbox": [ 106, 84, 250, 96 ], "score": 1.0, "content": "tasks are provided in the Appendix.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table", "bbox": [ 126, 144, 485, 311 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 99, 505, 143 ], "group_id": 0, "lines": [ { "bbox": [ 105, 100, 506, 113 ], "spans": [ { "bbox": [ 105, 100, 506, 113 ], "score": 1.0, "content": "Table 2: Emergence of Knowledge. By progressively fine-tuning LLaMA on datasets from different", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 111, 506, 122 ], "spans": [ { "bbox": [ 106, 111, 506, 122 ], "score": 1.0, "content": "domains, we observe the emergence of knowledge as the model transitions from AI society to code,", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 121, 506, 135 ], "spans": [ { "bbox": [ 105, 121, 506, 135 ], "score": 1.0, "content": "math, and science. This finding is indicated by the fact that Model 2 almost always performs better", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 132, 293, 144 ], "spans": [ { "bbox": [ 106, 132, 293, 144 ], "score": 1.0, "content": "than Model 1, especially on the added dataset.", "type": "text" } ], "index": 5 } ], "index": 3.5 }, { "type": "table_body", "bbox": [ 126, 144, 485, 311 ], "group_id": 0, "lines": [ { "bbox": [ 126, 144, 485, 311 ], "spans": [ { "bbox": [ 126, 144, 485, 311 ], "score": 0.983, "html": "
DatasetModel 1Model 2DrawModel 1Model 2
AI SocietyCodeMathScienceAI SocietyCode1Math Science
AI Society0614
Code交0020
Math√956
Science√01347
AI Society488
Code1910
Math√<<√587
Science√11940
AI Society√569
Code><>1910
Math√1316
Science√<<√3849
AI Society√3116
Code1811
Math√1055
Science√√<<<9249
AI Society√√0020
Code<<<0020
Math0020
Science√0060
", "type": "table", "image_path": "3104793f5dabcf16eda5448d63fcae3acab2fdaa3887aa76bb45863cd6fe8632.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 126, 144, 485, 199.66666666666666 ], "spans": [], "index": 6 }, { "bbox": [ 126, 199.66666666666666, 485, 255.33333333333331 ], "spans": [], "index": 7 }, { "bbox": [ 126, 255.33333333333331, 485, 311.0 ], "spans": [], "index": 8 } ] } ], "index": 5.25 }, { "type": "title", "bbox": [ 106, 328, 194, 341 ], "lines": [ { "bbox": [ 104, 328, 195, 342 ], "spans": [ { "bbox": [ 104, 328, 195, 342 ], "score": 1.0, "content": "5.3 HumanEval(+)", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "table", "bbox": [ 201, 392, 409, 452 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 359, 504, 393 ], "group_id": 1, "lines": [ { "bbox": [ 105, 358, 505, 372 ], "spans": [ { "bbox": [ 105, 358, 505, 372 ], "score": 1.0, "content": "Table 3: HumanEval(+) for Various Models. We test our CAMEL model, which is a LLaMa-7B", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 370, 506, 383 ], "spans": [ { "bbox": [ 105, 370, 506, 383 ], "score": 1.0, "content": "fine-tuned on all our datasets (AI Society, Code, Math, Science) on HumanEval and HumanEval+", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 380, 466, 394 ], "spans": [ { "bbox": [ 105, 380, 291, 394 ], "score": 1.0, "content": "benchmarks, where we show competitive pass", "type": "text" }, { "bbox": [ 291, 382, 305, 392 ], "score": 0.76, "content": "@ k", "type": "inline_equation" }, { "bbox": [ 305, 380, 466, 394 ], "score": 1.0, "content": "scores with LLaMa-7B and Vicuna-7B.", "type": "text" } ], "index": 12 } ], "index": 11 }, { "type": "table_body", "bbox": [ 201, 392, 409, 452 ], "group_id": 1, "lines": [ { "bbox": [ 201, 392, 409, 452 ], "spans": [ { "bbox": [ 201, 392, 409, 452 ], "score": 0.964, "html": "
HumanEvalHumanEval+
pass@k [%]k =1k=100k=1k=100
gpt-3.5-turbo69.494.061.789.8
LLaMA-7B10.536.51-
Vicuna-7B11.042.99.934.7
CAMEL-7B14.057.912.250.0
", "type": "table", "image_path": "7d534093ca29ed1d6ea5e069a01d88888165df6526f3f2a3f8cc972be6358fb0.jpg" } ] } ], "index": 14.5, "virtual_lines": [ { "bbox": [ 201, 392, 409, 407.0 ], "spans": [], "index": 13 }, { "bbox": [ 201, 407.0, 409, 422.0 ], "spans": [], "index": 14 }, { "bbox": [ 201, 422.0, 409, 437.0 ], "spans": [], "index": 15 }, { "bbox": [ 201, 437.0, 409, 452.0 ], "spans": [], "index": 16 } ] } ], "index": 12.75 }, { "type": "text", "bbox": [ 106, 463, 506, 529 ], "lines": [ { "bbox": [ 106, 464, 506, 475 ], "spans": [ { "bbox": [ 106, 464, 506, 475 ], "score": 1.0, "content": "To evaluate the coding task-solving capabilities of our CAMEL model, specifically the LLaMA-", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 474, 507, 487 ], "spans": [ { "bbox": [ 105, 474, 507, 487 ], "score": 1.0, "content": "7B fine-tuned on our comprehensive datasets, we rely on HumanEval [18] and HumanEval+ [69].", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 485, 506, 498 ], "spans": [ { "bbox": [ 106, 485, 506, 498 ], "score": 1.0, "content": "The results, as depicted in table 3, clearly demonstrate the remarkable performance of CAMEL. It", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 495, 505, 510 ], "spans": [ { "bbox": [ 105, 495, 505, 510 ], "score": 1.0, "content": "surpasses not only the LLaMA-7B model but also Vicuna-7B [21] by a big margin. These findings", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 507, 505, 520 ], "spans": [ { "bbox": [ 106, 507, 505, 520 ], "score": 1.0, "content": "underscore the critical role played by the generated datasets in enhancing LLaMA’s ability to tackle", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 518, 192, 531 ], "spans": [ { "bbox": [ 106, 518, 192, 531 ], "score": 1.0, "content": "coding-related tasks.", "type": "text" } ], "index": 22 } ], "index": 19.5 }, { "type": "title", "bbox": [ 107, 537, 183, 550 ], "lines": [ { "bbox": [ 104, 534, 185, 554 ], "spans": [ { "bbox": [ 104, 534, 185, 554 ], "score": 1.0, "content": "6 Conclusion", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "text", "bbox": [ 106, 554, 505, 675 ], "lines": [ { "bbox": [ 105, 554, 505, 568 ], "spans": [ { "bbox": [ 105, 554, 505, 568 ], "score": 1.0, "content": "In this paper, we explore the potential of autonomous cooperation among communicative agents", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 565, 505, 578 ], "spans": [ { "bbox": [ 105, 565, 505, 578 ], "score": 1.0, "content": "and propose a novel cooperative agent framework named role-playing . Our approach enables", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 576, 506, 590 ], "spans": [ { "bbox": [ 105, 576, 506, 590 ], "score": 1.0, "content": "communicative agents to collaborate autonomously toward completing tasks while requiring minimal", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 588, 506, 600 ], "spans": [ { "bbox": [ 106, 588, 506, 600 ], "score": 1.0, "content": "human intervention, leading to better solutions are per our thorough evaluations. Through our analysis,", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 597, 505, 610 ], "spans": [ { "bbox": [ 106, 597, 505, 610 ], "score": 1.0, "content": "we show that achieving autonomous cooperation is challenging due to issues like conversation", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 608, 506, 622 ], "spans": [ { "bbox": [ 105, 608, 506, 622 ], "score": 1.0, "content": "deviation, role flipping, and termination conditions. Our framework offers a scalable approach for", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 619, 505, 633 ], "spans": [ { "bbox": [ 105, 619, 505, 633 ], "score": 1.0, "content": "studying the cooperative behaviors and capabilities of multi-agent systems and provides strategies", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 630, 505, 644 ], "spans": [ { "bbox": [ 105, 630, 505, 644 ], "score": 1.0, "content": "for addressing these challenges. Furthermore, our open-sourced library includes implementations", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 642, 505, 654 ], "spans": [ { "bbox": [ 106, 642, 505, 654 ], "score": 1.0, "content": "of various agents, data generation pipelines, data analysis tools, and collected datasets, to support", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 653, 504, 664 ], "spans": [ { "bbox": [ 106, 653, 504, 664 ], "score": 1.0, "content": "research on communicative agents and beyond. Our contributions offer valuable insights into the", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 663, 435, 677 ], "spans": [ { "bbox": [ 105, 663, 435, 677 ], "score": 1.0, "content": "future of large language artificial intelligence models and cooperative AI systems.", "type": "text" } ], "index": 34 } ], "index": 29 }, { "type": "title", "bbox": [ 107, 683, 225, 696 ], "lines": [ { "bbox": [ 104, 681, 226, 699 ], "spans": [ { "bbox": [ 104, 681, 226, 699 ], "score": 1.0, "content": "7 Acknowledgements", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 107, 700, 503, 722 ], "lines": [ { "bbox": [ 105, 699, 505, 712 ], "spans": [ { "bbox": [ 105, 699, 505, 712 ], "score": 1.0, "content": "This work was supported by SDAIA-KAUST Center of Excellence in Data Science and Artificial", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 711, 247, 723 ], "spans": [ { "bbox": [ 106, 711, 247, 723 ], "score": 1.0, "content": "Intelligence (SDAIA-KAUST AI).", "type": "text" } ], "index": 37 } ], "index": 36.5 } ], "page_idx": 9, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 313, 754 ], "spans": [ { "bbox": [ 299, 740, 313, 754 ], "score": 1.0, "content": "10", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 72, 504, 95 ], "lines": [], "index": 0.5, "bbox_fs": [ 106, 73, 505, 96 ], "lines_deleted": true }, { "type": "table", "bbox": [ 126, 144, 485, 311 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 99, 505, 143 ], "group_id": 0, "lines": [ { "bbox": [ 105, 100, 506, 113 ], "spans": [ { "bbox": [ 105, 100, 506, 113 ], "score": 1.0, "content": "Table 2: Emergence of Knowledge. By progressively fine-tuning LLaMA on datasets from different", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 111, 506, 122 ], "spans": [ { "bbox": [ 106, 111, 506, 122 ], "score": 1.0, "content": "domains, we observe the emergence of knowledge as the model transitions from AI society to code,", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 121, 506, 135 ], "spans": [ { "bbox": [ 105, 121, 506, 135 ], "score": 1.0, "content": "math, and science. This finding is indicated by the fact that Model 2 almost always performs better", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 132, 293, 144 ], "spans": [ { "bbox": [ 106, 132, 293, 144 ], "score": 1.0, "content": "than Model 1, especially on the added dataset.", "type": "text" } ], "index": 5 } ], "index": 3.5 }, { "type": "table_body", "bbox": [ 126, 144, 485, 311 ], "group_id": 0, "lines": [ { "bbox": [ 126, 144, 485, 311 ], "spans": [ { "bbox": [ 126, 144, 485, 311 ], "score": 0.983, "html": "
DatasetModel 1Model 2DrawModel 1Model 2
AI SocietyCodeMathScienceAI SocietyCode1Math Science
AI Society0614
Code交0020
Math√956
Science√01347
AI Society488
Code1910
Math√<<√587
Science√11940
AI Society√569
Code><>1910
Math√1316
Science√<<√3849
AI Society√3116
Code1811
Math√1055
Science√√<<<9249
AI Society√√0020
Code<<<0020
Math0020
Science√0060
", "type": "table", "image_path": "3104793f5dabcf16eda5448d63fcae3acab2fdaa3887aa76bb45863cd6fe8632.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 126, 144, 485, 199.66666666666666 ], "spans": [], "index": 6 }, { "bbox": [ 126, 199.66666666666666, 485, 255.33333333333331 ], "spans": [], "index": 7 }, { "bbox": [ 126, 255.33333333333331, 485, 311.0 ], "spans": [], "index": 8 } ] } ], "index": 5.25 }, { "type": "title", "bbox": [ 106, 328, 194, 341 ], "lines": [ { "bbox": [ 104, 328, 195, 342 ], "spans": [ { "bbox": [ 104, 328, 195, 342 ], "score": 1.0, "content": "5.3 HumanEval(+)", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "table", "bbox": [ 201, 392, 409, 452 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 359, 504, 393 ], "group_id": 1, "lines": [ { "bbox": [ 105, 358, 505, 372 ], "spans": [ { "bbox": [ 105, 358, 505, 372 ], "score": 1.0, "content": "Table 3: HumanEval(+) for Various Models. We test our CAMEL model, which is a LLaMa-7B", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 370, 506, 383 ], "spans": [ { "bbox": [ 105, 370, 506, 383 ], "score": 1.0, "content": "fine-tuned on all our datasets (AI Society, Code, Math, Science) on HumanEval and HumanEval+", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 380, 466, 394 ], "spans": [ { "bbox": [ 105, 380, 291, 394 ], "score": 1.0, "content": "benchmarks, where we show competitive pass", "type": "text" }, { "bbox": [ 291, 382, 305, 392 ], "score": 0.76, "content": "@ k", "type": "inline_equation" }, { "bbox": [ 305, 380, 466, 394 ], "score": 1.0, "content": "scores with LLaMa-7B and Vicuna-7B.", "type": "text" } ], "index": 12 } ], "index": 11 }, { "type": "table_body", "bbox": [ 201, 392, 409, 452 ], "group_id": 1, "lines": [ { "bbox": [ 201, 392, 409, 452 ], "spans": [ { "bbox": [ 201, 392, 409, 452 ], "score": 0.964, "html": "
HumanEvalHumanEval+
pass@k [%]k =1k=100k=1k=100
gpt-3.5-turbo69.494.061.789.8
LLaMA-7B10.536.51-
Vicuna-7B11.042.99.934.7
CAMEL-7B14.057.912.250.0
", "type": "table", "image_path": "7d534093ca29ed1d6ea5e069a01d88888165df6526f3f2a3f8cc972be6358fb0.jpg" } ] } ], "index": 14.5, "virtual_lines": [ { "bbox": [ 201, 392, 409, 407.0 ], "spans": [], "index": 13 }, { "bbox": [ 201, 407.0, 409, 422.0 ], "spans": [], "index": 14 }, { "bbox": [ 201, 422.0, 409, 437.0 ], "spans": [], "index": 15 }, { "bbox": [ 201, 437.0, 409, 452.0 ], "spans": [], "index": 16 } ] } ], "index": 12.75 }, { "type": "text", "bbox": [ 106, 463, 506, 529 ], "lines": [ { "bbox": [ 106, 464, 506, 475 ], "spans": [ { "bbox": [ 106, 464, 506, 475 ], "score": 1.0, "content": "To evaluate the coding task-solving capabilities of our CAMEL model, specifically the LLaMA-", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 474, 507, 487 ], "spans": [ { "bbox": [ 105, 474, 507, 487 ], "score": 1.0, "content": "7B fine-tuned on our comprehensive datasets, we rely on HumanEval [18] and HumanEval+ [69].", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 485, 506, 498 ], "spans": [ { "bbox": [ 106, 485, 506, 498 ], "score": 1.0, "content": "The results, as depicted in table 3, clearly demonstrate the remarkable performance of CAMEL. It", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 495, 505, 510 ], "spans": [ { "bbox": [ 105, 495, 505, 510 ], "score": 1.0, "content": "surpasses not only the LLaMA-7B model but also Vicuna-7B [21] by a big margin. These findings", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 507, 505, 520 ], "spans": [ { "bbox": [ 106, 507, 505, 520 ], "score": 1.0, "content": "underscore the critical role played by the generated datasets in enhancing LLaMA’s ability to tackle", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 518, 192, 531 ], "spans": [ { "bbox": [ 106, 518, 192, 531 ], "score": 1.0, "content": "coding-related tasks.", "type": "text" } ], "index": 22 } ], "index": 19.5, "bbox_fs": [ 105, 464, 507, 531 ] }, { "type": "title", "bbox": [ 107, 537, 183, 550 ], "lines": [ { "bbox": [ 104, 534, 185, 554 ], "spans": [ { "bbox": [ 104, 534, 185, 554 ], "score": 1.0, "content": "6 Conclusion", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "text", "bbox": [ 106, 554, 505, 675 ], "lines": [ { "bbox": [ 105, 554, 505, 568 ], "spans": [ { "bbox": [ 105, 554, 505, 568 ], "score": 1.0, "content": "In this paper, we explore the potential of autonomous cooperation among communicative agents", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 565, 505, 578 ], "spans": [ { "bbox": [ 105, 565, 505, 578 ], "score": 1.0, "content": "and propose a novel cooperative agent framework named role-playing . Our approach enables", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 576, 506, 590 ], "spans": [ { "bbox": [ 105, 576, 506, 590 ], "score": 1.0, "content": "communicative agents to collaborate autonomously toward completing tasks while requiring minimal", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 588, 506, 600 ], "spans": [ { "bbox": [ 106, 588, 506, 600 ], "score": 1.0, "content": "human intervention, leading to better solutions are per our thorough evaluations. Through our analysis,", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 597, 505, 610 ], "spans": [ { "bbox": [ 106, 597, 505, 610 ], "score": 1.0, "content": "we show that achieving autonomous cooperation is challenging due to issues like conversation", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 608, 506, 622 ], "spans": [ { "bbox": [ 105, 608, 506, 622 ], "score": 1.0, "content": "deviation, role flipping, and termination conditions. Our framework offers a scalable approach for", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 619, 505, 633 ], "spans": [ { "bbox": [ 105, 619, 505, 633 ], "score": 1.0, "content": "studying the cooperative behaviors and capabilities of multi-agent systems and provides strategies", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 630, 505, 644 ], "spans": [ { "bbox": [ 105, 630, 505, 644 ], "score": 1.0, "content": "for addressing these challenges. Furthermore, our open-sourced library includes implementations", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 642, 505, 654 ], "spans": [ { "bbox": [ 106, 642, 505, 654 ], "score": 1.0, "content": "of various agents, data generation pipelines, data analysis tools, and collected datasets, to support", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 653, 504, 664 ], "spans": [ { "bbox": [ 106, 653, 504, 664 ], "score": 1.0, "content": "research on communicative agents and beyond. Our contributions offer valuable insights into the", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 663, 435, 677 ], "spans": [ { "bbox": [ 105, 663, 435, 677 ], "score": 1.0, "content": "future of large language artificial intelligence models and cooperative AI systems.", "type": "text" } ], "index": 34 } ], "index": 29, "bbox_fs": [ 105, 554, 506, 677 ] }, { "type": "title", "bbox": [ 107, 683, 225, 696 ], "lines": [ { "bbox": [ 104, 681, 226, 699 ], "spans": [ { "bbox": [ 104, 681, 226, 699 ], "score": 1.0, "content": "7 Acknowledgements", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 107, 700, 503, 722 ], "lines": [ { "bbox": [ 105, 699, 505, 712 ], "spans": [ { "bbox": [ 105, 699, 505, 712 ], "score": 1.0, "content": "This work was supported by SDAIA-KAUST Center of Excellence in Data Science and Artificial", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 711, 247, 723 ], "spans": [ { "bbox": [ 106, 711, 247, 723 ], "score": 1.0, "content": "Intelligence (SDAIA-KAUST AI).", "type": "text" } ], "index": 37 } ], "index": 36.5, "bbox_fs": [ 105, 699, 505, 723 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 107, 72, 163, 84 ], "lines": [ { "bbox": [ 106, 70, 165, 86 ], "spans": [ { "bbox": [ 106, 70, 165, 86 ], "score": 1.0, "content": "References", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 115, 90, 506, 140 ], "lines": [ { "bbox": [ 115, 90, 505, 101 ], "spans": [ { "bbox": [ 115, 90, 505, 101 ], "score": 1.0, "content": "[1] Josh Abramson, Arun Ahuja, Iain Barr, Arthur Brussee, Federico Carnevale, Mary Cassin, Rachita", "type": "text" } ], "index": 1 }, { "bbox": [ 129, 98, 506, 112 ], "spans": [ { "bbox": [ 129, 98, 506, 112 ], "score": 1.0, "content": "Chhaparia, Stephen Clark, Bogdan Damoc, Andrew Dudzik, Petko Georgiev, Aurelia Guy, Tim Harley,", "type": "text" } ], "index": 2 }, { "bbox": [ 129, 108, 505, 122 ], "spans": [ { "bbox": [ 129, 108, 505, 122 ], "score": 1.0, "content": "Felix Hill, Alden Hung, Zachary Kenton, Jessica Landon, Timothy Lillicrap, Kory Mathewson, Sonaˇ", "type": "text" } ], "index": 3 }, { "bbox": [ 129, 118, 507, 131 ], "spans": [ { "bbox": [ 129, 118, 507, 131 ], "score": 1.0, "content": "Mokrá, Alistair Muldal, Adam Santoro, Nikolay Savinov, Vikrant Varma, Greg Wayne, Duncan Williams,", "type": "text" } ], "index": 4 }, { "bbox": [ 129, 129, 427, 141 ], "spans": [ { "bbox": [ 129, 129, 427, 141 ], "score": 1.0, "content": "Nathaniel Wong, Chen Yan, and Rui Zhu. Imitating interactive intelligence, 2020.", "type": "text" } ], "index": 5 } ], "index": 3 }, { "type": "text", "bbox": [ 116, 147, 505, 228 ], "lines": [ { "bbox": [ 115, 148, 505, 159 ], "spans": [ { "bbox": [ 115, 148, 505, 159 ], "score": 1.0, "content": "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea", "type": "text" } ], "index": 6 }, { "bbox": [ 129, 156, 507, 170 ], "spans": [ { "bbox": [ 129, 156, 507, 170 ], "score": 1.0, "content": "Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu,", "type": "text" } ], "index": 7 }, { "bbox": [ 129, 166, 506, 180 ], "spans": [ { "bbox": [ 129, 166, 506, 180 ], "score": 1.0, "content": "Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth,", "type": "text" } ], "index": 8 }, { "bbox": [ 129, 177, 506, 190 ], "spans": [ { "bbox": [ 129, 177, 506, 190 ], "score": 1.0, "content": "Nikhil J Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao", "type": "text" } ], "index": 9 }, { "bbox": [ 129, 186, 506, 200 ], "spans": [ { "bbox": [ 129, 186, 506, 200 ], "score": 1.0, "content": "Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse,", "type": "text" } ], "index": 10 }, { "bbox": [ 130, 198, 505, 209 ], "spans": [ { "bbox": [ 130, 198, 505, 209 ], "score": 1.0, "content": "Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei", "type": "text" } ], "index": 11 }, { "bbox": [ 129, 205, 506, 221 ], "spans": [ { "bbox": [ 129, 205, 506, 221 ], "score": 1.0, "content": "Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do as i can, not as i say: Grounding", "type": "text" } ], "index": 12 }, { "bbox": [ 129, 218, 270, 228 ], "spans": [ { "bbox": [ 129, 218, 270, 228 ], "score": 1.0, "content": "language in robotic affordances, 2022.", "type": "text" } ], "index": 13 } ], "index": 9.5 }, { "type": "text", "bbox": [ 115, 235, 338, 246 ], "lines": [ { "bbox": [ 115, 235, 339, 247 ], "spans": [ { "bbox": [ 115, 235, 339, 247 ], "score": 1.0, "content": "[3] Jacob Andreas. Language models as agent models, 2022.", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 112, 253, 504, 274 ], "lines": [ { "bbox": [ 115, 253, 505, 266 ], "spans": [ { "bbox": [ 115, 253, 505, 266 ], "score": 1.0, "content": "[4] Jacob Andreas and Dan Klein. Alignment-based compositional semantics for instruction following. arXiv", "type": "text" } ], "index": 15 }, { "bbox": [ 129, 264, 255, 275 ], "spans": [ { "bbox": [ 129, 264, 255, 275 ], "score": 1.0, "content": "preprint arXiv:1508.06491, 2015.", "type": "text" } ], "index": 16 } ], "index": 15.5 }, { "type": "text", "bbox": [ 116, 282, 325, 293 ], "lines": [ { "bbox": [ 115, 281, 327, 294 ], "spans": [ { "bbox": [ 115, 281, 327, 294 ], "score": 1.0, "content": "[5] Anthropic. Introducing claude. Anthropic Blog, 2023.", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 115, 300, 328, 311 ], "lines": [ { "bbox": [ 115, 300, 329, 312 ], "spans": [ { "bbox": [ 115, 300, 329, 312 ], "score": 1.0, "content": "[6] Isaac Asimov. I. Robot. Narkaling Productions., 1940.", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 112, 318, 504, 339 ], "lines": [ { "bbox": [ 114, 318, 504, 330 ], "spans": [ { "bbox": [ 114, 318, 504, 330 ], "score": 1.0, "content": "[7] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? Advances in neural information", "type": "text" } ], "index": 19 }, { "bbox": [ 129, 330, 240, 339 ], "spans": [ { "bbox": [ 129, 330, 240, 339 ], "score": 1.0, "content": "processing systems, 27, 2014.", "type": "text" } ], "index": 20 } ], "index": 19.5 }, { "type": "text", "bbox": [ 118, 347, 504, 397 ], "lines": [ { "bbox": [ 116, 346, 506, 358 ], "spans": [ { "bbox": [ 116, 346, 506, 358 ], "score": 1.0, "content": "[8] Sanghwan Bae, Donghyun Kwak, Sungdong Kim, Donghoon Ham, Soyoung Kang, Sang-Woo Lee, and", "type": "text" } ], "index": 21 }, { "bbox": [ 130, 356, 505, 370 ], "spans": [ { "bbox": [ 130, 356, 505, 370 ], "score": 1.0, "content": "Woomyoung Park. Building a role specified open-domain dialogue system leveraging large-scale language", "type": "text" } ], "index": 22 }, { "bbox": [ 130, 366, 505, 379 ], "spans": [ { "bbox": [ 130, 366, 505, 379 ], "score": 1.0, "content": "models. In Proceedings of the 2022 Conference of the North American Chapter of the Association for", "type": "text" } ], "index": 23 }, { "bbox": [ 130, 376, 506, 389 ], "spans": [ { "bbox": [ 130, 376, 506, 389 ], "score": 1.0, "content": "Computational Linguistics: Human Language Technologies, pages 2128–2150, Seattle, United States,", "type": "text" } ], "index": 24 }, { "bbox": [ 130, 386, 327, 399 ], "spans": [ { "bbox": [ 130, 386, 327, 399 ], "score": 1.0, "content": "July 2022. Association for Computational Linguistics.", "type": "text" } ], "index": 25 } ], "index": 23 }, { "type": "text", "bbox": [ 114, 405, 504, 436 ], "lines": [ { "bbox": [ 114, 404, 506, 417 ], "spans": [ { "bbox": [ 114, 404, 506, 417 ], "score": 1.0, "content": "[9] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain,", "type": "text" } ], "index": 26 }, { "bbox": [ 131, 415, 505, 426 ], "spans": [ { "bbox": [ 131, 415, 505, 426 ], "score": 1.0, "content": "Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with", "type": "text" } ], "index": 27 }, { "bbox": [ 130, 425, 445, 436 ], "spans": [ { "bbox": [ 130, 425, 445, 436 ], "score": 1.0, "content": "reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022.", "type": "text" } ], "index": 28 } ], "index": 27 }, { "type": "text", "bbox": [ 111, 443, 504, 474 ], "lines": [ { "bbox": [ 109, 442, 505, 455 ], "spans": [ { "bbox": [ 109, 442, 505, 455 ], "score": 1.0, "content": "[10] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna", "type": "text" } ], "index": 29 }, { "bbox": [ 131, 453, 505, 464 ], "spans": [ { "bbox": [ 131, 453, 505, 464 ], "score": 1.0, "content": "Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from", "type": "text" } ], "index": 30 }, { "bbox": [ 130, 463, 323, 474 ], "spans": [ { "bbox": [ 130, 463, 323, 474 ], "score": 1.0, "content": "ai feedback. arXiv preprint arXiv:2212.08073, 2022.", "type": "text" } ], "index": 31 } ], "index": 30 }, { "type": "text", "bbox": [ 110, 481, 504, 512 ], "lines": [ { "bbox": [ 111, 481, 505, 493 ], "spans": [ { "bbox": [ 111, 481, 505, 493 ], "score": 1.0, "content": "[11] Nolan Bard, Jakob N Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H Francis Song, Emilio", "type": "text" } ], "index": 32 }, { "bbox": [ 129, 490, 505, 503 ], "spans": [ { "bbox": [ 129, 490, 505, 503 ], "score": 1.0, "content": "Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, et al. The hanabi challenge: A new", "type": "text" } ], "index": 33 }, { "bbox": [ 131, 502, 367, 512 ], "spans": [ { "bbox": [ 131, 502, 367, 512 ], "score": 1.0, "content": "frontier for ai research. Artificial Intelligence, 280:103216, 2020.", "type": "text" } ], "index": 34 } ], "index": 33 }, { "type": "text", "bbox": [ 110, 519, 504, 550 ], "lines": [ { "bbox": [ 110, 519, 505, 531 ], "spans": [ { "bbox": [ 110, 519, 505, 531 ], "score": 1.0, "content": "[12] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind", "type": "text" } ], "index": 35 }, { "bbox": [ 131, 529, 506, 540 ], "spans": [ { "bbox": [ 131, 529, 506, 540 ], "score": 1.0, "content": "Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.", "type": "text" } ], "index": 36 }, { "bbox": [ 129, 539, 398, 551 ], "spans": [ { "bbox": [ 129, 539, 398, 551 ], "score": 1.0, "content": "Advances in neural information processing systems, 33:1877–1901, 2020.", "type": "text" } ], "index": 37 } ], "index": 36 }, { "type": "text", "bbox": [ 110, 557, 504, 588 ], "lines": [ { "bbox": [ 110, 558, 506, 569 ], "spans": [ { "bbox": [ 110, 558, 506, 569 ], "score": 1.0, "content": "[13] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar,", "type": "text" } ], "index": 38 }, { "bbox": [ 129, 567, 505, 580 ], "spans": [ { "bbox": [ 129, 567, 505, 580 ], "score": 1.0, "content": "Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early", "type": "text" } ], "index": 39 }, { "bbox": [ 130, 578, 366, 588 ], "spans": [ { "bbox": [ 130, 578, 366, 588 ], "score": 1.0, "content": "experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023.", "type": "text" } ], "index": 40 } ], "index": 39 }, { "type": "text", "bbox": [ 109, 595, 504, 627 ], "lines": [ { "bbox": [ 109, 594, 506, 609 ], "spans": [ { "bbox": [ 109, 594, 506, 609 ], "score": 1.0, "content": "[14] N Carlini, F Tramer, E Wallace, M Jagielski, A Herbert-Voss, K Lee, A Roberts, T Brown, D Song,", "type": "text" } ], "index": 41 }, { "bbox": [ 129, 605, 506, 619 ], "spans": [ { "bbox": [ 129, 605, 506, 619 ], "score": 1.0, "content": "Ú Erlingsson, et al. Extracting training data from large language models. arxiv. Preprint posted online", "type": "text" } ], "index": 42 }, { "bbox": [ 130, 617, 208, 627 ], "spans": [ { "bbox": [ 130, 617, 208, 627 ], "score": 1.0, "content": "December, 14, 2020.", "type": "text" } ], "index": 43 } ], "index": 42 }, { "type": "text", "bbox": [ 109, 635, 504, 665 ], "lines": [ { "bbox": [ 110, 635, 505, 646 ], "spans": [ { "bbox": [ 110, 635, 505, 646 ], "score": 1.0, "content": "[15] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramèr, Borja", "type": "text" } ], "index": 44 }, { "bbox": [ 130, 644, 505, 657 ], "spans": [ { "bbox": [ 130, 644, 505, 657 ], "score": 1.0, "content": "Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. arXiv preprint", "type": "text" } ], "index": 45 }, { "bbox": [ 129, 655, 223, 665 ], "spans": [ { "bbox": [ 129, 655, 223, 665 ], "score": 1.0, "content": "arXiv:2301.13188, 2023.", "type": "text" } ], "index": 46 } ], "index": 45 }, { "type": "text", "bbox": [ 111, 673, 257, 684 ], "lines": [ { "bbox": [ 110, 673, 258, 685 ], "spans": [ { "bbox": [ 110, 673, 258, 685 ], "score": 1.0, "content": "[16] Harrison Chase. Langchain. 2022.", "type": "text" } ], "index": 47 } ], "index": 47 }, { "type": "text", "bbox": [ 109, 691, 504, 722 ], "lines": [ { "bbox": [ 110, 691, 506, 704 ], "spans": [ { "bbox": [ 110, 691, 506, 704 ], "score": 1.0, "content": "[17] Defang Chen, Jian-Ping Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Cross-layer", "type": "text" } ], "index": 48 }, { "bbox": [ 129, 701, 506, 713 ], "spans": [ { "bbox": [ 129, 701, 506, 713 ], "score": 1.0, "content": "distillation with semantic calibration. In Proceedings of the AAAI Conference on Artificial Intelligence,", "type": "text" } ], "index": 49 }, { "bbox": [ 131, 711, 263, 723 ], "spans": [ { "bbox": [ 131, 711, 263, 723 ], "score": 1.0, "content": "volume 35, pages 7028–7036, 2021.", "type": "text" } ], "index": 50 } ], "index": 49 } ], "page_idx": 10, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 310, 750 ], "lines": [ { "bbox": [ 299, 740, 312, 754 ], "spans": [ { "bbox": [ 299, 740, 312, 754 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 107, 72, 163, 84 ], "lines": [ { "bbox": [ 106, 70, 165, 86 ], "spans": [ { "bbox": [ 106, 70, 165, 86 ], "score": 1.0, "content": "References", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 115, 90, 506, 140 ], "lines": [ { "bbox": [ 115, 90, 505, 101 ], "spans": [ { "bbox": [ 115, 90, 505, 101 ], "score": 1.0, "content": "[1] Josh Abramson, Arun Ahuja, Iain Barr, Arthur Brussee, Federico Carnevale, Mary Cassin, Rachita", "type": "text" } ], "index": 1 }, { "bbox": [ 129, 98, 506, 112 ], "spans": [ { "bbox": [ 129, 98, 506, 112 ], "score": 1.0, "content": "Chhaparia, Stephen Clark, Bogdan Damoc, Andrew Dudzik, Petko Georgiev, Aurelia Guy, Tim Harley,", "type": "text" } ], "index": 2 }, { "bbox": [ 129, 108, 505, 122 ], "spans": [ { "bbox": [ 129, 108, 505, 122 ], "score": 1.0, "content": "Felix Hill, Alden Hung, Zachary Kenton, Jessica Landon, Timothy Lillicrap, Kory Mathewson, Sonaˇ", "type": "text" } ], "index": 3 }, { "bbox": [ 129, 118, 507, 131 ], "spans": [ { "bbox": [ 129, 118, 507, 131 ], "score": 1.0, "content": "Mokrá, Alistair Muldal, Adam Santoro, Nikolay Savinov, Vikrant Varma, Greg Wayne, Duncan Williams,", "type": "text" } ], "index": 4 }, { "bbox": [ 129, 129, 427, 141 ], "spans": [ { "bbox": [ 129, 129, 427, 141 ], "score": 1.0, "content": "Nathaniel Wong, Chen Yan, and Rui Zhu. Imitating interactive intelligence, 2020.", "type": "text" } ], "index": 5 } ], "index": 3, "bbox_fs": [ 115, 90, 507, 141 ] }, { "type": "text", "bbox": [ 116, 147, 505, 228 ], "lines": [ { "bbox": [ 115, 148, 505, 159 ], "spans": [ { "bbox": [ 115, 148, 505, 159 ], "score": 1.0, "content": "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea", "type": "text" } ], "index": 6 }, { "bbox": [ 129, 156, 507, 170 ], "spans": [ { "bbox": [ 129, 156, 507, 170 ], "score": 1.0, "content": "Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu,", "type": "text" } ], "index": 7 }, { "bbox": [ 129, 166, 506, 180 ], "spans": [ { "bbox": [ 129, 166, 506, 180 ], "score": 1.0, "content": "Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth,", "type": "text" } ], "index": 8 }, { "bbox": [ 129, 177, 506, 190 ], "spans": [ { "bbox": [ 129, 177, 506, 190 ], "score": 1.0, "content": "Nikhil J Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao", "type": "text" } ], "index": 9 }, { "bbox": [ 129, 186, 506, 200 ], "spans": [ { "bbox": [ 129, 186, 506, 200 ], "score": 1.0, "content": "Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse,", "type": "text" } ], "index": 10 }, { "bbox": [ 130, 198, 505, 209 ], "spans": [ { "bbox": [ 130, 198, 505, 209 ], "score": 1.0, "content": "Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei", "type": "text" } ], "index": 11 }, { "bbox": [ 129, 205, 506, 221 ], "spans": [ { "bbox": [ 129, 205, 506, 221 ], "score": 1.0, "content": "Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do as i can, not as i say: Grounding", "type": "text" } ], "index": 12 }, { "bbox": [ 129, 218, 270, 228 ], "spans": [ { "bbox": [ 129, 218, 270, 228 ], "score": 1.0, "content": "language in robotic affordances, 2022.", "type": "text" } ], "index": 13 } ], "index": 9.5, "bbox_fs": [ 115, 148, 507, 228 ] }, { "type": "text", "bbox": [ 115, 235, 338, 246 ], "lines": [ { "bbox": [ 115, 235, 339, 247 ], "spans": [ { "bbox": [ 115, 235, 339, 247 ], "score": 1.0, "content": "[3] Jacob Andreas. Language models as agent models, 2022.", "type": "text" } ], "index": 14 } ], "index": 14, "bbox_fs": [ 115, 235, 339, 247 ] }, { "type": "text", "bbox": [ 112, 253, 504, 274 ], "lines": [ { "bbox": [ 115, 253, 505, 266 ], "spans": [ { "bbox": [ 115, 253, 505, 266 ], "score": 1.0, "content": "[4] Jacob Andreas and Dan Klein. Alignment-based compositional semantics for instruction following. arXiv", "type": "text" } ], "index": 15 }, { "bbox": [ 129, 264, 255, 275 ], "spans": [ { "bbox": [ 129, 264, 255, 275 ], "score": 1.0, "content": "preprint arXiv:1508.06491, 2015.", "type": "text" } ], "index": 16 } ], "index": 15.5, "bbox_fs": [ 115, 253, 505, 275 ] }, { "type": "text", "bbox": [ 116, 282, 325, 293 ], "lines": [ { "bbox": [ 115, 281, 327, 294 ], "spans": [ { "bbox": [ 115, 281, 327, 294 ], "score": 1.0, "content": "[5] Anthropic. Introducing claude. Anthropic Blog, 2023.", "type": "text" } ], "index": 17 } ], "index": 17, "bbox_fs": [ 115, 281, 327, 294 ] }, { "type": "text", "bbox": [ 115, 300, 328, 311 ], "lines": [ { "bbox": [ 115, 300, 329, 312 ], "spans": [ { "bbox": [ 115, 300, 329, 312 ], "score": 1.0, "content": "[6] Isaac Asimov. I. Robot. Narkaling Productions., 1940.", "type": "text" } ], "index": 18 } ], "index": 18, "bbox_fs": [ 115, 300, 329, 312 ] }, { "type": "text", "bbox": [ 112, 318, 504, 339 ], "lines": [ { "bbox": [ 114, 318, 504, 330 ], "spans": [ { "bbox": [ 114, 318, 504, 330 ], "score": 1.0, "content": "[7] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? Advances in neural information", "type": "text" } ], "index": 19 }, { "bbox": [ 129, 330, 240, 339 ], "spans": [ { "bbox": [ 129, 330, 240, 339 ], "score": 1.0, "content": "processing systems, 27, 2014.", "type": "text" } ], "index": 20 } ], "index": 19.5, "bbox_fs": [ 114, 318, 504, 339 ] }, { "type": "text", "bbox": [ 118, 347, 504, 397 ], "lines": [ { "bbox": [ 116, 346, 506, 358 ], "spans": [ { "bbox": [ 116, 346, 506, 358 ], "score": 1.0, "content": "[8] Sanghwan Bae, Donghyun Kwak, Sungdong Kim, Donghoon Ham, Soyoung Kang, Sang-Woo Lee, and", "type": "text" } ], "index": 21 }, { "bbox": [ 130, 356, 505, 370 ], "spans": [ { "bbox": [ 130, 356, 505, 370 ], "score": 1.0, "content": "Woomyoung Park. Building a role specified open-domain dialogue system leveraging large-scale language", "type": "text" } ], "index": 22 }, { "bbox": [ 130, 366, 505, 379 ], "spans": [ { "bbox": [ 130, 366, 505, 379 ], "score": 1.0, "content": "models. In Proceedings of the 2022 Conference of the North American Chapter of the Association for", "type": "text" } ], "index": 23 }, { "bbox": [ 130, 376, 506, 389 ], "spans": [ { "bbox": [ 130, 376, 506, 389 ], "score": 1.0, "content": "Computational Linguistics: Human Language Technologies, pages 2128–2150, Seattle, United States,", "type": "text" } ], "index": 24 }, { "bbox": [ 130, 386, 327, 399 ], "spans": [ { "bbox": [ 130, 386, 327, 399 ], "score": 1.0, "content": "July 2022. Association for Computational Linguistics.", "type": "text" } ], "index": 25 } ], "index": 23, "bbox_fs": [ 116, 346, 506, 399 ] }, { "type": "text", "bbox": [ 114, 405, 504, 436 ], "lines": [ { "bbox": [ 114, 404, 506, 417 ], "spans": [ { "bbox": [ 114, 404, 506, 417 ], "score": 1.0, "content": "[9] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain,", "type": "text" } ], "index": 26 }, { "bbox": [ 131, 415, 505, 426 ], "spans": [ { "bbox": [ 131, 415, 505, 426 ], "score": 1.0, "content": "Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with", "type": "text" } ], "index": 27 }, { "bbox": [ 130, 425, 445, 436 ], "spans": [ { "bbox": [ 130, 425, 445, 436 ], "score": 1.0, "content": "reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022.", "type": "text" } ], "index": 28 } ], "index": 27, "bbox_fs": [ 114, 404, 506, 436 ] }, { "type": "text", "bbox": [ 111, 443, 504, 474 ], "lines": [ { "bbox": [ 109, 442, 505, 455 ], "spans": [ { "bbox": [ 109, 442, 505, 455 ], "score": 1.0, "content": "[10] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna", "type": "text" } ], "index": 29 }, { "bbox": [ 131, 453, 505, 464 ], "spans": [ { "bbox": [ 131, 453, 505, 464 ], "score": 1.0, "content": "Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from", "type": "text" } ], "index": 30 }, { "bbox": [ 130, 463, 323, 474 ], "spans": [ { "bbox": [ 130, 463, 323, 474 ], "score": 1.0, "content": "ai feedback. arXiv preprint arXiv:2212.08073, 2022.", "type": "text" } ], "index": 31 } ], "index": 30, "bbox_fs": [ 109, 442, 505, 474 ] }, { "type": "text", "bbox": [ 110, 481, 504, 512 ], "lines": [ { "bbox": [ 111, 481, 505, 493 ], "spans": [ { "bbox": [ 111, 481, 505, 493 ], "score": 1.0, "content": "[11] Nolan Bard, Jakob N Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H Francis Song, Emilio", "type": "text" } ], "index": 32 }, { "bbox": [ 129, 490, 505, 503 ], "spans": [ { "bbox": [ 129, 490, 505, 503 ], "score": 1.0, "content": "Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, et al. The hanabi challenge: A new", "type": "text" } ], "index": 33 }, { "bbox": [ 131, 502, 367, 512 ], "spans": [ { "bbox": [ 131, 502, 367, 512 ], "score": 1.0, "content": "frontier for ai research. Artificial Intelligence, 280:103216, 2020.", "type": "text" } ], "index": 34 } ], "index": 33, "bbox_fs": [ 111, 481, 505, 512 ] }, { "type": "text", "bbox": [ 110, 519, 504, 550 ], "lines": [ { "bbox": [ 110, 519, 505, 531 ], "spans": [ { "bbox": [ 110, 519, 505, 531 ], "score": 1.0, "content": "[12] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind", "type": "text" } ], "index": 35 }, { "bbox": [ 131, 529, 506, 540 ], "spans": [ { "bbox": [ 131, 529, 506, 540 ], "score": 1.0, "content": "Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.", "type": "text" } ], "index": 36 }, { "bbox": [ 129, 539, 398, 551 ], "spans": [ { "bbox": [ 129, 539, 398, 551 ], "score": 1.0, "content": "Advances in neural information processing systems, 33:1877–1901, 2020.", "type": "text" } ], "index": 37 } ], "index": 36, "bbox_fs": [ 110, 519, 506, 551 ] }, { "type": "text", "bbox": [ 110, 557, 504, 588 ], "lines": [ { "bbox": [ 110, 558, 506, 569 ], "spans": [ { "bbox": [ 110, 558, 506, 569 ], "score": 1.0, "content": "[13] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar,", "type": "text" } ], "index": 38 }, { "bbox": [ 129, 567, 505, 580 ], "spans": [ { "bbox": [ 129, 567, 505, 580 ], "score": 1.0, "content": "Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early", "type": "text" } ], "index": 39 }, { "bbox": [ 130, 578, 366, 588 ], "spans": [ { "bbox": [ 130, 578, 366, 588 ], "score": 1.0, "content": "experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023.", "type": "text" } ], "index": 40 } ], "index": 39, "bbox_fs": [ 110, 558, 506, 588 ] }, { "type": "text", "bbox": [ 109, 595, 504, 627 ], "lines": [ { "bbox": [ 109, 594, 506, 609 ], "spans": [ { "bbox": [ 109, 594, 506, 609 ], "score": 1.0, "content": "[14] N Carlini, F Tramer, E Wallace, M Jagielski, A Herbert-Voss, K Lee, A Roberts, T Brown, D Song,", "type": "text" } ], "index": 41 }, { "bbox": [ 129, 605, 506, 619 ], "spans": [ { "bbox": [ 129, 605, 506, 619 ], "score": 1.0, "content": "Ú Erlingsson, et al. Extracting training data from large language models. arxiv. Preprint posted online", "type": "text" } ], "index": 42 }, { "bbox": [ 130, 617, 208, 627 ], "spans": [ { "bbox": [ 130, 617, 208, 627 ], "score": 1.0, "content": "December, 14, 2020.", "type": "text" } ], "index": 43 } ], "index": 42, "bbox_fs": [ 109, 594, 506, 627 ] }, { "type": "text", "bbox": [ 109, 635, 504, 665 ], "lines": [ { "bbox": [ 110, 635, 505, 646 ], "spans": [ { "bbox": [ 110, 635, 505, 646 ], "score": 1.0, "content": "[15] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramèr, Borja", "type": "text" } ], "index": 44 }, { "bbox": [ 130, 644, 505, 657 ], "spans": [ { "bbox": [ 130, 644, 505, 657 ], "score": 1.0, "content": "Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. arXiv preprint", "type": "text" } ], "index": 45 }, { "bbox": [ 129, 655, 223, 665 ], "spans": [ { "bbox": [ 129, 655, 223, 665 ], "score": 1.0, "content": "arXiv:2301.13188, 2023.", "type": "text" } ], "index": 46 } ], "index": 45, "bbox_fs": [ 110, 635, 505, 665 ] }, { "type": "text", "bbox": [ 111, 673, 257, 684 ], "lines": [ { "bbox": [ 110, 673, 258, 685 ], "spans": [ { "bbox": [ 110, 673, 258, 685 ], "score": 1.0, "content": "[16] Harrison Chase. Langchain. 2022.", "type": "text" } ], "index": 47 } ], "index": 47, "bbox_fs": [ 110, 673, 258, 685 ] }, { "type": "text", "bbox": [ 109, 691, 504, 722 ], "lines": [ { "bbox": [ 110, 691, 506, 704 ], "spans": [ { "bbox": [ 110, 691, 506, 704 ], "score": 1.0, "content": "[17] Defang Chen, Jian-Ping Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Cross-layer", "type": "text" } ], "index": 48 }, { "bbox": [ 129, 701, 506, 713 ], "spans": [ { "bbox": [ 129, 701, 506, 713 ], "score": 1.0, "content": "distillation with semantic calibration. In Proceedings of the AAAI Conference on Artificial Intelligence,", "type": "text" } ], "index": 49 }, { "bbox": [ 131, 711, 263, 723 ], "spans": [ { "bbox": [ 131, 711, 263, 723 ], "score": 1.0, "content": "volume 35, pages 7028–7036, 2021.", "type": "text" } ], "index": 50 } ], "index": 49, "bbox_fs": [ 110, 691, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 109, 43, 507, 726 ], "lines": [ { "bbox": [ 110, 73, 507, 86 ], "spans": [ { "bbox": [ 110, 73, 507, 86 ], "score": 1.0, "content": "[18] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan,", "type": "text" } ], "index": 0 }, { "bbox": [ 129, 82, 507, 97 ], "spans": [ { "bbox": [ 129, 82, 507, 97 ], "score": 1.0, "content": "Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models", "type": "text" } ], "index": 1 }, { "bbox": [ 128, 92, 339, 106 ], "spans": [ { "bbox": [ 128, 92, 339, 106 ], "score": 1.0, "content": "trained on code. arXiv preprint arXiv:2107.03374, 2021.", "type": "text" } ], "index": 2 }, { "bbox": [ 108, 110, 507, 124 ], "spans": [ { "bbox": [ 108, 110, 507, 124 ], "score": 1.0, "content": "[19] Maximillian Chen, Alexandros Papangelis, Chenyang Tao, Seokhwan Kim, Andy Rosenbaum, Yang Liu,", "type": "text" } ], "index": 3 }, { "bbox": [ 130, 121, 505, 133 ], "spans": [ { "bbox": [ 130, 121, 505, 133 ], "score": 1.0, "content": "Zhou Yu, and Dilek Hakkani-Tur. Places: Prompting language models for social conversation synthesis.", "type": "text" } ], "index": 4 }, { "bbox": [ 128, 130, 278, 143 ], "spans": [ { "bbox": [ 128, 130, 278, 143 ], "score": 1.0, "content": "arXiv preprint arXiv:2302.03269, 2023.", "type": "text" } ], "index": 5 }, { "bbox": [ 110, 149, 506, 160 ], "spans": [ { "bbox": [ 110, 149, 506, 160 ], "score": 1.0, "content": "[20] Maximillian Chen, Alexandros Papangelis, Chenyang Tao, Andy Rosenbaum, Seokhwan Kim, Yang Liu,", "type": "text" } ], "index": 6 }, { "bbox": [ 128, 156, 506, 172 ], "spans": [ { "bbox": [ 128, 156, 506, 172 ], "score": 1.0, "content": "Zhou Yu, and Dilek Hakkani-Tur. Weakly supervised data augmentation through prompting for dialogue", "type": "text" } ], "index": 7 }, { "bbox": [ 130, 169, 482, 181 ], "spans": [ { "bbox": [ 130, 169, 482, 181 ], "score": 1.0, "content": "understanding. NeurIPS 2022 Workshop on Synthetic Data for Empowering ML Research, 2022.", "type": "text" } ], "index": 8 }, { "bbox": [ 110, 186, 504, 199 ], "spans": [ { "bbox": [ 110, 186, 504, 199 ], "score": 1.0, "content": "[21] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan", "type": "text" } ], "index": 9 }, { "bbox": [ 130, 196, 504, 208 ], "spans": [ { "bbox": [ 130, 196, 504, 208 ], "score": 1.0, "content": "Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source", "type": "text" } ], "index": 10 }, { "bbox": [ 130, 205, 370, 219 ], "spans": [ { "bbox": [ 130, 205, 240, 219 ], "score": 1.0, "content": "chatbot impressing gpt-4 with", "type": "text" }, { "bbox": [ 240, 206, 263, 216 ], "score": 0.85, "content": "9 0 \\% ^ { \\ast }", "type": "inline_equation" }, { "bbox": [ 263, 205, 370, 219 ], "score": 1.0, "content": "chatgpt quality, March 2023.", "type": "text" } ], "index": 11 }, { "bbox": [ 109, 223, 507, 237 ], "spans": [ { "bbox": [ 109, 223, 507, 237 ], "score": 1.0, "content": "[22] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts,", "type": "text" } ], "index": 12 }, { "bbox": [ 128, 232, 505, 248 ], "spans": [ { "bbox": [ 128, 232, 505, 248 ], "score": 1.0, "content": "Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language", "type": "text" } ], "index": 13 }, { "bbox": [ 130, 244, 371, 255 ], "spans": [ { "bbox": [ 130, 244, 371, 255 ], "score": 1.0, "content": "modeling with pathways. arXiv preprint arXiv:2204.02311, 2022.", "type": "text" } ], "index": 14 }, { "bbox": [ 109, 260, 506, 274 ], "spans": [ { "bbox": [ 109, 260, 506, 274 ], "score": 1.0, "content": "[23] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep", "type": "text" } ], "index": 15 }, { "bbox": [ 130, 271, 506, 283 ], "spans": [ { "bbox": [ 130, 271, 506, 283 ], "score": 1.0, "content": "reinforcement learning from human preferences. Advances in neural information processing systems, 30,", "type": "text" } ], "index": 16 }, { "bbox": [ 130, 279, 156, 294 ], "spans": [ { "bbox": [ 130, 279, 156, 294 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 17 }, { "bbox": [ 109, 297, 506, 312 ], "spans": [ { "bbox": [ 109, 297, 506, 312 ], "score": 1.0, "content": "[24] Caroline Claus and Craig Boutilier. The dynamics of reinforcement learning in cooperative multiagent", "type": "text" } ], "index": 18 }, { "bbox": [ 129, 308, 241, 321 ], "spans": [ { "bbox": [ 129, 308, 241, 321 ], "score": 1.0, "content": "systems. In AAAI/IAAI, 1998.", "type": "text" } ], "index": 19 }, { "bbox": [ 109, 325, 506, 340 ], "spans": [ { "bbox": [ 109, 325, 506, 340 ], "score": 1.0, "content": "[25] Antonia Creswell, Murray Shanahan, and Irina Higgins. Selection-inference: Exploiting large language", "type": "text" } ], "index": 20 }, { "bbox": [ 130, 336, 308, 348 ], "spans": [ { "bbox": [ 130, 336, 308, 348 ], "score": 1.0, "content": "models for interpretable logical reasoning, 2022.", "type": "text" } ], "index": 21 }, { "bbox": [ 109, 352, 506, 367 ], "spans": [ { "bbox": [ 109, 352, 506, 367 ], "score": 1.0, "content": "[26] Allan Dafoe, Yoram Bachrach, Gillian Hadfield, Eric Horvitz, Kate Larson, and Thore Graepel. Coopera-", "type": "text" } ], "index": 22 }, { "bbox": [ 130, 364, 443, 376 ], "spans": [ { "bbox": [ 130, 364, 443, 376 ], "score": 1.0, "content": "tive ai: machines must learn to find common ground. Nature, 593(7857):33–36, 2021.", "type": "text" } ], "index": 23 }, { "bbox": [ 109, 380, 506, 394 ], "spans": [ { "bbox": [ 109, 380, 506, 394 ], "score": 1.0, "content": "[27] Allan Dafoe, Edward Hughes, Yoram Bachrach, Tantum Collins, Kevin R McKee, Joel Z Leibo, Kate", "type": "text" } ], "index": 24 }, { "bbox": [ 128, 390, 503, 404 ], "spans": [ { "bbox": [ 128, 390, 503, 404 ], "score": 1.0, "content": "Larson, and Thore Graepel. Open problems in cooperative ai. arXiv preprint arXiv:2012.08630, 2020.", "type": "text" } ], "index": 25 }, { "bbox": [ 109, 408, 507, 423 ], "spans": [ { "bbox": [ 109, 408, 507, 423 ], "score": 1.0, "content": "[28] Yali Du, Bo Liu, Vincent Moens, Ziqi Liu, Zhicheng Ren, Jun Wang, Xu Chen, and Haifeng Zhang.", "type": "text" } ], "index": 26 }, { "bbox": [ 130, 419, 506, 431 ], "spans": [ { "bbox": [ 130, 419, 506, 431 ], "score": 1.0, "content": "Learning correlated communication topology in multi-agent reinforcement learning. In Proceedings of", "type": "text" } ], "index": 27 }, { "bbox": [ 128, 428, 507, 442 ], "spans": [ { "bbox": [ 128, 428, 507, 442 ], "score": 1.0, "content": "the 20th International Conference on Autonomous Agents and MultiAgent Systems, pages 456–464, 2021.", "type": "text" } ], "index": 28 }, { "bbox": [ 110, 447, 505, 459 ], "spans": [ { "bbox": [ 110, 447, 505, 459 ], "score": 1.0, "content": "[29] Tim Finin, Richard Fritzson, Don McKay, and Robin McEntire. Kqml as an agent communication", "type": "text" } ], "index": 29 }, { "bbox": [ 131, 456, 507, 469 ], "spans": [ { "bbox": [ 131, 456, 507, 469 ], "score": 1.0, "content": "language. In Proceedings of the third international conference on Information and knowledge management,", "type": "text" } ], "index": 30 }, { "bbox": [ 128, 466, 213, 478 ], "spans": [ { "bbox": [ 128, 466, 213, 478 ], "score": 1.0, "content": "pages 456–463, 1994.", "type": "text" } ], "index": 31 }, { "bbox": [ 109, 482, 507, 497 ], "spans": [ { "bbox": [ 109, 482, 507, 497 ], "score": 1.0, "content": "[30] Jakob Foerster, Ioannis Alexandros Assael, Nando De Freitas, and Shimon Whiteson. Learning to", "type": "text" } ], "index": 32 }, { "bbox": [ 128, 492, 507, 507 ], "spans": [ { "bbox": [ 128, 492, 507, 507 ], "score": 1.0, "content": "communicate with deep multi-agent reinforcement learning. Advances in neural information processing", "type": "text" } ], "index": 33 }, { "bbox": [ 129, 504, 200, 516 ], "spans": [ { "bbox": [ 129, 504, 200, 516 ], "score": 1.0, "content": "systems, 29, 2016.", "type": "text" } ], "index": 34 }, { "bbox": [ 109, 520, 507, 534 ], "spans": [ { "bbox": [ 109, 520, 507, 534 ], "score": 1.0, "content": "[31] Yao Fu, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. Complexity-based prompting for", "type": "text" } ], "index": 35 }, { "bbox": [ 128, 531, 357, 544 ], "spans": [ { "bbox": [ 128, 531, 357, 544 ], "score": 1.0, "content": "multi-step reasoning. arXiv preprint arXiv:2210.00720, 2022.", "type": "text" } ], "index": 36 }, { "bbox": [ 109, 547, 504, 561 ], "spans": [ { "bbox": [ 109, 547, 504, 561 ], "score": 1.0, "content": "[32] Iason Gabriel. Artificial intelligence, values, and alignment. Minds and Machines, 30:411 – 437, 2020.", "type": "text" } ], "index": 37 }, { "bbox": [ 109, 566, 507, 579 ], "spans": [ { "bbox": [ 109, 566, 507, 579 ], "score": 1.0, "content": "[33] Jianfeng Gao, Michel Galley, and Lihong Li. Neural approaches to conversational ai. In The 41st", "type": "text" } ], "index": 38 }, { "bbox": [ 128, 575, 506, 590 ], "spans": [ { "bbox": [ 128, 575, 506, 590 ], "score": 1.0, "content": "International ACM SIGIR Conference on Research & Development in Information Retrieval, pages", "type": "text" } ], "index": 39 }, { "bbox": [ 130, 586, 198, 598 ], "spans": [ { "bbox": [ 130, 586, 198, 598 ], "score": 1.0, "content": "1371–1374, 2018.", "type": "text" } ], "index": 40 }, { "bbox": [ 109, 603, 507, 618 ], "spans": [ { "bbox": [ 109, 603, 507, 618 ], "score": 1.0, "content": "[34] Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding,", "type": "text" } ], "index": 41 }, { "bbox": [ 129, 613, 507, 627 ], "spans": [ { "bbox": [ 129, 613, 507, 627 ], "score": 1.0, "content": "Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric Tang, Anish Thite,", "type": "text" } ], "index": 42 }, { "bbox": [ 128, 622, 507, 637 ], "spans": [ { "bbox": [ 128, 622, 507, 637 ], "score": 1.0, "content": "Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language model evaluation, September", "type": "text" } ], "index": 43 }, { "bbox": [ 129, 633, 155, 646 ], "spans": [ { "bbox": [ 129, 633, 155, 646 ], "score": 1.0, "content": "2021.", "type": "text" } ], "index": 44 }, { "bbox": [ 110, 651, 506, 663 ], "spans": [ { "bbox": [ 110, 651, 506, 663 ], "score": 1.0, "content": "[35] Amelia Glaese, Nat McAleese, Maja Trebacz, John Aslanides, Vlad Firoiu, Timo Ewalds, Maribeth Rauh,", "type": "text" } ], "index": 45 }, { "bbox": [ 130, 662, 506, 674 ], "spans": [ { "bbox": [ 130, 662, 506, 674 ], "score": 1.0, "content": "Laura Weidinger, Martin Chadwick, Phoebe Thacker, Lucy Campbell-Gillingham, Jonathan Uesato,", "type": "text" } ], "index": 46 }, { "bbox": [ 130, 672, 505, 684 ], "spans": [ { "bbox": [ 130, 672, 505, 684 ], "score": 1.0, "content": "Po-Sen Huang, Ramona Comanescu, Fan Yang, Abigail See, Sumanth Dathathri, Rory Greig, Charlie", "type": "text" } ], "index": 47 }, { "bbox": [ 129, 681, 506, 695 ], "spans": [ { "bbox": [ 129, 681, 506, 695 ], "score": 1.0, "content": "Chen, Doug Fritz, Jaume Sanchez Elias, Richard Green, Sona Mokrá, Nicholas Fernando, Boxi Wu, ˇ", "type": "text" } ], "index": 48 }, { "bbox": [ 128, 690, 506, 705 ], "spans": [ { "bbox": [ 128, 690, 506, 705 ], "score": 1.0, "content": "Rachel Foley, Susannah Young, Iason Gabriel, William Isaac, John Mellor, Demis Hassabis, Koray", "type": "text" } ], "index": 49 }, { "bbox": [ 128, 699, 507, 714 ], "spans": [ { "bbox": [ 128, 699, 507, 714 ], "score": 1.0, "content": "Kavukcuoglu, Lisa Anne Hendricks, and Geoffrey Irving. Improving alignment of dialogue agents via", "type": "text" } ], "index": 50 }, { "bbox": [ 128, 711, 258, 724 ], "spans": [ { "bbox": [ 128, 711, 258, 724 ], "score": 1.0, "content": "targeted human judgements, 2022.", "type": "text" } ], "index": 51 } ], "index": 25.5 } ], "page_idx": 11, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 313, 754 ], "spans": [ { "bbox": [ 299, 740, 313, 754 ], "score": 1.0, "content": "12", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 109, 43, 507, 726 ], "lines": [ { "bbox": [ 110, 73, 507, 86 ], "spans": [ { "bbox": [ 110, 73, 507, 86 ], "score": 1.0, "content": "[18] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan,", "type": "text" } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 129, 82, 507, 97 ], "spans": [ { "bbox": [ 129, 82, 507, 97 ], "score": 1.0, "content": "Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models", "type": "text" } ], "index": 1 }, { "bbox": [ 128, 92, 339, 106 ], "spans": [ { "bbox": [ 128, 92, 339, 106 ], "score": 1.0, "content": "trained on code. arXiv preprint arXiv:2107.03374, 2021.", "type": "text" } ], "index": 2, "is_list_end_line": true }, { "bbox": [ 108, 110, 507, 124 ], "spans": [ { "bbox": [ 108, 110, 507, 124 ], "score": 1.0, "content": "[19] Maximillian Chen, Alexandros Papangelis, Chenyang Tao, Seokhwan Kim, Andy Rosenbaum, Yang Liu,", "type": "text" } ], "index": 3, "is_list_start_line": true }, { "bbox": [ 130, 121, 505, 133 ], "spans": [ { "bbox": [ 130, 121, 505, 133 ], "score": 1.0, "content": "Zhou Yu, and Dilek Hakkani-Tur. Places: Prompting language models for social conversation synthesis.", "type": "text" } ], "index": 4 }, { "bbox": [ 128, 130, 278, 143 ], "spans": [ { "bbox": [ 128, 130, 278, 143 ], "score": 1.0, "content": "arXiv preprint arXiv:2302.03269, 2023.", "type": "text" } ], "index": 5, "is_list_end_line": true }, { "bbox": [ 110, 149, 506, 160 ], "spans": [ { "bbox": [ 110, 149, 506, 160 ], "score": 1.0, "content": "[20] Maximillian Chen, Alexandros Papangelis, Chenyang Tao, Andy Rosenbaum, Seokhwan Kim, Yang Liu,", "type": "text" } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 128, 156, 506, 172 ], "spans": [ { "bbox": [ 128, 156, 506, 172 ], "score": 1.0, "content": "Zhou Yu, and Dilek Hakkani-Tur. Weakly supervised data augmentation through prompting for dialogue", "type": "text" } ], "index": 7 }, { "bbox": [ 130, 169, 482, 181 ], "spans": [ { "bbox": [ 130, 169, 482, 181 ], "score": 1.0, "content": "understanding. NeurIPS 2022 Workshop on Synthetic Data for Empowering ML Research, 2022.", "type": "text" } ], "index": 8, "is_list_end_line": true }, { "bbox": [ 110, 186, 504, 199 ], "spans": [ { "bbox": [ 110, 186, 504, 199 ], "score": 1.0, "content": "[21] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan", "type": "text" } ], "index": 9, "is_list_start_line": true }, { "bbox": [ 130, 196, 504, 208 ], "spans": [ { "bbox": [ 130, 196, 504, 208 ], "score": 1.0, "content": "Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source", "type": "text" } ], "index": 10 }, { "bbox": [ 130, 205, 370, 219 ], "spans": [ { "bbox": [ 130, 205, 240, 219 ], "score": 1.0, "content": "chatbot impressing gpt-4 with", "type": "text" }, { "bbox": [ 240, 206, 263, 216 ], "score": 0.85, "content": "9 0 \\% ^ { \\ast }", "type": "inline_equation" }, { "bbox": [ 263, 205, 370, 219 ], "score": 1.0, "content": "chatgpt quality, March 2023.", "type": "text" } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 109, 223, 507, 237 ], "spans": [ { "bbox": [ 109, 223, 507, 237 ], "score": 1.0, "content": "[22] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts,", "type": "text" } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 128, 232, 505, 248 ], "spans": [ { "bbox": [ 128, 232, 505, 248 ], "score": 1.0, "content": "Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language", "type": "text" } ], "index": 13 }, { "bbox": [ 130, 244, 371, 255 ], "spans": [ { "bbox": [ 130, 244, 371, 255 ], "score": 1.0, "content": "modeling with pathways. arXiv preprint arXiv:2204.02311, 2022.", "type": "text" } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 109, 260, 506, 274 ], "spans": [ { "bbox": [ 109, 260, 506, 274 ], "score": 1.0, "content": "[23] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep", "type": "text" } ], "index": 15, "is_list_start_line": true }, { "bbox": [ 130, 271, 506, 283 ], "spans": [ { "bbox": [ 130, 271, 506, 283 ], "score": 1.0, "content": "reinforcement learning from human preferences. Advances in neural information processing systems, 30,", "type": "text" } ], "index": 16 }, { "bbox": [ 130, 279, 156, 294 ], "spans": [ { "bbox": [ 130, 279, 156, 294 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 17, "is_list_end_line": true }, { "bbox": [ 109, 297, 506, 312 ], "spans": [ { "bbox": [ 109, 297, 506, 312 ], "score": 1.0, "content": "[24] Caroline Claus and Craig Boutilier. The dynamics of reinforcement learning in cooperative multiagent", "type": "text" } ], "index": 18, "is_list_start_line": true }, { "bbox": [ 129, 308, 241, 321 ], "spans": [ { "bbox": [ 129, 308, 241, 321 ], "score": 1.0, "content": "systems. In AAAI/IAAI, 1998.", "type": "text" } ], "index": 19, "is_list_end_line": true }, { "bbox": [ 109, 325, 506, 340 ], "spans": [ { "bbox": [ 109, 325, 506, 340 ], "score": 1.0, "content": "[25] Antonia Creswell, Murray Shanahan, and Irina Higgins. Selection-inference: Exploiting large language", "type": "text" } ], "index": 20, "is_list_start_line": true }, { "bbox": [ 130, 336, 308, 348 ], "spans": [ { "bbox": [ 130, 336, 308, 348 ], "score": 1.0, "content": "models for interpretable logical reasoning, 2022.", "type": "text" } ], "index": 21, "is_list_end_line": true }, { "bbox": [ 109, 352, 506, 367 ], "spans": [ { "bbox": [ 109, 352, 506, 367 ], "score": 1.0, "content": "[26] Allan Dafoe, Yoram Bachrach, Gillian Hadfield, Eric Horvitz, Kate Larson, and Thore Graepel. Coopera-", "type": "text" } ], "index": 22, "is_list_start_line": true }, { "bbox": [ 130, 364, 443, 376 ], "spans": [ { "bbox": [ 130, 364, 443, 376 ], "score": 1.0, "content": "tive ai: machines must learn to find common ground. Nature, 593(7857):33–36, 2021.", "type": "text" } ], "index": 23, "is_list_end_line": true }, { "bbox": [ 109, 380, 506, 394 ], "spans": [ { "bbox": [ 109, 380, 506, 394 ], "score": 1.0, "content": "[27] Allan Dafoe, Edward Hughes, Yoram Bachrach, Tantum Collins, Kevin R McKee, Joel Z Leibo, Kate", "type": "text" } ], "index": 24, "is_list_start_line": true }, { "bbox": [ 128, 390, 503, 404 ], "spans": [ { "bbox": [ 128, 390, 503, 404 ], "score": 1.0, "content": "Larson, and Thore Graepel. Open problems in cooperative ai. arXiv preprint arXiv:2012.08630, 2020.", "type": "text" } ], "index": 25 }, { "bbox": [ 109, 408, 507, 423 ], "spans": [ { "bbox": [ 109, 408, 507, 423 ], "score": 1.0, "content": "[28] Yali Du, Bo Liu, Vincent Moens, Ziqi Liu, Zhicheng Ren, Jun Wang, Xu Chen, and Haifeng Zhang.", "type": "text" } ], "index": 26, "is_list_start_line": true }, { "bbox": [ 130, 419, 506, 431 ], "spans": [ { "bbox": [ 130, 419, 506, 431 ], "score": 1.0, "content": "Learning correlated communication topology in multi-agent reinforcement learning. In Proceedings of", "type": "text" } ], "index": 27 }, { "bbox": [ 128, 428, 507, 442 ], "spans": [ { "bbox": [ 128, 428, 507, 442 ], "score": 1.0, "content": "the 20th International Conference on Autonomous Agents and MultiAgent Systems, pages 456–464, 2021.", "type": "text" } ], "index": 28 }, { "bbox": [ 110, 447, 505, 459 ], "spans": [ { "bbox": [ 110, 447, 505, 459 ], "score": 1.0, "content": "[29] Tim Finin, Richard Fritzson, Don McKay, and Robin McEntire. Kqml as an agent communication", "type": "text" } ], "index": 29, "is_list_start_line": true }, { "bbox": [ 131, 456, 507, 469 ], "spans": [ { "bbox": [ 131, 456, 507, 469 ], "score": 1.0, "content": "language. In Proceedings of the third international conference on Information and knowledge management,", "type": "text" } ], "index": 30 }, { "bbox": [ 128, 466, 213, 478 ], "spans": [ { "bbox": [ 128, 466, 213, 478 ], "score": 1.0, "content": "pages 456–463, 1994.", "type": "text" } ], "index": 31, "is_list_end_line": true }, { "bbox": [ 109, 482, 507, 497 ], "spans": [ { "bbox": [ 109, 482, 507, 497 ], "score": 1.0, "content": "[30] Jakob Foerster, Ioannis Alexandros Assael, Nando De Freitas, and Shimon Whiteson. Learning to", "type": "text" } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 128, 492, 507, 507 ], "spans": [ { "bbox": [ 128, 492, 507, 507 ], "score": 1.0, "content": "communicate with deep multi-agent reinforcement learning. Advances in neural information processing", "type": "text" } ], "index": 33 }, { "bbox": [ 129, 504, 200, 516 ], "spans": [ { "bbox": [ 129, 504, 200, 516 ], "score": 1.0, "content": "systems, 29, 2016.", "type": "text" } ], "index": 34, "is_list_end_line": true }, { "bbox": [ 109, 520, 507, 534 ], "spans": [ { "bbox": [ 109, 520, 507, 534 ], "score": 1.0, "content": "[31] Yao Fu, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. Complexity-based prompting for", "type": "text" } ], "index": 35, "is_list_start_line": true }, { "bbox": [ 128, 531, 357, 544 ], "spans": [ { "bbox": [ 128, 531, 357, 544 ], "score": 1.0, "content": "multi-step reasoning. arXiv preprint arXiv:2210.00720, 2022.", "type": "text" } ], "index": 36, "is_list_end_line": true }, { "bbox": [ 109, 547, 504, 561 ], "spans": [ { "bbox": [ 109, 547, 504, 561 ], "score": 1.0, "content": "[32] Iason Gabriel. Artificial intelligence, values, and alignment. Minds and Machines, 30:411 – 437, 2020.", "type": "text" } ], "index": 37, "is_list_start_line": true }, { "bbox": [ 109, 566, 507, 579 ], "spans": [ { "bbox": [ 109, 566, 507, 579 ], "score": 1.0, "content": "[33] Jianfeng Gao, Michel Galley, and Lihong Li. Neural approaches to conversational ai. In The 41st", "type": "text" } ], "index": 38, "is_list_start_line": true }, { "bbox": [ 128, 575, 506, 590 ], "spans": [ { "bbox": [ 128, 575, 506, 590 ], "score": 1.0, "content": "International ACM SIGIR Conference on Research & Development in Information Retrieval, pages", "type": "text" } ], "index": 39 }, { "bbox": [ 130, 586, 198, 598 ], "spans": [ { "bbox": [ 130, 586, 198, 598 ], "score": 1.0, "content": "1371–1374, 2018.", "type": "text" } ], "index": 40, "is_list_end_line": true }, { "bbox": [ 109, 603, 507, 618 ], "spans": [ { "bbox": [ 109, 603, 507, 618 ], "score": 1.0, "content": "[34] Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding,", "type": "text" } ], "index": 41, "is_list_start_line": true }, { "bbox": [ 129, 613, 507, 627 ], "spans": [ { "bbox": [ 129, 613, 507, 627 ], "score": 1.0, "content": "Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric Tang, Anish Thite,", "type": "text" } ], "index": 42 }, { "bbox": [ 128, 622, 507, 637 ], "spans": [ { "bbox": [ 128, 622, 507, 637 ], "score": 1.0, "content": "Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language model evaluation, September", "type": "text" } ], "index": 43 }, { "bbox": [ 129, 633, 155, 646 ], "spans": [ { "bbox": [ 129, 633, 155, 646 ], "score": 1.0, "content": "2021.", "type": "text" } ], "index": 44, "is_list_end_line": true }, { "bbox": [ 110, 651, 506, 663 ], "spans": [ { "bbox": [ 110, 651, 506, 663 ], "score": 1.0, "content": "[35] Amelia Glaese, Nat McAleese, Maja Trebacz, John Aslanides, Vlad Firoiu, Timo Ewalds, Maribeth Rauh,", "type": "text" } ], "index": 45, "is_list_start_line": true }, { "bbox": [ 130, 662, 506, 674 ], "spans": [ { "bbox": [ 130, 662, 506, 674 ], "score": 1.0, "content": "Laura Weidinger, Martin Chadwick, Phoebe Thacker, Lucy Campbell-Gillingham, Jonathan Uesato,", "type": "text" } ], "index": 46 }, { "bbox": [ 130, 672, 505, 684 ], "spans": [ { "bbox": [ 130, 672, 505, 684 ], "score": 1.0, "content": "Po-Sen Huang, Ramona Comanescu, Fan Yang, Abigail See, Sumanth Dathathri, Rory Greig, Charlie", "type": "text" } ], "index": 47 }, { "bbox": [ 129, 681, 506, 695 ], "spans": [ { "bbox": [ 129, 681, 506, 695 ], "score": 1.0, "content": "Chen, Doug Fritz, Jaume Sanchez Elias, Richard Green, Sona Mokrá, Nicholas Fernando, Boxi Wu, ˇ", "type": "text" } ], "index": 48 }, { "bbox": [ 128, 690, 506, 705 ], "spans": [ { "bbox": [ 128, 690, 506, 705 ], "score": 1.0, "content": "Rachel Foley, Susannah Young, Iason Gabriel, William Isaac, John Mellor, Demis Hassabis, Koray", "type": "text" } ], "index": 49 }, { "bbox": [ 128, 699, 507, 714 ], "spans": [ { "bbox": [ 128, 699, 507, 714 ], "score": 1.0, "content": "Kavukcuoglu, Lisa Anne Hendricks, and Geoffrey Irving. Improving alignment of dialogue agents via", "type": "text" } ], "index": 50 }, { "bbox": [ 128, 711, 258, 724 ], "spans": [ { "bbox": [ 128, 711, 258, 724 ], "score": 1.0, "content": "targeted human judgements, 2022.", "type": "text" } ], "index": 51, "is_list_end_line": true }, { "bbox": [ 110, 73, 506, 85 ], "spans": [ { "bbox": [ 110, 73, 506, 85 ], "score": 1.0, "content": "[36] Amelia Glaese, Nat McAleese, Maja Tr˛ebacz, John Aslanides, Vlad Firoiu, Timo Ewalds, Maribeth Rauh,", "type": "text", "cross_page": true } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 128, 82, 507, 96 ], "spans": [ { "bbox": [ 128, 82, 507, 96 ], "score": 1.0, "content": "Laura Weidinger, Martin Chadwick, Phoebe Thacker, et al. Improving alignment of dialogue agents via", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 128, 92, 383, 106 ], "spans": [ { "bbox": [ 128, 92, 383, 106 ], "score": 1.0, "content": "targeted human judgements. arXiv preprint arXiv:2209.14375, 2022.", "type": "text", "cross_page": true } ], "index": 2, "is_list_end_line": true }, { "bbox": [ 111, 111, 506, 123 ], "spans": [ { "bbox": [ 111, 111, 506, 123 ], "score": 1.0, "content": "[37] Josh A Goldstein, Girish Sastry, Micah Musser, Renee DiResta, Matthew Gentzel, and Katerina Se-", "type": "text", "cross_page": true } ], "index": 3, "is_list_start_line": true }, { "bbox": [ 129, 120, 506, 133 ], "spans": [ { "bbox": [ 129, 120, 506, 133 ], "score": 1.0, "content": "dova. Generative language models and automated influence operations: Emerging threats and potential", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 129, 131, 323, 142 ], "spans": [ { "bbox": [ 129, 131, 323, 142 ], "score": 1.0, "content": "mitigations. arXiv preprint arXiv:2301.04246, 2023.", "type": "text", "cross_page": true } ], "index": 5, "is_list_end_line": true }, { "bbox": [ 110, 149, 506, 161 ], "spans": [ { "bbox": [ 110, 149, 506, 161 ], "score": 1.0, "content": "[38] Dylan Hadfield-Menell. The principal-agent alignment problem in artificial intelligence. Ph. D. disserta-", "type": "text", "cross_page": true } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 129, 159, 173, 171 ], "spans": [ { "bbox": [ 129, 159, 173, 171 ], "score": 1.0, "content": "tion, 2021.", "type": "text", "cross_page": true } ], "index": 7, "is_list_end_line": true }, { "bbox": [ 109, 176, 507, 190 ], "spans": [ { "bbox": [ 109, 176, 507, 190 ], "score": 1.0, "content": "[39] Dylan Hadfield-Menell, McKane Andrus, and Gillian Hadfield. Legible normativity for ai alignment:", "type": "text", "cross_page": true } ], "index": 8, "is_list_start_line": true }, { "bbox": [ 129, 185, 507, 200 ], "spans": [ { "bbox": [ 129, 185, 507, 200 ], "score": 1.0, "content": "The value of silly rules. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society,", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 128, 197, 213, 208 ], "spans": [ { "bbox": [ 128, 197, 213, 208 ], "score": 1.0, "content": "pages 115–121, 2019.", "type": "text", "cross_page": true } ], "index": 10, "is_list_end_line": true }, { "bbox": [ 110, 215, 506, 227 ], "spans": [ { "bbox": [ 110, 215, 506, 227 ], "score": 1.0, "content": "[40] Dylan Hadfield-Menell, Stuart J Russell, Pieter Abbeel, and Anca Dragan. Cooperative inverse reinforce-", "type": "text", "cross_page": true } ], "index": 11, "is_list_start_line": true }, { "bbox": [ 128, 224, 412, 237 ], "spans": [ { "bbox": [ 128, 224, 412, 237 ], "score": 1.0, "content": "ment learning. Advances in neural information processing systems, 29, 2016.", "type": "text", "cross_page": true } ], "index": 12, "is_list_end_line": true }, { "bbox": [ 110, 243, 506, 254 ], "spans": [ { "bbox": [ 110, 243, 506, 254 ], "score": 1.0, "content": "[41] Serhii Havrylov and Ivan Titov. Emergence of language with multi-agent games: Learning to communicate", "type": "text", "cross_page": true } ], "index": 13, "is_list_start_line": true }, { "bbox": [ 130, 253, 457, 265 ], "spans": [ { "bbox": [ 130, 253, 457, 265 ], "score": 1.0, "content": "with sequences of symbols. Advances in neural information processing systems, 30, 2017.", "type": "text", "cross_page": true } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 109, 269, 506, 282 ], "spans": [ { "bbox": [ 109, 269, 506, 282 ], "score": 1.0, "content": "[42] Peter Henderson, Koustuv Sinha, Nicolas Angelard-Gontier, Nan Rosemary Ke, Genevieve Fried, Ryan", "type": "text", "cross_page": true } ], "index": 15, "is_list_start_line": true }, { "bbox": [ 128, 279, 506, 293 ], "spans": [ { "bbox": [ 128, 279, 506, 293 ], "score": 1.0, "content": "Lowe, and Joelle Pineau. Ethical challenges in data-driven dialogue systems. In Proceedings of the 2018", "type": "text", "cross_page": true } ], "index": 16 }, { "bbox": [ 128, 290, 396, 304 ], "spans": [ { "bbox": [ 128, 290, 396, 304 ], "score": 1.0, "content": "AAAI/ACM Conference on AI, Ethics, and Society, pages 123–129, 2018.", "type": "text", "cross_page": true } ], "index": 17, "is_list_end_line": true }, { "bbox": [ 108, 305, 508, 323 ], "spans": [ { "bbox": [ 108, 305, 508, 323 ], "score": 1.0, "content": "[43] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song,", "type": "text", "cross_page": true } ], "index": 18, "is_list_start_line": true }, { "bbox": [ 128, 317, 506, 331 ], "spans": [ { "bbox": [ 128, 317, 506, 331 ], "score": 1.0, "content": "and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint", "type": "text", "cross_page": true } ], "index": 19 }, { "bbox": [ 129, 329, 225, 339 ], "spans": [ { "bbox": [ 129, 329, 225, 339 ], "score": 1.0, "content": "arXiv:2103.03874, 2021.", "type": "text", "cross_page": true } ], "index": 20, "is_list_end_line": true }, { "bbox": [ 110, 346, 506, 358 ], "spans": [ { "bbox": [ 110, 346, 506, 358 ], "score": 1.0, "content": "[44] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of", "type": "text", "cross_page": true } ], "index": 21, "is_list_start_line": true }, { "bbox": [ 130, 356, 506, 368 ], "spans": [ { "bbox": [ 130, 356, 506, 368 ], "score": 1.0, "content": "activation boundaries formed by hidden neurons. In Proceedings of the AAAI Conference on Artificial", "type": "text", "cross_page": true } ], "index": 22 }, { "bbox": [ 130, 366, 310, 378 ], "spans": [ { "bbox": [ 130, 366, 310, 378 ], "score": 1.0, "content": "Intelligence, volume 33, pages 3779–3787, 2019.", "type": "text", "cross_page": true } ], "index": 23, "is_list_end_line": true }, { "bbox": [ 110, 383, 506, 395 ], "spans": [ { "bbox": [ 110, 383, 506, 395 ], "score": 1.0, "content": "[45] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv", "type": "text", "cross_page": true } ], "index": 24, "is_list_start_line": true }, { "bbox": [ 128, 393, 257, 406 ], "spans": [ { "bbox": [ 128, 393, 257, 406 ], "score": 1.0, "content": "preprint arXiv:1503.02531, 2015.", "type": "text", "cross_page": true } ], "index": 25, "is_list_end_line": true }, { "bbox": [ 110, 411, 506, 423 ], "spans": [ { "bbox": [ 110, 411, 506, 423 ], "score": 1.0, "content": "[46] Namgyu Ho, Laura Schmid, and Se-Young Yun. Large language models are reasoning teachers. arXiv", "type": "text", "cross_page": true } ], "index": 26, "is_list_start_line": true }, { "bbox": [ 128, 421, 256, 434 ], "spans": [ { "bbox": [ 128, 421, 256, 434 ], "score": 1.0, "content": "preprint arXiv:2212.10071, 2022.", "type": "text", "cross_page": true } ], "index": 27, "is_list_end_line": true }, { "bbox": [ 110, 439, 506, 451 ], "spans": [ { "bbox": [ 110, 439, 506, 451 ], "score": 1.0, "content": "[47] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford,", "type": "text", "cross_page": true } ], "index": 28, "is_list_start_line": true }, { "bbox": [ 131, 449, 505, 460 ], "spans": [ { "bbox": [ 131, 449, 505, 460 ], "score": 1.0, "content": "Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal", "type": "text", "cross_page": true } ], "index": 29 }, { "bbox": [ 129, 458, 363, 471 ], "spans": [ { "bbox": [ 129, 458, 363, 471 ], "score": 1.0, "content": "large language models. arXiv preprint arXiv:2203.15556, 2022.", "type": "text", "cross_page": true } ], "index": 30, "is_list_end_line": true }, { "bbox": [ 109, 475, 506, 491 ], "spans": [ { "bbox": [ 109, 475, 506, 491 ], "score": 1.0, "content": "[48] Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. Unnatural instructions: Tuning language", "type": "text", "cross_page": true } ], "index": 31, "is_list_start_line": true }, { "bbox": [ 128, 486, 416, 499 ], "spans": [ { "bbox": [ 128, 486, 416, 499 ], "score": 1.0, "content": "models with (almost) no human labor. arXiv preprint arXiv:2212.09689, 2022.", "type": "text", "cross_page": true } ], "index": 32, "is_list_end_line": true }, { "bbox": [ 110, 505, 506, 517 ], "spans": [ { "bbox": [ 110, 505, 506, 517 ], "score": 1.0, "content": "[49] Ehsan Hosseini-Asl, Bryan McCann, Chien-Sheng Wu, Semih Yavuz, and Richard Socher. A sim-", "type": "text", "cross_page": true } ], "index": 33, "is_list_start_line": true }, { "bbox": [ 130, 514, 507, 528 ], "spans": [ { "bbox": [ 130, 514, 507, 528 ], "score": 1.0, "content": "ple language model for task-oriented dialogue. Advances in Neural Information Processing Systems,", "type": "text", "cross_page": true } ], "index": 34 }, { "bbox": [ 131, 525, 219, 537 ], "spans": [ { "bbox": [ 131, 525, 219, 537 ], "score": 1.0, "content": "33:20179–20191, 2020.", "type": "text", "cross_page": true } ], "index": 35, "is_list_end_line": true }, { "bbox": [ 109, 542, 506, 556 ], "spans": [ { "bbox": [ 109, 542, 506, 556 ], "score": 1.0, "content": "[50] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot", "type": "text", "cross_page": true } ], "index": 36, "is_list_start_line": true }, { "bbox": [ 128, 552, 507, 565 ], "spans": [ { "bbox": [ 128, 552, 507, 565 ], "score": 1.0, "content": "planners: Extracting actionable knowledge for embodied agents. arXiv preprint arXiv:2201.07207, 2022.", "type": "text", "cross_page": true } ], "index": 37 }, { "bbox": [ 109, 569, 506, 583 ], "spans": [ { "bbox": [ 109, 569, 506, 583 ], "score": 1.0, "content": "[51] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan", "type": "text", "cross_page": true } ], "index": 38, "is_list_start_line": true }, { "bbox": [ 128, 579, 506, 594 ], "spans": [ { "bbox": [ 128, 579, 506, 594 ], "score": 1.0, "content": "Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through", "type": "text", "cross_page": true } ], "index": 39 }, { "bbox": [ 129, 590, 394, 602 ], "spans": [ { "bbox": [ 129, 590, 394, 602 ], "score": 1.0, "content": "planning with language models. arXiv preprint arXiv:2207.05608, 2022.", "type": "text", "cross_page": true } ], "index": 40, "is_list_end_line": true }, { "bbox": [ 109, 607, 506, 622 ], "spans": [ { "bbox": [ 109, 607, 506, 622 ], "score": 1.0, "content": "[52] Shima Imani, Liang Du, and Harsh Shrivastava. Mathprompter: Mathematical reasoning using large", "type": "text", "cross_page": true } ], "index": 41, "is_list_start_line": true }, { "bbox": [ 128, 618, 344, 630 ], "spans": [ { "bbox": [ 128, 618, 344, 630 ], "score": 1.0, "content": "language models. arXiv preprint arXiv:2303.05398, 2023.", "type": "text", "cross_page": true } ], "index": 42, "is_list_end_line": true }, { "bbox": [ 110, 636, 506, 648 ], "spans": [ { "bbox": [ 110, 636, 506, 648 ], "score": 1.0, "content": "[53] Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru, Todor Mihaylov, Dániel Simig, Ping Yu, Kurt", "type": "text", "cross_page": true } ], "index": 43, "is_list_start_line": true }, { "bbox": [ 130, 646, 506, 658 ], "spans": [ { "bbox": [ 130, 646, 506, 658 ], "score": 1.0, "content": "Shuster, Tianlu Wang, Qing Liu, Punit Singh Koura, et al. Opt-iml: Scaling language model instruction", "type": "text", "cross_page": true } ], "index": 44 }, { "bbox": [ 130, 656, 453, 668 ], "spans": [ { "bbox": [ 130, 656, 453, 668 ], "score": 1.0, "content": "meta learning through the lens of generalization. arXiv preprint arXiv:2212.12017, 2022.", "type": "text", "cross_page": true } ], "index": 45, "is_list_end_line": true }, { "bbox": [ 109, 673, 506, 687 ], "spans": [ { "bbox": [ 109, 673, 506, 687 ], "score": 1.0, "content": "[54] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models", "type": "text", "cross_page": true } ], "index": 46, "is_list_start_line": true }, { "bbox": [ 130, 684, 452, 696 ], "spans": [ { "bbox": [ 130, 684, 452, 696 ], "score": 1.0, "content": "know? Transactions of the Association for Computational Linguistics, 8:423–438, 2020.", "type": "text", "cross_page": true } ], "index": 47, "is_list_end_line": true }, { "bbox": [ 110, 702, 506, 713 ], "spans": [ { "bbox": [ 110, 702, 506, 713 ], "score": 1.0, "content": "[55] Siddharth Karamcheti, Megha Srivastava, Percy Liang, and Dorsa Sadigh. Lila: Language-informed", "type": "text", "cross_page": true } ], "index": 48, "is_list_start_line": true }, { "bbox": [ 130, 712, 309, 723 ], "spans": [ { "bbox": [ 130, 712, 309, 723 ], "score": 1.0, "content": "latent actions. In CoRL, pages 1379–1390, 2021.", "type": "text", "cross_page": true } ], "index": 49, "is_list_end_line": true }, { "bbox": [ 108, 71, 507, 88 ], "spans": [ { "bbox": [ 108, 71, 507, 88 ], "score": 1.0, "content": "[56] Zachary Kenton, Tom Everitt, Laura Weidinger, Iason Gabriel, Vladimir Mikulik, and Geoffrey Irving.", "type": "text", "cross_page": true } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 129, 83, 391, 97 ], "spans": [ { "bbox": [ 129, 83, 391, 97 ], "score": 1.0, "content": "Alignment of language agents. arXiv preprint arXiv:2103.14659, 2021.", "type": "text", "cross_page": true } ], "index": 1, "is_list_end_line": true }, { "bbox": [ 110, 102, 505, 114 ], "spans": [ { "bbox": [ 110, 102, 505, 114 ], "score": 1.0, "content": "[57] Hyunwoo Kim, Jack Hessel, Liwei Jiang, Ximing Lu, Youngjae Yu, Pei Zhou, Ronan Le Bras, Mal-", "type": "text", "cross_page": true } ], "index": 2, "is_list_start_line": true }, { "bbox": [ 128, 110, 506, 125 ], "spans": [ { "bbox": [ 128, 110, 506, 125 ], "score": 1.0, "content": "ihe Alikhani, Gunhee Kim, Maarten Sap, et al. Soda: Million-scale dialogue distillation with social", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 128, 120, 397, 134 ], "spans": [ { "bbox": [ 128, 120, 397, 134 ], "score": 1.0, "content": "commonsense contextualization. arXiv preprint arXiv:2212.10465, 2022.", "type": "text", "cross_page": true } ], "index": 4, "is_list_end_line": true }, { "bbox": [ 110, 139, 505, 151 ], "spans": [ { "bbox": [ 110, 139, 505, 151 ], "score": 1.0, "content": "[58] Jangho Kim, Seonguk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via", "type": "text", "cross_page": true } ], "index": 5, "is_list_start_line": true }, { "bbox": [ 130, 150, 506, 162 ], "spans": [ { "bbox": [ 130, 150, 506, 162 ], "score": 1.0, "content": "factor transfer. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett,", "type": "text", "cross_page": true } ], "index": 6 }, { "bbox": [ 130, 160, 505, 172 ], "spans": [ { "bbox": [ 130, 160, 505, 172 ], "score": 1.0, "content": "editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018.", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 110, 178, 506, 191 ], "spans": [ { "bbox": [ 110, 178, 506, 191 ], "score": 1.0, "content": "[59] Yekyung Kim, Seohyeong Jeong, and Kyunghyun Cho. Linda: Unsupervised learning to interpolate in", "type": "text", "cross_page": true } ], "index": 8, "is_list_start_line": true }, { "bbox": [ 129, 187, 384, 202 ], "spans": [ { "bbox": [ 129, 187, 384, 202 ], "score": 1.0, "content": "natural language processing. arXiv preprint arXiv:2112.13969, 2021.", "type": "text", "cross_page": true } ], "index": 9, "is_list_end_line": true }, { "bbox": [ 110, 207, 506, 218 ], "spans": [ { "bbox": [ 110, 207, 506, 218 ], "score": 1.0, "content": "[60] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In", "type": "text", "cross_page": true } ], "index": 10, "is_list_start_line": true }, { "bbox": [ 129, 216, 421, 231 ], "spans": [ { "bbox": [ 129, 216, 421, 231 ], "score": 1.0, "content": "International conference on machine learning, pages 1885–1894. PMLR, 2017.", "type": "text", "cross_page": true } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 109, 233, 506, 248 ], "spans": [ { "bbox": [ 109, 233, 506, 248 ], "score": 1.0, "content": "[61] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language", "type": "text", "cross_page": true } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 130, 245, 393, 257 ], "spans": [ { "bbox": [ 130, 245, 393, 257 ], "score": 1.0, "content": "models are zero-shot reasoners. arXiv preprint arXiv:2205.11916, 2022.", "type": "text", "cross_page": true } ], "index": 13, "is_list_end_line": true }, { "bbox": [ 110, 263, 505, 275 ], "spans": [ { "bbox": [ 110, 263, 505, 275 ], "score": 1.0, "content": "[62] Jonáš Kulhánek, Vojtech Hude ˇ cek, Tomá ˇ š Nekvinda, and Ondˇrej Dušek. Augpt: Auxiliary tasks and", "type": "text", "cross_page": true } ], "index": 14, "is_list_start_line": true }, { "bbox": [ 129, 271, 505, 286 ], "spans": [ { "bbox": [ 129, 271, 505, 286 ], "score": 1.0, "content": "data augmentation for end-to-end dialogue with pre-trained language models. In Proceedings of the 3rd", "type": "text", "cross_page": true } ], "index": 15 }, { "bbox": [ 128, 282, 453, 295 ], "spans": [ { "bbox": [ 128, 282, 453, 295 ], "score": 1.0, "content": "Workshop on Natural Language Processing for Conversational AI, pages 198–210, 2021.", "type": "text", "cross_page": true } ], "index": 16, "is_list_end_line": true }, { "bbox": [ 110, 300, 507, 314 ], "spans": [ { "bbox": [ 110, 300, 507, 314 ], "score": 1.0, "content": "[63] Kenton Lee, Kelvin Guu, Luheng He, Tim Dozat, and Hyung Won Chung. Neural data augmentation via", "type": "text", "cross_page": true } ], "index": 17, "is_list_start_line": true }, { "bbox": [ 128, 311, 364, 324 ], "spans": [ { "bbox": [ 128, 311, 364, 324 ], "score": 1.0, "content": "example extrapolation. arXiv preprint arXiv:2102.01335, 2021.", "type": "text", "cross_page": true } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 109, 329, 507, 342 ], "spans": [ { "bbox": [ 109, 329, 507, 342 ], "score": 1.0, "content": "[64] Aitor Lewkowycz, Anders Johan Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski,", "type": "text", "cross_page": true } ], "index": 19, "is_list_start_line": true }, { "bbox": [ 130, 338, 507, 353 ], "spans": [ { "bbox": [ 130, 338, 507, 353 ], "score": 1.0, "content": "Vinay Venkatesh Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, et al.", "type": "text", "cross_page": true } ], "index": 20 }, { "bbox": [ 130, 350, 385, 362 ], "spans": [ { "bbox": [ 130, 350, 385, 362 ], "score": 1.0, "content": "Solving quantitative reasoning problems with language models. 2022.", "type": "text", "cross_page": true } ], "index": 21, "is_list_end_line": true }, { "bbox": [ 109, 366, 507, 381 ], "spans": [ { "bbox": [ 109, 366, 507, 381 ], "score": 1.0, "content": "[65] Shuang Li, Xavier Puig, Chris Paxton, Yilun Du, Clinton Wang, Linxi Fan, Tao Chen, De-An Huang,", "type": "text", "cross_page": true } ], "index": 22, "is_list_start_line": true }, { "bbox": [ 131, 378, 506, 390 ], "spans": [ { "bbox": [ 131, 378, 506, 390 ], "score": 1.0, "content": "Ekin Akyürek, Anima Anandkumar, Jacob Andreas, Igor Mordatch, Antonio Torralba, and Yuke Zhu.", "type": "text", "cross_page": true } ], "index": 23 }, { "bbox": [ 129, 387, 375, 401 ], "spans": [ { "bbox": [ 129, 387, 375, 401 ], "score": 1.0, "content": "Pre-trained language models for interactive decision-making, 2022.", "type": "text", "cross_page": true } ], "index": 24, "is_list_end_line": true }, { "bbox": [ 110, 406, 505, 418 ], "spans": [ { "bbox": [ 110, 406, 505, 418 ], "score": 1.0, "content": "[66] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv", "type": "text", "cross_page": true } ], "index": 25, "is_list_start_line": true }, { "bbox": [ 128, 415, 257, 428 ], "spans": [ { "bbox": [ 128, 415, 257, 428 ], "score": 1.0, "content": "preprint arXiv:2101.00190, 2021.", "type": "text", "cross_page": true } ], "index": 26, "is_list_end_line": true }, { "bbox": [ 110, 433, 506, 447 ], "spans": [ { "bbox": [ 110, 433, 506, 447 ], "score": 1.0, "content": "[67] Zekun Li, Wenhu Chen, Shiyang Li, Hong Wang, Jing Qian, and Xifeng Yan. Controllable dialogue", "type": "text", "cross_page": true } ], "index": 27, "is_list_start_line": true }, { "bbox": [ 130, 445, 408, 457 ], "spans": [ { "bbox": [ 130, 445, 408, 457 ], "score": 1.0, "content": "simulation with in-context learning. arXiv preprint arXiv:2210.04185, 2022.", "type": "text", "cross_page": true } ], "index": 28, "is_list_end_line": true }, { "bbox": [ 111, 464, 505, 473 ], "spans": [ { "bbox": [ 111, 464, 505, 473 ], "score": 1.0, "content": "[68] Alisa Liu, Swabha Swayamdipta, Noah A. Smith, and Yejin Choi. WANLI: Worker and AI collaboration", "type": "text", "cross_page": true } ], "index": 29, "is_list_start_line": true }, { "bbox": [ 131, 473, 506, 485 ], "spans": [ { "bbox": [ 131, 473, 506, 485 ], "score": 1.0, "content": "for natural language inference dataset creation. In Findings of the Association for Computational", "type": "text", "cross_page": true } ], "index": 30 }, { "bbox": [ 131, 482, 506, 495 ], "spans": [ { "bbox": [ 131, 482, 506, 495 ], "score": 1.0, "content": "Linguistics: EMNLP 2022, pages 6826–6847, Abu Dhabi, United Arab Emirates, December 2022.", "type": "text", "cross_page": true } ], "index": 31 }, { "bbox": [ 130, 494, 289, 504 ], "spans": [ { "bbox": [ 130, 494, 289, 504 ], "score": 1.0, "content": "Association for Computational Linguistics.", "type": "text", "cross_page": true } ], "index": 32, "is_list_end_line": true }, { "bbox": [ 110, 511, 505, 523 ], "spans": [ { "bbox": [ 110, 511, 505, 523 ], "score": 1.0, "content": "[69] Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. Is your code generated by chatgpt", "type": "text", "cross_page": true } ], "index": 33, "is_list_start_line": true }, { "bbox": [ 129, 519, 506, 534 ], "spans": [ { "bbox": [ 129, 519, 506, 534 ], "score": 1.0, "content": "really correct? rigorous evaluation of large language models for code generation. arXiv preprint", "type": "text", "cross_page": true } ], "index": 34 }, { "bbox": [ 131, 532, 223, 541 ], "spans": [ { "bbox": [ 131, 532, 223, 541 ], "score": 1.0, "content": "arXiv:2305.01210, 2023.", "type": "text", "cross_page": true } ], "index": 35, "is_list_end_line": true }, { "bbox": [ 110, 549, 506, 561 ], "spans": [ { "bbox": [ 110, 549, 506, 561 ], "score": 1.0, "content": "[70] Yat Long Lo, Christian Schroeder de Witt, Samuel Sokota, Jakob Nicolaus Foerster, and Shimon Whiteson.", "type": "text", "cross_page": true } ], "index": 36, "is_list_start_line": true }, { "bbox": [ 131, 559, 505, 570 ], "spans": [ { "bbox": [ 131, 559, 505, 570 ], "score": 1.0, "content": "Cheap talk discovery and utilization in multi-agent reinforcement learning. In The Eleventh International", "type": "text", "cross_page": true } ], "index": 37 }, { "bbox": [ 129, 568, 306, 582 ], "spans": [ { "bbox": [ 129, 568, 306, 582 ], "score": 1.0, "content": "Conference on Learning Representations, 2023.", "type": "text", "cross_page": true } ], "index": 38, "is_list_end_line": true }, { "bbox": [ 110, 586, 507, 600 ], "spans": [ { "bbox": [ 110, 586, 507, 600 ], "score": 1.0, "content": "[71] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le,", "type": "text", "cross_page": true } ], "index": 39, "is_list_start_line": true }, { "bbox": [ 129, 596, 507, 611 ], "spans": [ { "bbox": [ 129, 596, 507, 611 ], "score": 1.0, "content": "Barret Zoph, Jason Wei, et al. The flan collection: Designing data and methods for effective instruction", "type": "text", "cross_page": true } ], "index": 40 }, { "bbox": [ 130, 607, 305, 619 ], "spans": [ { "bbox": [ 130, 607, 305, 619 ], "score": 1.0, "content": "tuning. arXiv preprint arXiv:2301.13688, 2023.", "type": "text", "cross_page": true } ], "index": 41, "is_list_end_line": true }, { "bbox": [ 109, 624, 507, 638 ], "spans": [ { "bbox": [ 109, 624, 507, 638 ], "score": 1.0, "content": "[72] Ryan Lowe, Yi I Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent", "type": "text", "cross_page": true } ], "index": 42, "is_list_start_line": true }, { "bbox": [ 128, 634, 507, 649 ], "spans": [ { "bbox": [ 128, 634, 507, 649 ], "score": 1.0, "content": "actor-critic for mixed cooperative-competitive environments. Advances in neural information processing", "type": "text", "cross_page": true } ], "index": 43 }, { "bbox": [ 130, 645, 199, 657 ], "spans": [ { "bbox": [ 130, 645, 199, 657 ], "score": 1.0, "content": "systems, 30, 2017.", "type": "text", "cross_page": true } ], "index": 44, "is_list_end_line": true }, { "bbox": [ 110, 663, 506, 676 ], "spans": [ { "bbox": [ 110, 663, 506, 676 ], "score": 1.0, "content": "[73] Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark,", "type": "text", "cross_page": true } ], "index": 45, "is_list_start_line": true }, { "bbox": [ 128, 672, 506, 687 ], "spans": [ { "bbox": [ 128, 672, 506, 687 ], "score": 1.0, "content": "and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical", "type": "text", "cross_page": true } ], "index": 46 }, { "bbox": [ 129, 682, 229, 696 ], "spans": [ { "bbox": [ 129, 682, 229, 696 ], "score": 1.0, "content": "reasoning. In ICLR, 2023.", "type": "text", "cross_page": true } ], "index": 47, "is_list_end_line": true }, { "bbox": [ 109, 700, 506, 714 ], "spans": [ { "bbox": [ 109, 700, 506, 714 ], "score": 1.0, "content": "[74] Michael J. Matthews, Samuel H. Matthews, and Thomas K. Kelemen. The alignment problem: Machine", "type": "text", "cross_page": true } ], "index": 48, "is_list_start_line": true }, { "bbox": [ 131, 712, 339, 724 ], "spans": [ { "bbox": [ 131, 712, 339, 724 ], "score": 1.0, "content": "learning and human values. Personnel Psychology, 2022.", "type": "text", "cross_page": true } ], "index": 49, "is_list_end_line": true } ], "index": 25.5, "bbox_fs": [ 108, 73, 507, 724 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 109, 61, 507, 728 ], "lines": [ { "bbox": [ 110, 73, 506, 85 ], "spans": [ { "bbox": [ 110, 73, 506, 85 ], "score": 1.0, "content": "[36] Amelia Glaese, Nat McAleese, Maja Tr˛ebacz, John Aslanides, Vlad Firoiu, Timo Ewalds, Maribeth Rauh,", "type": "text" } ], "index": 0 }, { "bbox": [ 128, 82, 507, 96 ], "spans": [ { "bbox": [ 128, 82, 507, 96 ], "score": 1.0, "content": "Laura Weidinger, Martin Chadwick, Phoebe Thacker, et al. Improving alignment of dialogue agents via", "type": "text" } ], "index": 1 }, { "bbox": [ 128, 92, 383, 106 ], "spans": [ { "bbox": [ 128, 92, 383, 106 ], "score": 1.0, "content": "targeted human judgements. arXiv preprint arXiv:2209.14375, 2022.", "type": "text" } ], "index": 2 }, { "bbox": [ 111, 111, 506, 123 ], "spans": [ { "bbox": [ 111, 111, 506, 123 ], "score": 1.0, "content": "[37] Josh A Goldstein, Girish Sastry, Micah Musser, Renee DiResta, Matthew Gentzel, and Katerina Se-", "type": "text" } ], "index": 3 }, { "bbox": [ 129, 120, 506, 133 ], "spans": [ { "bbox": [ 129, 120, 506, 133 ], "score": 1.0, "content": "dova. Generative language models and automated influence operations: Emerging threats and potential", "type": "text" } ], "index": 4 }, { "bbox": [ 129, 131, 323, 142 ], "spans": [ { "bbox": [ 129, 131, 323, 142 ], "score": 1.0, "content": "mitigations. arXiv preprint arXiv:2301.04246, 2023.", "type": "text" } ], "index": 5 }, { "bbox": [ 110, 149, 506, 161 ], "spans": [ { "bbox": [ 110, 149, 506, 161 ], "score": 1.0, "content": "[38] Dylan Hadfield-Menell. The principal-agent alignment problem in artificial intelligence. Ph. D. disserta-", "type": "text" } ], "index": 6 }, { "bbox": [ 129, 159, 173, 171 ], "spans": [ { "bbox": [ 129, 159, 173, 171 ], "score": 1.0, "content": "tion, 2021.", "type": "text" } ], "index": 7 }, { "bbox": [ 109, 176, 507, 190 ], "spans": [ { "bbox": [ 109, 176, 507, 190 ], "score": 1.0, "content": "[39] Dylan Hadfield-Menell, McKane Andrus, and Gillian Hadfield. Legible normativity for ai alignment:", "type": "text" } ], "index": 8 }, { "bbox": [ 129, 185, 507, 200 ], "spans": [ { "bbox": [ 129, 185, 507, 200 ], "score": 1.0, "content": "The value of silly rules. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society,", "type": "text" } ], "index": 9 }, { "bbox": [ 128, 197, 213, 208 ], "spans": [ { "bbox": [ 128, 197, 213, 208 ], "score": 1.0, "content": "pages 115–121, 2019.", "type": "text" } ], "index": 10 }, { "bbox": [ 110, 215, 506, 227 ], "spans": [ { "bbox": [ 110, 215, 506, 227 ], "score": 1.0, "content": "[40] Dylan Hadfield-Menell, Stuart J Russell, Pieter Abbeel, and Anca Dragan. Cooperative inverse reinforce-", "type": "text" } ], "index": 11 }, { "bbox": [ 128, 224, 412, 237 ], "spans": [ { "bbox": [ 128, 224, 412, 237 ], "score": 1.0, "content": "ment learning. Advances in neural information processing systems, 29, 2016.", "type": "text" } ], "index": 12 }, { "bbox": [ 110, 243, 506, 254 ], "spans": [ { "bbox": [ 110, 243, 506, 254 ], "score": 1.0, "content": "[41] Serhii Havrylov and Ivan Titov. Emergence of language with multi-agent games: Learning to communicate", "type": "text" } ], "index": 13 }, { "bbox": [ 130, 253, 457, 265 ], "spans": [ { "bbox": [ 130, 253, 457, 265 ], "score": 1.0, "content": "with sequences of symbols. Advances in neural information processing systems, 30, 2017.", "type": "text" } ], "index": 14 }, { "bbox": [ 109, 269, 506, 282 ], "spans": [ { "bbox": [ 109, 269, 506, 282 ], "score": 1.0, "content": "[42] Peter Henderson, Koustuv Sinha, Nicolas Angelard-Gontier, Nan Rosemary Ke, Genevieve Fried, Ryan", "type": "text" } ], "index": 15 }, { "bbox": [ 128, 279, 506, 293 ], "spans": [ { "bbox": [ 128, 279, 506, 293 ], "score": 1.0, "content": "Lowe, and Joelle Pineau. Ethical challenges in data-driven dialogue systems. In Proceedings of the 2018", "type": "text" } ], "index": 16 }, { "bbox": [ 128, 290, 396, 304 ], "spans": [ { "bbox": [ 128, 290, 396, 304 ], "score": 1.0, "content": "AAAI/ACM Conference on AI, Ethics, and Society, pages 123–129, 2018.", "type": "text" } ], "index": 17 }, { "bbox": [ 108, 305, 508, 323 ], "spans": [ { "bbox": [ 108, 305, 508, 323 ], "score": 1.0, "content": "[43] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song,", "type": "text" } ], "index": 18 }, { "bbox": [ 128, 317, 506, 331 ], "spans": [ { "bbox": [ 128, 317, 506, 331 ], "score": 1.0, "content": "and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint", "type": "text" } ], "index": 19 }, { "bbox": [ 129, 329, 225, 339 ], "spans": [ { "bbox": [ 129, 329, 225, 339 ], "score": 1.0, "content": "arXiv:2103.03874, 2021.", "type": "text" } ], "index": 20 }, { "bbox": [ 110, 346, 506, 358 ], "spans": [ { "bbox": [ 110, 346, 506, 358 ], "score": 1.0, "content": "[44] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of", "type": "text" } ], "index": 21 }, { "bbox": [ 130, 356, 506, 368 ], "spans": [ { "bbox": [ 130, 356, 506, 368 ], "score": 1.0, "content": "activation boundaries formed by hidden neurons. In Proceedings of the AAAI Conference on Artificial", "type": "text" } ], "index": 22 }, { "bbox": [ 130, 366, 310, 378 ], "spans": [ { "bbox": [ 130, 366, 310, 378 ], "score": 1.0, "content": "Intelligence, volume 33, pages 3779–3787, 2019.", "type": "text" } ], "index": 23 }, { "bbox": [ 110, 383, 506, 395 ], "spans": [ { "bbox": [ 110, 383, 506, 395 ], "score": 1.0, "content": "[45] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv", "type": "text" } ], "index": 24 }, { "bbox": [ 128, 393, 257, 406 ], "spans": [ { "bbox": [ 128, 393, 257, 406 ], "score": 1.0, "content": "preprint arXiv:1503.02531, 2015.", "type": "text" } ], "index": 25 }, { "bbox": [ 110, 411, 506, 423 ], "spans": [ { "bbox": [ 110, 411, 506, 423 ], "score": 1.0, "content": "[46] Namgyu Ho, Laura Schmid, and Se-Young Yun. Large language models are reasoning teachers. arXiv", "type": "text" } ], "index": 26 }, { "bbox": [ 128, 421, 256, 434 ], "spans": [ { "bbox": [ 128, 421, 256, 434 ], "score": 1.0, "content": "preprint arXiv:2212.10071, 2022.", "type": "text" } ], "index": 27 }, { "bbox": [ 110, 439, 506, 451 ], "spans": [ { "bbox": [ 110, 439, 506, 451 ], "score": 1.0, "content": "[47] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford,", "type": "text" } ], "index": 28 }, { "bbox": [ 131, 449, 505, 460 ], "spans": [ { "bbox": [ 131, 449, 505, 460 ], "score": 1.0, "content": "Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal", "type": "text" } ], "index": 29 }, { "bbox": [ 129, 458, 363, 471 ], "spans": [ { "bbox": [ 129, 458, 363, 471 ], "score": 1.0, "content": "large language models. arXiv preprint arXiv:2203.15556, 2022.", "type": "text" } ], "index": 30 }, { "bbox": [ 109, 475, 506, 491 ], "spans": [ { "bbox": [ 109, 475, 506, 491 ], "score": 1.0, "content": "[48] Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. Unnatural instructions: Tuning language", "type": "text" } ], "index": 31 }, { "bbox": [ 128, 486, 416, 499 ], "spans": [ { "bbox": [ 128, 486, 416, 499 ], "score": 1.0, "content": "models with (almost) no human labor. arXiv preprint arXiv:2212.09689, 2022.", "type": "text" } ], "index": 32 }, { "bbox": [ 110, 505, 506, 517 ], "spans": [ { "bbox": [ 110, 505, 506, 517 ], "score": 1.0, "content": "[49] Ehsan Hosseini-Asl, Bryan McCann, Chien-Sheng Wu, Semih Yavuz, and Richard Socher. A sim-", "type": "text" } ], "index": 33 }, { "bbox": [ 130, 514, 507, 528 ], "spans": [ { "bbox": [ 130, 514, 507, 528 ], "score": 1.0, "content": "ple language model for task-oriented dialogue. Advances in Neural Information Processing Systems,", "type": "text" } ], "index": 34 }, { "bbox": [ 131, 525, 219, 537 ], "spans": [ { "bbox": [ 131, 525, 219, 537 ], "score": 1.0, "content": "33:20179–20191, 2020.", "type": "text" } ], "index": 35 }, { "bbox": [ 109, 542, 506, 556 ], "spans": [ { "bbox": [ 109, 542, 506, 556 ], "score": 1.0, "content": "[50] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot", "type": "text" } ], "index": 36 }, { "bbox": [ 128, 552, 507, 565 ], "spans": [ { "bbox": [ 128, 552, 507, 565 ], "score": 1.0, "content": "planners: Extracting actionable knowledge for embodied agents. arXiv preprint arXiv:2201.07207, 2022.", "type": "text" } ], "index": 37 }, { "bbox": [ 109, 569, 506, 583 ], "spans": [ { "bbox": [ 109, 569, 506, 583 ], "score": 1.0, "content": "[51] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan", "type": "text" } ], "index": 38 }, { "bbox": [ 128, 579, 506, 594 ], "spans": [ { "bbox": [ 128, 579, 506, 594 ], "score": 1.0, "content": "Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through", "type": "text" } ], "index": 39 }, { "bbox": [ 129, 590, 394, 602 ], "spans": [ { "bbox": [ 129, 590, 394, 602 ], "score": 1.0, "content": "planning with language models. arXiv preprint arXiv:2207.05608, 2022.", "type": "text" } ], "index": 40 }, { "bbox": [ 109, 607, 506, 622 ], "spans": [ { "bbox": [ 109, 607, 506, 622 ], "score": 1.0, "content": "[52] Shima Imani, Liang Du, and Harsh Shrivastava. Mathprompter: Mathematical reasoning using large", "type": "text" } ], "index": 41 }, { "bbox": [ 128, 618, 344, 630 ], "spans": [ { "bbox": [ 128, 618, 344, 630 ], "score": 1.0, "content": "language models. arXiv preprint arXiv:2303.05398, 2023.", "type": "text" } ], "index": 42 }, { "bbox": [ 110, 636, 506, 648 ], "spans": [ { "bbox": [ 110, 636, 506, 648 ], "score": 1.0, "content": "[53] Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru, Todor Mihaylov, Dániel Simig, Ping Yu, Kurt", "type": "text" } ], "index": 43 }, { "bbox": [ 130, 646, 506, 658 ], "spans": [ { "bbox": [ 130, 646, 506, 658 ], "score": 1.0, "content": "Shuster, Tianlu Wang, Qing Liu, Punit Singh Koura, et al. Opt-iml: Scaling language model instruction", "type": "text" } ], "index": 44 }, { "bbox": [ 130, 656, 453, 668 ], "spans": [ { "bbox": [ 130, 656, 453, 668 ], "score": 1.0, "content": "meta learning through the lens of generalization. arXiv preprint arXiv:2212.12017, 2022.", "type": "text" } ], "index": 45 }, { "bbox": [ 109, 673, 506, 687 ], "spans": [ { "bbox": [ 109, 673, 506, 687 ], "score": 1.0, "content": "[54] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models", "type": "text" } ], "index": 46 }, { "bbox": [ 130, 684, 452, 696 ], "spans": [ { "bbox": [ 130, 684, 452, 696 ], "score": 1.0, "content": "know? Transactions of the Association for Computational Linguistics, 8:423–438, 2020.", "type": "text" } ], "index": 47 }, { "bbox": [ 110, 702, 506, 713 ], "spans": [ { "bbox": [ 110, 702, 506, 713 ], "score": 1.0, "content": "[55] Siddharth Karamcheti, Megha Srivastava, Percy Liang, and Dorsa Sadigh. Lila: Language-informed", "type": "text" } ], "index": 48 }, { "bbox": [ 130, 712, 309, 723 ], "spans": [ { "bbox": [ 130, 712, 309, 723 ], "score": 1.0, "content": "latent actions. In CoRL, pages 1379–1390, 2021.", "type": "text" } ], "index": 49 } ], "index": 24.5 } ], "page_idx": 12, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 313, 754 ], "spans": [ { "bbox": [ 299, 740, 313, 754 ], "score": 1.0, "content": "13", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 109, 61, 507, 728 ], "lines": [], "index": 24.5, "bbox_fs": [ 108, 73, 508, 723 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 108, 36, 507, 724 ], "lines": [ { "bbox": [ 108, 71, 507, 88 ], "spans": [ { "bbox": [ 108, 71, 507, 88 ], "score": 1.0, "content": "[56] Zachary Kenton, Tom Everitt, Laura Weidinger, Iason Gabriel, Vladimir Mikulik, and Geoffrey Irving.", "type": "text" } ], "index": 0 }, { "bbox": [ 129, 83, 391, 97 ], "spans": [ { "bbox": [ 129, 83, 391, 97 ], "score": 1.0, "content": "Alignment of language agents. arXiv preprint arXiv:2103.14659, 2021.", "type": "text" } ], "index": 1 }, { "bbox": [ 110, 102, 505, 114 ], "spans": [ { "bbox": [ 110, 102, 505, 114 ], "score": 1.0, "content": "[57] Hyunwoo Kim, Jack Hessel, Liwei Jiang, Ximing Lu, Youngjae Yu, Pei Zhou, Ronan Le Bras, Mal-", "type": "text" } ], "index": 2 }, { "bbox": [ 128, 110, 506, 125 ], "spans": [ { "bbox": [ 128, 110, 506, 125 ], "score": 1.0, "content": "ihe Alikhani, Gunhee Kim, Maarten Sap, et al. Soda: Million-scale dialogue distillation with social", "type": "text" } ], "index": 3 }, { "bbox": [ 128, 120, 397, 134 ], "spans": [ { "bbox": [ 128, 120, 397, 134 ], "score": 1.0, "content": "commonsense contextualization. arXiv preprint arXiv:2212.10465, 2022.", "type": "text" } ], "index": 4 }, { "bbox": [ 110, 139, 505, 151 ], "spans": [ { "bbox": [ 110, 139, 505, 151 ], "score": 1.0, "content": "[58] Jangho Kim, Seonguk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via", "type": "text" } ], "index": 5 }, { "bbox": [ 130, 150, 506, 162 ], "spans": [ { "bbox": [ 130, 150, 506, 162 ], "score": 1.0, "content": "factor transfer. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett,", "type": "text" } ], "index": 6 }, { "bbox": [ 130, 160, 505, 172 ], "spans": [ { "bbox": [ 130, 160, 505, 172 ], "score": 1.0, "content": "editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018.", "type": "text" } ], "index": 7 }, { "bbox": [ 110, 178, 506, 191 ], "spans": [ { "bbox": [ 110, 178, 506, 191 ], "score": 1.0, "content": "[59] Yekyung Kim, Seohyeong Jeong, and Kyunghyun Cho. Linda: Unsupervised learning to interpolate in", "type": "text" } ], "index": 8 }, { "bbox": [ 129, 187, 384, 202 ], "spans": [ { "bbox": [ 129, 187, 384, 202 ], "score": 1.0, "content": "natural language processing. arXiv preprint arXiv:2112.13969, 2021.", "type": "text" } ], "index": 9 }, { "bbox": [ 110, 207, 506, 218 ], "spans": [ { "bbox": [ 110, 207, 506, 218 ], "score": 1.0, "content": "[60] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In", "type": "text" } ], "index": 10 }, { "bbox": [ 129, 216, 421, 231 ], "spans": [ { "bbox": [ 129, 216, 421, 231 ], "score": 1.0, "content": "International conference on machine learning, pages 1885–1894. PMLR, 2017.", "type": "text" } ], "index": 11 }, { "bbox": [ 109, 233, 506, 248 ], "spans": [ { "bbox": [ 109, 233, 506, 248 ], "score": 1.0, "content": "[61] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language", "type": "text" } ], "index": 12 }, { "bbox": [ 130, 245, 393, 257 ], "spans": [ { "bbox": [ 130, 245, 393, 257 ], "score": 1.0, "content": "models are zero-shot reasoners. arXiv preprint arXiv:2205.11916, 2022.", "type": "text" } ], "index": 13 }, { "bbox": [ 110, 263, 505, 275 ], "spans": [ { "bbox": [ 110, 263, 505, 275 ], "score": 1.0, "content": "[62] Jonáš Kulhánek, Vojtech Hude ˇ cek, Tomá ˇ š Nekvinda, and Ondˇrej Dušek. Augpt: Auxiliary tasks and", "type": "text" } ], "index": 14 }, { "bbox": [ 129, 271, 505, 286 ], "spans": [ { "bbox": [ 129, 271, 505, 286 ], "score": 1.0, "content": "data augmentation for end-to-end dialogue with pre-trained language models. In Proceedings of the 3rd", "type": "text" } ], "index": 15 }, { "bbox": [ 128, 282, 453, 295 ], "spans": [ { "bbox": [ 128, 282, 453, 295 ], "score": 1.0, "content": "Workshop on Natural Language Processing for Conversational AI, pages 198–210, 2021.", "type": "text" } ], "index": 16 }, { "bbox": [ 110, 300, 507, 314 ], "spans": [ { "bbox": [ 110, 300, 507, 314 ], "score": 1.0, "content": "[63] Kenton Lee, Kelvin Guu, Luheng He, Tim Dozat, and Hyung Won Chung. Neural data augmentation via", "type": "text" } ], "index": 17 }, { "bbox": [ 128, 311, 364, 324 ], "spans": [ { "bbox": [ 128, 311, 364, 324 ], "score": 1.0, "content": "example extrapolation. arXiv preprint arXiv:2102.01335, 2021.", "type": "text" } ], "index": 18 }, { "bbox": [ 109, 329, 507, 342 ], "spans": [ { "bbox": [ 109, 329, 507, 342 ], "score": 1.0, "content": "[64] Aitor Lewkowycz, Anders Johan Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski,", "type": "text" } ], "index": 19 }, { "bbox": [ 130, 338, 507, 353 ], "spans": [ { "bbox": [ 130, 338, 507, 353 ], "score": 1.0, "content": "Vinay Venkatesh Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, et al.", "type": "text" } ], "index": 20 }, { "bbox": [ 130, 350, 385, 362 ], "spans": [ { "bbox": [ 130, 350, 385, 362 ], "score": 1.0, "content": "Solving quantitative reasoning problems with language models. 2022.", "type": "text" } ], "index": 21 }, { "bbox": [ 109, 366, 507, 381 ], "spans": [ { "bbox": [ 109, 366, 507, 381 ], "score": 1.0, "content": "[65] Shuang Li, Xavier Puig, Chris Paxton, Yilun Du, Clinton Wang, Linxi Fan, Tao Chen, De-An Huang,", "type": "text" } ], "index": 22 }, { "bbox": [ 131, 378, 506, 390 ], "spans": [ { "bbox": [ 131, 378, 506, 390 ], "score": 1.0, "content": "Ekin Akyürek, Anima Anandkumar, Jacob Andreas, Igor Mordatch, Antonio Torralba, and Yuke Zhu.", "type": "text" } ], "index": 23 }, { "bbox": [ 129, 387, 375, 401 ], "spans": [ { "bbox": [ 129, 387, 375, 401 ], "score": 1.0, "content": "Pre-trained language models for interactive decision-making, 2022.", "type": "text" } ], "index": 24 }, { "bbox": [ 110, 406, 505, 418 ], "spans": [ { "bbox": [ 110, 406, 505, 418 ], "score": 1.0, "content": "[66] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv", "type": "text" } ], "index": 25 }, { "bbox": [ 128, 415, 257, 428 ], "spans": [ { "bbox": [ 128, 415, 257, 428 ], "score": 1.0, "content": "preprint arXiv:2101.00190, 2021.", "type": "text" } ], "index": 26 }, { "bbox": [ 110, 433, 506, 447 ], "spans": [ { "bbox": [ 110, 433, 506, 447 ], "score": 1.0, "content": "[67] Zekun Li, Wenhu Chen, Shiyang Li, Hong Wang, Jing Qian, and Xifeng Yan. Controllable dialogue", "type": "text" } ], "index": 27 }, { "bbox": [ 130, 445, 408, 457 ], "spans": [ { "bbox": [ 130, 445, 408, 457 ], "score": 1.0, "content": "simulation with in-context learning. arXiv preprint arXiv:2210.04185, 2022.", "type": "text" } ], "index": 28 }, { "bbox": [ 111, 464, 505, 473 ], "spans": [ { "bbox": [ 111, 464, 505, 473 ], "score": 1.0, "content": "[68] Alisa Liu, Swabha Swayamdipta, Noah A. Smith, and Yejin Choi. WANLI: Worker and AI collaboration", "type": "text" } ], "index": 29 }, { "bbox": [ 131, 473, 506, 485 ], "spans": [ { "bbox": [ 131, 473, 506, 485 ], "score": 1.0, "content": "for natural language inference dataset creation. In Findings of the Association for Computational", "type": "text" } ], "index": 30 }, { "bbox": [ 131, 482, 506, 495 ], "spans": [ { "bbox": [ 131, 482, 506, 495 ], "score": 1.0, "content": "Linguistics: EMNLP 2022, pages 6826–6847, Abu Dhabi, United Arab Emirates, December 2022.", "type": "text" } ], "index": 31 }, { "bbox": [ 130, 494, 289, 504 ], "spans": [ { "bbox": [ 130, 494, 289, 504 ], "score": 1.0, "content": "Association for Computational Linguistics.", "type": "text" } ], "index": 32 }, { "bbox": [ 110, 511, 505, 523 ], "spans": [ { "bbox": [ 110, 511, 505, 523 ], "score": 1.0, "content": "[69] Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. Is your code generated by chatgpt", "type": "text" } ], "index": 33 }, { "bbox": [ 129, 519, 506, 534 ], "spans": [ { "bbox": [ 129, 519, 506, 534 ], "score": 1.0, "content": "really correct? rigorous evaluation of large language models for code generation. arXiv preprint", "type": "text" } ], "index": 34 }, { "bbox": [ 131, 532, 223, 541 ], "spans": [ { "bbox": [ 131, 532, 223, 541 ], "score": 1.0, "content": "arXiv:2305.01210, 2023.", "type": "text" } ], "index": 35 }, { "bbox": [ 110, 549, 506, 561 ], "spans": [ { "bbox": [ 110, 549, 506, 561 ], "score": 1.0, "content": "[70] Yat Long Lo, Christian Schroeder de Witt, Samuel Sokota, Jakob Nicolaus Foerster, and Shimon Whiteson.", "type": "text" } ], "index": 36 }, { "bbox": [ 131, 559, 505, 570 ], "spans": [ { "bbox": [ 131, 559, 505, 570 ], "score": 1.0, "content": "Cheap talk discovery and utilization in multi-agent reinforcement learning. In The Eleventh International", "type": "text" } ], "index": 37 }, { "bbox": [ 129, 568, 306, 582 ], "spans": [ { "bbox": [ 129, 568, 306, 582 ], "score": 1.0, "content": "Conference on Learning Representations, 2023.", "type": "text" } ], "index": 38 }, { "bbox": [ 110, 586, 507, 600 ], "spans": [ { "bbox": [ 110, 586, 507, 600 ], "score": 1.0, "content": "[71] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le,", "type": "text" } ], "index": 39 }, { "bbox": [ 129, 596, 507, 611 ], "spans": [ { "bbox": [ 129, 596, 507, 611 ], "score": 1.0, "content": "Barret Zoph, Jason Wei, et al. The flan collection: Designing data and methods for effective instruction", "type": "text" } ], "index": 40 }, { "bbox": [ 130, 607, 305, 619 ], "spans": [ { "bbox": [ 130, 607, 305, 619 ], "score": 1.0, "content": "tuning. arXiv preprint arXiv:2301.13688, 2023.", "type": "text" } ], "index": 41 }, { "bbox": [ 109, 624, 507, 638 ], "spans": [ { "bbox": [ 109, 624, 507, 638 ], "score": 1.0, "content": "[72] Ryan Lowe, Yi I Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent", "type": "text" } ], "index": 42 }, { "bbox": [ 128, 634, 507, 649 ], "spans": [ { "bbox": [ 128, 634, 507, 649 ], "score": 1.0, "content": "actor-critic for mixed cooperative-competitive environments. Advances in neural information processing", "type": "text" } ], "index": 43 }, { "bbox": [ 130, 645, 199, 657 ], "spans": [ { "bbox": [ 130, 645, 199, 657 ], "score": 1.0, "content": "systems, 30, 2017.", "type": "text" } ], "index": 44 }, { "bbox": [ 110, 663, 506, 676 ], "spans": [ { "bbox": [ 110, 663, 506, 676 ], "score": 1.0, "content": "[73] Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark,", "type": "text" } ], "index": 45 }, { "bbox": [ 128, 672, 506, 687 ], "spans": [ { "bbox": [ 128, 672, 506, 687 ], "score": 1.0, "content": "and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical", "type": "text" } ], "index": 46 }, { "bbox": [ 129, 682, 229, 696 ], "spans": [ { "bbox": [ 129, 682, 229, 696 ], "score": 1.0, "content": "reasoning. In ICLR, 2023.", "type": "text" } ], "index": 47 }, { "bbox": [ 109, 700, 506, 714 ], "spans": [ { "bbox": [ 109, 700, 506, 714 ], "score": 1.0, "content": "[74] Michael J. Matthews, Samuel H. Matthews, and Thomas K. Kelemen. The alignment problem: Machine", "type": "text" } ], "index": 48 }, { "bbox": [ 131, 712, 339, 724 ], "spans": [ { "bbox": [ 131, 712, 339, 724 ], "score": 1.0, "content": "learning and human values. Personnel Psychology, 2022.", "type": "text" } ], "index": 49 } ], "index": 24.5 } ], "page_idx": 13, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 313, 754 ], "spans": [ { "bbox": [ 299, 740, 313, 754 ], "score": 1.0, "content": "14", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 108, 36, 507, 724 ], "lines": [], "index": 24.5, "bbox_fs": [ 108, 71, 507, 724 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 109, 73, 505, 94 ], "lines": [ { "bbox": [ 111, 74, 506, 84 ], "spans": [ { "bbox": [ 111, 74, 506, 84 ], "score": 1.0, "content": "[75] Yu Meng, Jiaxin Huang, Yu Zhang, and Jiawei Han. Generating training data with language models:", "type": "text" } ], "index": 0 }, { "bbox": [ 131, 83, 506, 95 ], "spans": [ { "bbox": [ 131, 83, 506, 95 ], "score": 1.0, "content": "Towards zero-shot language understanding. In Advances in Neural Information Processing Systems, 2022.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 111, 100, 351, 112 ], "lines": [ { "bbox": [ 110, 100, 351, 113 ], "spans": [ { "bbox": [ 110, 100, 351, 113 ], "score": 1.0, "content": "[76] Marvin Minsky. Society of mind. Simon and Schuster, 1988.", "type": "text" } ], "index": 2 } ], "index": 2 }, { "type": "text", "bbox": [ 110, 118, 504, 139 ], "lines": [ { "bbox": [ 110, 117, 506, 130 ], "spans": [ { "bbox": [ 110, 117, 506, 130 ], "score": 1.0, "content": "[77] Marvin Minsky. The emotion machine: Commonsense thinking, artificial intelligence, and the future of", "type": "text" } ], "index": 3 }, { "bbox": [ 130, 128, 293, 139 ], "spans": [ { "bbox": [ 130, 128, 293, 139 ], "score": 1.0, "content": "the human mind. Simon and Schuster, 2007.", "type": "text" } ], "index": 4 } ], "index": 3.5 }, { "type": "text", "bbox": [ 110, 145, 503, 167 ], "lines": [ { "bbox": [ 110, 145, 506, 158 ], "spans": [ { "bbox": [ 110, 145, 506, 158 ], "score": 1.0, "content": "[78] Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task generalization via", "type": "text" } ], "index": 5 }, { "bbox": [ 130, 156, 348, 167 ], "spans": [ { "bbox": [ 130, 156, 348, 167 ], "score": 1.0, "content": "natural language crowdsourcing instructions. In ACL, 2022.", "type": "text" } ], "index": 6 } ], "index": 5.5 }, { "type": "text", "bbox": [ 109, 173, 504, 195 ], "lines": [ { "bbox": [ 110, 173, 505, 186 ], "spans": [ { "bbox": [ 110, 173, 505, 186 ], "score": 1.0, "content": "[79] Igor Mordatch and Pieter Abbeel. Emergence of grounded compositional language in multi-agent", "type": "text" } ], "index": 7 }, { "bbox": [ 129, 183, 477, 195 ], "spans": [ { "bbox": [ 129, 183, 477, 195 ], "score": 1.0, "content": "populations. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018.", "type": "text" } ], "index": 8 } ], "index": 7.5 }, { "type": "text", "bbox": [ 111, 200, 505, 241 ], "lines": [ { "bbox": [ 111, 201, 506, 213 ], "spans": [ { "bbox": [ 111, 201, 506, 213 ], "score": 1.0, "content": "[80] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse,", "type": "text" } ], "index": 9 }, { "bbox": [ 131, 211, 505, 223 ], "spans": [ { "bbox": [ 131, 211, 505, 223 ], "score": 1.0, "content": "Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen", "type": "text" } ], "index": 10 }, { "bbox": [ 131, 221, 505, 232 ], "spans": [ { "bbox": [ 131, 221, 505, 232 ], "score": 1.0, "content": "Krueger, Kevin Button, Matthew Knight, Benjamin Chess, and John Schulman. Webgpt: Browser-assisted", "type": "text" } ], "index": 11 }, { "bbox": [ 131, 231, 308, 242 ], "spans": [ { "bbox": [ 131, 231, 308, 242 ], "score": 1.0, "content": "question-answering with human feedback, 2021.", "type": "text" } ], "index": 12 } ], "index": 10.5 }, { "type": "text", "bbox": [ 109, 248, 504, 279 ], "lines": [ { "bbox": [ 110, 248, 506, 261 ], "spans": [ { "bbox": [ 110, 248, 506, 261 ], "score": 1.0, "content": "[81] Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber,", "type": "text" } ], "index": 13 }, { "bbox": [ 129, 258, 506, 270 ], "spans": [ { "bbox": [ 129, 258, 506, 270 ], "score": 1.0, "content": "David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, Charles Sutton, and Augustus Odena.", "type": "text" } ], "index": 14 }, { "bbox": [ 130, 268, 454, 280 ], "spans": [ { "bbox": [ 130, 268, 454, 280 ], "score": 1.0, "content": "Show your work: Scratchpads for intermediate computation with language models, 2021.", "type": "text" } ], "index": 15 } ], "index": 14 }, { "type": "text", "bbox": [ 111, 285, 317, 297 ], "lines": [ { "bbox": [ 110, 285, 317, 299 ], "spans": [ { "bbox": [ 110, 285, 317, 299 ], "score": 1.0, "content": "[82] OpenAI. Introducing chatgpt. Open AI Blog, 2022.", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "text", "bbox": [ 112, 303, 300, 315 ], "lines": [ { "bbox": [ 110, 302, 301, 316 ], "spans": [ { "bbox": [ 110, 302, 301, 316 ], "score": 1.0, "content": "[83] OpenAI. Chatgpt plugins. OpenAI blog, 2023.", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 110, 321, 504, 352 ], "lines": [ { "bbox": [ 110, 320, 506, 334 ], "spans": [ { "bbox": [ 110, 320, 506, 334 ], "score": 1.0, "content": "[84] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang,", "type": "text" } ], "index": 18 }, { "bbox": [ 130, 331, 505, 343 ], "spans": [ { "bbox": [ 130, 331, 505, 343 ], "score": 1.0, "content": "Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with", "type": "text" } ], "index": 19 }, { "bbox": [ 130, 341, 475, 353 ], "spans": [ { "bbox": [ 130, 341, 475, 353 ], "score": 1.0, "content": "human feedback. Advances in Neural Information Processing Systems, 35:27730–27744, 2022.", "type": "text" } ], "index": 20 } ], "index": 19 }, { "type": "text", "bbox": [ 109, 358, 504, 379 ], "lines": [ { "bbox": [ 110, 357, 505, 370 ], "spans": [ { "bbox": [ 110, 357, 505, 370 ], "score": 1.0, "content": "[85] Liviu Panait and Sean Luke. Cooperative multi-agent learning: The state of the art. Autonomous Agents", "type": "text" } ], "index": 21 }, { "bbox": [ 129, 368, 295, 380 ], "spans": [ { "bbox": [ 129, 368, 295, 380 ], "score": 1.0, "content": "and Multi-Agent Systems, 11:387–434, 2005.", "type": "text" } ], "index": 22 } ], "index": 21.5 }, { "type": "text", "bbox": [ 110, 386, 504, 407 ], "lines": [ { "bbox": [ 110, 385, 506, 398 ], "spans": [ { "bbox": [ 110, 385, 506, 398 ], "score": 1.0, "content": "[86] Alexandros Papangelis, Karthik Gopalakrishnan, Aishwarya Padmakumar, Seokhwan Kim, Gokhan Tur,", "type": "text" } ], "index": 23 }, { "bbox": [ 129, 396, 434, 407 ], "spans": [ { "bbox": [ 129, 396, 434, 407 ], "score": 1.0, "content": "and Dilek Z. Hakkani-Tür. Generative conversational networks. In SIGDIAL, 2021.", "type": "text" } ], "index": 24 } ], "index": 23.5 }, { "type": "text", "bbox": [ 110, 413, 504, 435 ], "lines": [ { "bbox": [ 110, 413, 506, 426 ], "spans": [ { "bbox": [ 110, 413, 506, 426 ], "score": 1.0, "content": "[87] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings", "type": "text" } ], "index": 25 }, { "bbox": [ 130, 423, 493, 435 ], "spans": [ { "bbox": [ 130, 423, 493, 435 ], "score": 1.0, "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967–3976, 2019.", "type": "text" } ], "index": 26 } ], "index": 25.5 }, { "type": "text", "bbox": [ 110, 441, 505, 472 ], "lines": [ { "bbox": [ 110, 441, 506, 453 ], "spans": [ { "bbox": [ 110, 441, 506, 453 ], "score": 1.0, "content": "[88] Peyman Passban, Yimeng Wu, Mehdi Rezagholizadeh, and Qun Liu. Alp-kd: Attention-based layer", "type": "text" } ], "index": 27 }, { "bbox": [ 129, 451, 507, 464 ], "spans": [ { "bbox": [ 129, 451, 507, 464 ], "score": 1.0, "content": "projection for knowledge distillation. In Proceedings of the AAAI Conference on artificial intelligence,", "type": "text" } ], "index": 28 }, { "bbox": [ 131, 461, 272, 473 ], "spans": [ { "bbox": [ 131, 461, 272, 473 ], "score": 1.0, "content": "volume 35, pages 13657–13665, 2021.", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "text", "bbox": [ 111, 478, 409, 490 ], "lines": [ { "bbox": [ 110, 477, 410, 492 ], "spans": [ { "bbox": [ 110, 477, 410, 492 ], "score": 1.0, "content": "[89] Sundar Pichai. An important next step on our ai journey. Google Blog, 2023.", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 110, 496, 505, 517 ], "lines": [ { "bbox": [ 110, 495, 506, 507 ], "spans": [ { "bbox": [ 110, 495, 506, 507 ], "score": 1.0, "content": "[90] Stefan Poslad. Specifying protocols for multi-agent systems interaction. ACM Transactions on Au-", "type": "text" } ], "index": 31 }, { "bbox": [ 130, 506, 344, 518 ], "spans": [ { "bbox": [ 130, 506, 344, 518 ], "score": 1.0, "content": "tonomous and Adaptive Systems (TAAS), 2(4):15–es, 2007.", "type": "text" } ], "index": 32 } ], "index": 31.5 }, { "type": "text", "bbox": [ 110, 523, 504, 545 ], "lines": [ { "bbox": [ 110, 522, 505, 537 ], "spans": [ { "bbox": [ 110, 522, 505, 537 ], "score": 1.0, "content": "[91] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language", "type": "text" } ], "index": 33 }, { "bbox": [ 129, 533, 391, 546 ], "spans": [ { "bbox": [ 129, 533, 391, 546 ], "score": 1.0, "content": "models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.", "type": "text" } ], "index": 34 } ], "index": 33.5 }, { "type": "text", "bbox": [ 111, 551, 506, 592 ], "lines": [ { "bbox": [ 111, 551, 505, 563 ], "spans": [ { "bbox": [ 111, 551, 505, 563 ], "score": 1.0, "content": "[92] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel", "type": "text" } ], "index": 35 }, { "bbox": [ 129, 560, 506, 573 ], "spans": [ { "bbox": [ 129, 560, 506, 573 ], "score": 1.0, "content": "Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce,", "type": "text" } ], "index": 36 }, { "bbox": [ 129, 570, 507, 584 ], "spans": [ { "bbox": [ 129, 570, 507, 584 ], "score": 1.0, "content": "Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar,", "type": "text" } ], "index": 37 }, { "bbox": [ 129, 581, 304, 593 ], "spans": [ { "bbox": [ 129, 581, 304, 593 ], "score": 1.0, "content": "and Nando de Freitas. A generalist agent, 2022.", "type": "text" } ], "index": 38 } ], "index": 36.5 }, { "type": "text", "bbox": [ 109, 599, 504, 630 ], "lines": [ { "bbox": [ 110, 598, 505, 611 ], "spans": [ { "bbox": [ 110, 598, 505, 611 ], "score": 1.0, "content": "[93] Laria Reynolds and Kyle McDonell. Prompt programming for large language models: Beyond the", "type": "text" } ], "index": 39 }, { "bbox": [ 129, 607, 505, 622 ], "spans": [ { "bbox": [ 129, 607, 505, 622 ], "score": 1.0, "content": "few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing", "type": "text" } ], "index": 40 }, { "bbox": [ 130, 619, 227, 630 ], "spans": [ { "bbox": [ 130, 619, 227, 630 ], "score": 1.0, "content": "Systems, pages 1–7, 2021.", "type": "text" } ], "index": 41 } ], "index": 40 }, { "type": "text", "bbox": [ 108, 636, 504, 657 ], "lines": [ { "bbox": [ 110, 636, 505, 648 ], "spans": [ { "bbox": [ 110, 636, 505, 648 ], "score": 1.0, "content": "[94] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution", "type": "text" } ], "index": 42 }, { "bbox": [ 130, 646, 317, 657 ], "spans": [ { "bbox": [ 130, 646, 317, 657 ], "score": 1.0, "content": "image synthesis with latent diffusion models, 2021.", "type": "text" } ], "index": 43 } ], "index": 42.5 }, { "type": "text", "bbox": [ 109, 663, 504, 685 ], "lines": [ { "bbox": [ 111, 664, 505, 675 ], "spans": [ { "bbox": [ 111, 664, 505, 675 ], "score": 1.0, "content": "[95] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua", "type": "text" } ], "index": 44 }, { "bbox": [ 129, 673, 424, 685 ], "spans": [ { "bbox": [ 129, 673, 424, 685 ], "score": 1.0, "content": "Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.", "type": "text" } ], "index": 45 } ], "index": 44.5 }, { "type": "text", "bbox": [ 110, 691, 504, 722 ], "lines": [ { "bbox": [ 110, 690, 505, 705 ], "spans": [ { "bbox": [ 110, 690, 505, 705 ], "score": 1.0, "content": "[96] Andy Rosenbaum, Saleh Soltan, Wael Hamza, Yannick Versley, and Markus Boese. Linguist: Language", "type": "text" } ], "index": 46 }, { "bbox": [ 130, 702, 505, 713 ], "spans": [ { "bbox": [ 130, 702, 505, 713 ], "score": 1.0, "content": "model instruction tuning to generate annotated utterances for intent classification and slot tagging. arXiv", "type": "text" } ], "index": 47 }, { "bbox": [ 129, 712, 254, 722 ], "spans": [ { "bbox": [ 129, 712, 254, 722 ], "score": 1.0, "content": "preprint arXiv:2209.09900, 2022.", "type": "text" } ], "index": 48 } ], "index": 47 } ], "page_idx": 14, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 313, 754 ], "spans": [ { "bbox": [ 299, 740, 313, 754 ], "score": 1.0, "content": "15", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 109, 73, 505, 94 ], "lines": [ { "bbox": [ 111, 74, 506, 84 ], "spans": [ { "bbox": [ 111, 74, 506, 84 ], "score": 1.0, "content": "[75] Yu Meng, Jiaxin Huang, Yu Zhang, and Jiawei Han. Generating training data with language models:", "type": "text" } ], "index": 0 }, { "bbox": [ 131, 83, 506, 95 ], "spans": [ { "bbox": [ 131, 83, 506, 95 ], "score": 1.0, "content": "Towards zero-shot language understanding. In Advances in Neural Information Processing Systems, 2022.", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 111, 74, 506, 95 ] }, { "type": "text", "bbox": [ 111, 100, 351, 112 ], "lines": [ { "bbox": [ 110, 100, 351, 113 ], "spans": [ { "bbox": [ 110, 100, 351, 113 ], "score": 1.0, "content": "[76] Marvin Minsky. Society of mind. Simon and Schuster, 1988.", "type": "text" } ], "index": 2 } ], "index": 2, "bbox_fs": [ 110, 100, 351, 113 ] }, { "type": "text", "bbox": [ 110, 118, 504, 139 ], "lines": [ { "bbox": [ 110, 117, 506, 130 ], "spans": [ { "bbox": [ 110, 117, 506, 130 ], "score": 1.0, "content": "[77] Marvin Minsky. The emotion machine: Commonsense thinking, artificial intelligence, and the future of", "type": "text" } ], "index": 3 }, { "bbox": [ 130, 128, 293, 139 ], "spans": [ { "bbox": [ 130, 128, 293, 139 ], "score": 1.0, "content": "the human mind. Simon and Schuster, 2007.", "type": "text" } ], "index": 4 } ], "index": 3.5, "bbox_fs": [ 110, 117, 506, 139 ] }, { "type": "text", "bbox": [ 110, 145, 503, 167 ], "lines": [ { "bbox": [ 110, 145, 506, 158 ], "spans": [ { "bbox": [ 110, 145, 506, 158 ], "score": 1.0, "content": "[78] Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task generalization via", "type": "text" } ], "index": 5 }, { "bbox": [ 130, 156, 348, 167 ], "spans": [ { "bbox": [ 130, 156, 348, 167 ], "score": 1.0, "content": "natural language crowdsourcing instructions. In ACL, 2022.", "type": "text" } ], "index": 6 } ], "index": 5.5, "bbox_fs": [ 110, 145, 506, 167 ] }, { "type": "text", "bbox": [ 109, 173, 504, 195 ], "lines": [ { "bbox": [ 110, 173, 505, 186 ], "spans": [ { "bbox": [ 110, 173, 505, 186 ], "score": 1.0, "content": "[79] Igor Mordatch and Pieter Abbeel. Emergence of grounded compositional language in multi-agent", "type": "text" } ], "index": 7 }, { "bbox": [ 129, 183, 477, 195 ], "spans": [ { "bbox": [ 129, 183, 477, 195 ], "score": 1.0, "content": "populations. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018.", "type": "text" } ], "index": 8 } ], "index": 7.5, "bbox_fs": [ 110, 173, 505, 195 ] }, { "type": "text", "bbox": [ 111, 200, 505, 241 ], "lines": [ { "bbox": [ 111, 201, 506, 213 ], "spans": [ { "bbox": [ 111, 201, 506, 213 ], "score": 1.0, "content": "[80] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse,", "type": "text" } ], "index": 9 }, { "bbox": [ 131, 211, 505, 223 ], "spans": [ { "bbox": [ 131, 211, 505, 223 ], "score": 1.0, "content": "Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen", "type": "text" } ], "index": 10 }, { "bbox": [ 131, 221, 505, 232 ], "spans": [ { "bbox": [ 131, 221, 505, 232 ], "score": 1.0, "content": "Krueger, Kevin Button, Matthew Knight, Benjamin Chess, and John Schulman. Webgpt: Browser-assisted", "type": "text" } ], "index": 11 }, { "bbox": [ 131, 231, 308, 242 ], "spans": [ { "bbox": [ 131, 231, 308, 242 ], "score": 1.0, "content": "question-answering with human feedback, 2021.", "type": "text" } ], "index": 12 } ], "index": 10.5, "bbox_fs": [ 111, 201, 506, 242 ] }, { "type": "text", "bbox": [ 109, 248, 504, 279 ], "lines": [ { "bbox": [ 110, 248, 506, 261 ], "spans": [ { "bbox": [ 110, 248, 506, 261 ], "score": 1.0, "content": "[81] Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber,", "type": "text" } ], "index": 13 }, { "bbox": [ 129, 258, 506, 270 ], "spans": [ { "bbox": [ 129, 258, 506, 270 ], "score": 1.0, "content": "David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, Charles Sutton, and Augustus Odena.", "type": "text" } ], "index": 14 }, { "bbox": [ 130, 268, 454, 280 ], "spans": [ { "bbox": [ 130, 268, 454, 280 ], "score": 1.0, "content": "Show your work: Scratchpads for intermediate computation with language models, 2021.", "type": "text" } ], "index": 15 } ], "index": 14, "bbox_fs": [ 110, 248, 506, 280 ] }, { "type": "text", "bbox": [ 111, 285, 317, 297 ], "lines": [ { "bbox": [ 110, 285, 317, 299 ], "spans": [ { "bbox": [ 110, 285, 317, 299 ], "score": 1.0, "content": "[82] OpenAI. Introducing chatgpt. Open AI Blog, 2022.", "type": "text" } ], "index": 16 } ], "index": 16, "bbox_fs": [ 110, 285, 317, 299 ] }, { "type": "text", "bbox": [ 112, 303, 300, 315 ], "lines": [ { "bbox": [ 110, 302, 301, 316 ], "spans": [ { "bbox": [ 110, 302, 301, 316 ], "score": 1.0, "content": "[83] OpenAI. Chatgpt plugins. OpenAI blog, 2023.", "type": "text" } ], "index": 17 } ], "index": 17, "bbox_fs": [ 110, 302, 301, 316 ] }, { "type": "text", "bbox": [ 110, 321, 504, 352 ], "lines": [ { "bbox": [ 110, 320, 506, 334 ], "spans": [ { "bbox": [ 110, 320, 506, 334 ], "score": 1.0, "content": "[84] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang,", "type": "text" } ], "index": 18 }, { "bbox": [ 130, 331, 505, 343 ], "spans": [ { "bbox": [ 130, 331, 505, 343 ], "score": 1.0, "content": "Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with", "type": "text" } ], "index": 19 }, { "bbox": [ 130, 341, 475, 353 ], "spans": [ { "bbox": [ 130, 341, 475, 353 ], "score": 1.0, "content": "human feedback. Advances in Neural Information Processing Systems, 35:27730–27744, 2022.", "type": "text" } ], "index": 20 } ], "index": 19, "bbox_fs": [ 110, 320, 506, 353 ] }, { "type": "text", "bbox": [ 109, 358, 504, 379 ], "lines": [ { "bbox": [ 110, 357, 505, 370 ], "spans": [ { "bbox": [ 110, 357, 505, 370 ], "score": 1.0, "content": "[85] Liviu Panait and Sean Luke. Cooperative multi-agent learning: The state of the art. Autonomous Agents", "type": "text" } ], "index": 21 }, { "bbox": [ 129, 368, 295, 380 ], "spans": [ { "bbox": [ 129, 368, 295, 380 ], "score": 1.0, "content": "and Multi-Agent Systems, 11:387–434, 2005.", "type": "text" } ], "index": 22 } ], "index": 21.5, "bbox_fs": [ 110, 357, 505, 380 ] }, { "type": "text", "bbox": [ 110, 386, 504, 407 ], "lines": [ { "bbox": [ 110, 385, 506, 398 ], "spans": [ { "bbox": [ 110, 385, 506, 398 ], "score": 1.0, "content": "[86] Alexandros Papangelis, Karthik Gopalakrishnan, Aishwarya Padmakumar, Seokhwan Kim, Gokhan Tur,", "type": "text" } ], "index": 23 }, { "bbox": [ 129, 396, 434, 407 ], "spans": [ { "bbox": [ 129, 396, 434, 407 ], "score": 1.0, "content": "and Dilek Z. Hakkani-Tür. Generative conversational networks. In SIGDIAL, 2021.", "type": "text" } ], "index": 24 } ], "index": 23.5, "bbox_fs": [ 110, 385, 506, 407 ] }, { "type": "text", "bbox": [ 110, 413, 504, 435 ], "lines": [ { "bbox": [ 110, 413, 506, 426 ], "spans": [ { "bbox": [ 110, 413, 506, 426 ], "score": 1.0, "content": "[87] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings", "type": "text" } ], "index": 25 }, { "bbox": [ 130, 423, 493, 435 ], "spans": [ { "bbox": [ 130, 423, 493, 435 ], "score": 1.0, "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967–3976, 2019.", "type": "text" } ], "index": 26 } ], "index": 25.5, "bbox_fs": [ 110, 413, 506, 435 ] }, { "type": "text", "bbox": [ 110, 441, 505, 472 ], "lines": [ { "bbox": [ 110, 441, 506, 453 ], "spans": [ { "bbox": [ 110, 441, 506, 453 ], "score": 1.0, "content": "[88] Peyman Passban, Yimeng Wu, Mehdi Rezagholizadeh, and Qun Liu. Alp-kd: Attention-based layer", "type": "text" } ], "index": 27 }, { "bbox": [ 129, 451, 507, 464 ], "spans": [ { "bbox": [ 129, 451, 507, 464 ], "score": 1.0, "content": "projection for knowledge distillation. In Proceedings of the AAAI Conference on artificial intelligence,", "type": "text" } ], "index": 28 }, { "bbox": [ 131, 461, 272, 473 ], "spans": [ { "bbox": [ 131, 461, 272, 473 ], "score": 1.0, "content": "volume 35, pages 13657–13665, 2021.", "type": "text" } ], "index": 29 } ], "index": 28, "bbox_fs": [ 110, 441, 507, 473 ] }, { "type": "text", "bbox": [ 111, 478, 409, 490 ], "lines": [ { "bbox": [ 110, 477, 410, 492 ], "spans": [ { "bbox": [ 110, 477, 410, 492 ], "score": 1.0, "content": "[89] Sundar Pichai. An important next step on our ai journey. Google Blog, 2023.", "type": "text" } ], "index": 30 } ], "index": 30, "bbox_fs": [ 110, 477, 410, 492 ] }, { "type": "text", "bbox": [ 110, 496, 505, 517 ], "lines": [ { "bbox": [ 110, 495, 506, 507 ], "spans": [ { "bbox": [ 110, 495, 506, 507 ], "score": 1.0, "content": "[90] Stefan Poslad. Specifying protocols for multi-agent systems interaction. ACM Transactions on Au-", "type": "text" } ], "index": 31 }, { "bbox": [ 130, 506, 344, 518 ], "spans": [ { "bbox": [ 130, 506, 344, 518 ], "score": 1.0, "content": "tonomous and Adaptive Systems (TAAS), 2(4):15–es, 2007.", "type": "text" } ], "index": 32 } ], "index": 31.5, "bbox_fs": [ 110, 495, 506, 518 ] }, { "type": "text", "bbox": [ 110, 523, 504, 545 ], "lines": [ { "bbox": [ 110, 522, 505, 537 ], "spans": [ { "bbox": [ 110, 522, 505, 537 ], "score": 1.0, "content": "[91] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language", "type": "text" } ], "index": 33 }, { "bbox": [ 129, 533, 391, 546 ], "spans": [ { "bbox": [ 129, 533, 391, 546 ], "score": 1.0, "content": "models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.", "type": "text" } ], "index": 34 } ], "index": 33.5, "bbox_fs": [ 110, 522, 505, 546 ] }, { "type": "text", "bbox": [ 111, 551, 506, 592 ], "lines": [ { "bbox": [ 111, 551, 505, 563 ], "spans": [ { "bbox": [ 111, 551, 505, 563 ], "score": 1.0, "content": "[92] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel", "type": "text" } ], "index": 35 }, { "bbox": [ 129, 560, 506, 573 ], "spans": [ { "bbox": [ 129, 560, 506, 573 ], "score": 1.0, "content": "Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce,", "type": "text" } ], "index": 36 }, { "bbox": [ 129, 570, 507, 584 ], "spans": [ { "bbox": [ 129, 570, 507, 584 ], "score": 1.0, "content": "Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar,", "type": "text" } ], "index": 37 }, { "bbox": [ 129, 581, 304, 593 ], "spans": [ { "bbox": [ 129, 581, 304, 593 ], "score": 1.0, "content": "and Nando de Freitas. A generalist agent, 2022.", "type": "text" } ], "index": 38 } ], "index": 36.5, "bbox_fs": [ 111, 551, 507, 593 ] }, { "type": "text", "bbox": [ 109, 599, 504, 630 ], "lines": [ { "bbox": [ 110, 598, 505, 611 ], "spans": [ { "bbox": [ 110, 598, 505, 611 ], "score": 1.0, "content": "[93] Laria Reynolds and Kyle McDonell. Prompt programming for large language models: Beyond the", "type": "text" } ], "index": 39 }, { "bbox": [ 129, 607, 505, 622 ], "spans": [ { "bbox": [ 129, 607, 505, 622 ], "score": 1.0, "content": "few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing", "type": "text" } ], "index": 40 }, { "bbox": [ 130, 619, 227, 630 ], "spans": [ { "bbox": [ 130, 619, 227, 630 ], "score": 1.0, "content": "Systems, pages 1–7, 2021.", "type": "text" } ], "index": 41 } ], "index": 40, "bbox_fs": [ 110, 598, 505, 630 ] }, { "type": "text", "bbox": [ 108, 636, 504, 657 ], "lines": [ { "bbox": [ 110, 636, 505, 648 ], "spans": [ { "bbox": [ 110, 636, 505, 648 ], "score": 1.0, "content": "[94] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution", "type": "text" } ], "index": 42 }, { "bbox": [ 130, 646, 317, 657 ], "spans": [ { "bbox": [ 130, 646, 317, 657 ], "score": 1.0, "content": "image synthesis with latent diffusion models, 2021.", "type": "text" } ], "index": 43 } ], "index": 42.5, "bbox_fs": [ 110, 636, 505, 657 ] }, { "type": "text", "bbox": [ 109, 663, 504, 685 ], "lines": [ { "bbox": [ 111, 664, 505, 675 ], "spans": [ { "bbox": [ 111, 664, 505, 675 ], "score": 1.0, "content": "[95] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua", "type": "text" } ], "index": 44 }, { "bbox": [ 129, 673, 424, 685 ], "spans": [ { "bbox": [ 129, 673, 424, 685 ], "score": 1.0, "content": "Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.", "type": "text" } ], "index": 45 } ], "index": 44.5, "bbox_fs": [ 111, 664, 505, 685 ] }, { "type": "text", "bbox": [ 110, 691, 504, 722 ], "lines": [ { "bbox": [ 110, 690, 505, 705 ], "spans": [ { "bbox": [ 110, 690, 505, 705 ], "score": 1.0, "content": "[96] Andy Rosenbaum, Saleh Soltan, Wael Hamza, Yannick Versley, and Markus Boese. Linguist: Language", "type": "text" } ], "index": 46 }, { "bbox": [ 130, 702, 505, 713 ], "spans": [ { "bbox": [ 130, 702, 505, 713 ], "score": 1.0, "content": "model instruction tuning to generate annotated utterances for intent classification and slot tagging. arXiv", "type": "text" } ], "index": 47 }, { "bbox": [ 129, 712, 254, 722 ], "spans": [ { "bbox": [ 129, 712, 254, 722 ], "score": 1.0, "content": "preprint arXiv:2209.09900, 2022.", "type": "text" } ], "index": 48 } ], "index": 47, "bbox_fs": [ 110, 690, 505, 722 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 41, 507, 728 ], "lines": [ { "bbox": [ 110, 73, 459, 86 ], "spans": [ { "bbox": [ 110, 73, 459, 86 ], "score": 1.0, "content": "[97] Stuart J Russell. Artificial intelligence a modern approach. Pearson Education, Inc., 2010.", "type": "text" } ], "index": 0 }, { "bbox": [ 109, 90, 506, 104 ], "spans": [ { "bbox": [ 109, 90, 506, 104 ], "score": 1.0, "content": "[98] Gaurav Sahu, Pau Rodriguez, Issam H Laradji, Parmida Atighehchian, David Vazquez, and Dzmitry", "type": "text" } ], "index": 1 }, { "bbox": [ 129, 101, 506, 115 ], "spans": [ { "bbox": [ 129, 101, 506, 115 ], "score": 1.0, "content": "Bahdanau. Data augmentation for intent classification with off-the-shelf large language models. ACL,", "type": "text" } ], "index": 2 }, { "bbox": [ 131, 112, 154, 122 ], "spans": [ { "bbox": [ 131, 112, 154, 122 ], "score": 1.0, "content": "2022.", "type": "text" } ], "index": 3 }, { "bbox": [ 110, 129, 506, 141 ], "spans": [ { "bbox": [ 110, 129, 506, 141 ], "score": 1.0, "content": "[99] William Saunders, Catherine Yeh, Jeff Wu, Steven Bills, Long Ouyang, Jonathan Ward, and Jan Leike.", "type": "text" } ], "index": 4 }, { "bbox": [ 129, 138, 474, 153 ], "spans": [ { "bbox": [ 129, 138, 474, 153 ], "score": 1.0, "content": "Self-critiquing models for assisting human evaluators. arXiv preprint arXiv:2206.05802, 2022.", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 156, 506, 168 ], "spans": [ { "bbox": [ 106, 156, 506, 168 ], "score": 1.0, "content": "[100] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer,", "type": "text" } ], "index": 6 }, { "bbox": [ 129, 167, 506, 179 ], "spans": [ { "bbox": [ 129, 167, 506, 179 ], "score": 1.0, "content": "Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools.", "type": "text" } ], "index": 7 }, { "bbox": [ 128, 176, 277, 189 ], "spans": [ { "bbox": [ 128, 176, 277, 189 ], "score": 1.0, "content": "arXiv preprint arXiv:2302.04761, 2023.", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 195, 506, 207 ], "spans": [ { "bbox": [ 105, 195, 506, 207 ], "score": 1.0, "content": "[101] Timo Schick and Hinrich Schütze. Generating datasets with pretrained language models. In Proceedings", "type": "text" } ], "index": 9 }, { "bbox": [ 129, 204, 506, 218 ], "spans": [ { "bbox": [ 129, 204, 506, 218 ], "score": 1.0, "content": "of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943–6951, Online", "type": "text" } ], "index": 10 }, { "bbox": [ 130, 215, 491, 227 ], "spans": [ { "bbox": [ 130, 215, 491, 227 ], "score": 1.0, "content": "and Punta Cana, Dominican Republic, November 2021. Association for Computational Linguistics.", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 232, 506, 244 ], "spans": [ { "bbox": [ 106, 232, 506, 244 ], "score": 1.0, "content": "[102] Junjie Sheng, Xiangfeng Wang, Bo Jin, Junchi Yan, Wenhao Li, Tsung-Hui Chang, Jun Wang, and", "type": "text" } ], "index": 12 }, { "bbox": [ 130, 243, 506, 255 ], "spans": [ { "bbox": [ 130, 243, 506, 255 ], "score": 1.0, "content": "Hongyuan Zha. Learning structured communication for multi-agent reinforcement learning. Autonomous", "type": "text" } ], "index": 13 }, { "bbox": [ 128, 253, 309, 264 ], "spans": [ { "bbox": [ 128, 253, 309, 264 ], "score": 1.0, "content": "Agents and Multi-Agent Systems, 36(2):50, 2022.", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 270, 506, 282 ], "spans": [ { "bbox": [ 106, 270, 506, 282 ], "score": 1.0, "content": "[103] Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won", "type": "text" } ], "index": 15 }, { "bbox": [ 130, 279, 506, 293 ], "spans": [ { "bbox": [ 130, 279, 506, 293 ], "score": 1.0, "content": "Chung, Yi Tay, Sebastian Ruder, Denny Zhou, et al. Language models are multilingual chain-of-thought", "type": "text" } ], "index": 16 }, { "bbox": [ 129, 290, 226, 302 ], "spans": [ { "bbox": [ 129, 290, 226, 302 ], "score": 1.0, "content": "reasoners. In ICLR, 2023.", "type": "text" } ], "index": 17 }, { "bbox": [ 103, 306, 507, 321 ], "spans": [ { "bbox": [ 103, 306, 507, 321 ], "score": 1.0, "content": "[104] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt:", "type": "text" } ], "index": 18 }, { "bbox": [ 130, 318, 505, 330 ], "spans": [ { "bbox": [ 130, 318, 505, 330 ], "score": 1.0, "content": "Eliciting knowledge from language models with automatically generated prompts. arXiv preprint", "type": "text" } ], "index": 19 }, { "bbox": [ 129, 328, 224, 339 ], "spans": [ { "bbox": [ 129, 328, 224, 339 ], "score": 1.0, "content": "arXiv:2010.15980, 2020.", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 344, 506, 359 ], "spans": [ { "bbox": [ 104, 344, 506, 359 ], "score": 1.0, "content": "[105] Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflexion: an autonomous agent with dynamic memory", "type": "text" } ], "index": 21 }, { "bbox": [ 130, 356, 348, 368 ], "spans": [ { "bbox": [ 130, 356, 348, 368 ], "score": 1.0, "content": "and self-reflection. arXiv preprint arXiv:2303.11366, 2023.", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 373, 506, 385 ], "spans": [ { "bbox": [ 106, 373, 506, 385 ], "score": 1.0, "content": "[106] Kurt Shuster, Jing Xu, Mojtaba Komeili, Da Ju, Eric Michael Smith, Stephen Roller, Megan Ung, Moya", "type": "text" } ], "index": 23 }, { "bbox": [ 130, 384, 505, 396 ], "spans": [ { "bbox": [ 130, 384, 505, 396 ], "score": 1.0, "content": "Chen, Kushal Arora, Joshua Lane, et al. Blenderbot 3: a deployed conversational agent that continually", "type": "text" } ], "index": 24 }, { "bbox": [ 129, 393, 385, 407 ], "spans": [ { "bbox": [ 129, 393, 385, 407 ], "score": 1.0, "content": "learns to responsibly engage. arXiv preprint arXiv:2208.03188, 2022.", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 411, 506, 423 ], "spans": [ { "bbox": [ 105, 411, 506, 423 ], "score": 1.0, "content": "[107] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", "type": "text" } ], "index": 26 }, { "bbox": [ 129, 421, 506, 434 ], "spans": [ { "bbox": [ 129, 421, 506, 434 ], "score": 1.0, "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game", "type": "text" } ], "index": 27 }, { "bbox": [ 128, 430, 434, 444 ], "spans": [ { "bbox": [ 128, 430, 434, 444 ], "score": 1.0, "content": "of go with deep neural networks and tree search. nature, 529(7587):484–489, 2016.", "type": "text" } ], "index": 28 }, { "bbox": [ 103, 447, 507, 462 ], "spans": [ { "bbox": [ 103, 447, 507, 462 ], "score": 1.0, "content": "[108] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez,", "type": "text" } ], "index": 29 }, { "bbox": [ 128, 457, 506, 473 ], "spans": [ { "bbox": [ 128, 457, 506, 473 ], "score": 1.0, "content": "Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without", "type": "text" } ], "index": 30 }, { "bbox": [ 130, 470, 327, 479 ], "spans": [ { "bbox": [ 130, 470, 327, 479 ], "score": 1.0, "content": "human knowledge. nature, 550(7676):354–359, 2017.", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 486, 506, 500 ], "spans": [ { "bbox": [ 105, 486, 506, 500 ], "score": 1.0, "content": "[109] Abishek Sridhar, Robert Lo, Frank F. Xu, Hao Zhu, and Shuyan Zhou. Hierarchical prompting assists", "type": "text" } ], "index": 32 }, { "bbox": [ 130, 497, 351, 510 ], "spans": [ { "bbox": [ 130, 497, 351, 510 ], "score": 1.0, "content": "large language model on web navigation. In ArXiv, preprint.", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 515, 505, 527 ], "spans": [ { "bbox": [ 105, 515, 505, 527 ], "score": 1.0, "content": "[110] Jonathan Stray. Aligning ai optimization to community well-being. International Journal of Community", "type": "text" } ], "index": 34 }, { "bbox": [ 130, 525, 245, 537 ], "spans": [ { "bbox": [ 130, 525, 245, 537 ], "score": 1.0, "content": "Well-Being, 3:443 – 463, 2020.", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 543, 506, 555 ], "spans": [ { "bbox": [ 106, 543, 506, 555 ], "score": 1.0, "content": "[111] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation.", "type": "text" } ], "index": 36 }, { "bbox": [ 129, 552, 355, 565 ], "spans": [ { "bbox": [ 129, 552, 355, 565 ], "score": 1.0, "content": "Advances in neural information processing systems, 29, 2016.", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 570, 506, 582 ], "spans": [ { "bbox": [ 105, 570, 506, 582 ], "score": 1.0, "content": "[112] Alex Tamkin, Miles Brundage, Jack Clark, and Deep Ganguli. Understanding the capabilities, limitations,", "type": "text" } ], "index": 38 }, { "bbox": [ 129, 580, 445, 594 ], "spans": [ { "bbox": [ 129, 580, 445, 594 ], "score": 1.0, "content": "and societal impact of large language models. arXiv preprint arXiv:2102.02503, 2021.", "type": "text" } ], "index": 39 }, { "bbox": [ 104, 597, 506, 611 ], "spans": [ { "bbox": [ 104, 597, 506, 611 ], "score": 1.0, "content": "[113] Ming Tan. Multi-agent reinforcement learning: Independent versus cooperative agents. In International", "type": "text" } ], "index": 40 }, { "bbox": [ 130, 608, 279, 621 ], "spans": [ { "bbox": [ 130, 608, 279, 621 ], "score": 1.0, "content": "Conference on Machine Learning, 1997.", "type": "text" } ], "index": 41 }, { "bbox": [ 104, 624, 507, 639 ], "spans": [ { "bbox": [ 104, 624, 507, 639 ], "score": 1.0, "content": "[114] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang,", "type": "text" } ], "index": 42 }, { "bbox": [ 128, 634, 506, 649 ], "spans": [ { "bbox": [ 128, 634, 506, 649 ], "score": 1.0, "content": "and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.", "type": "text" } ], "index": 43 }, { "bbox": [ 130, 646, 296, 658 ], "spans": [ { "bbox": [ 130, 646, 296, 658 ], "score": 1.0, "content": "com/tatsu-lab/stanford_alpaca, 2023.", "type": "text" } ], "index": 44 }, { "bbox": [ 104, 663, 507, 677 ], "spans": [ { "bbox": [ 104, 663, 507, 677 ], "score": 1.0, "content": "[115] Gerald Tesauro et al. Temporal difference learning and td-gammon. Communications of the ACM,", "type": "text" } ], "index": 45 }, { "bbox": [ 130, 673, 203, 685 ], "spans": [ { "bbox": [ 130, 673, 203, 685 ], "score": 1.0, "content": "38(3):58–68, 1995.", "type": "text" } ], "index": 46 }, { "bbox": [ 102, 688, 508, 706 ], "spans": [ { "bbox": [ 102, 688, 508, 706 ], "score": 1.0, "content": "[116] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng,", "type": "text" } ], "index": 47 }, { "bbox": [ 129, 699, 507, 714 ], "spans": [ { "bbox": [ 129, 699, 507, 714 ], "score": 1.0, "content": "Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications.", "type": "text" } ], "index": 48 }, { "bbox": [ 130, 712, 277, 723 ], "spans": [ { "bbox": [ 130, 712, 277, 723 ], "score": 1.0, "content": "arXiv preprint arXiv:2201.08239, 2022.", "type": "text" } ], "index": 49 } ], "index": 24.5 } ], "page_idx": 15, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 313, 754 ], "spans": [ { "bbox": [ 299, 740, 313, 754 ], "score": 1.0, "content": "16", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 105, 41, 507, 728 ], "lines": [ { "bbox": [ 110, 73, 459, 86 ], "spans": [ { "bbox": [ 110, 73, 459, 86 ], "score": 1.0, "content": "[97] Stuart J Russell. Artificial intelligence a modern approach. Pearson Education, Inc., 2010.", "type": "text" } ], "index": 0, "is_list_end_line": true }, { "bbox": [ 109, 90, 506, 104 ], "spans": [ { "bbox": [ 109, 90, 506, 104 ], "score": 1.0, "content": "[98] Gaurav Sahu, Pau Rodriguez, Issam H Laradji, Parmida Atighehchian, David Vazquez, and Dzmitry", "type": "text" } ], "index": 1 }, { "bbox": [ 129, 101, 506, 115 ], "spans": [ { "bbox": [ 129, 101, 506, 115 ], "score": 1.0, "content": "Bahdanau. Data augmentation for intent classification with off-the-shelf large language models. ACL,", "type": "text" } ], "index": 2 }, { "bbox": [ 131, 112, 154, 122 ], "spans": [ { "bbox": [ 131, 112, 154, 122 ], "score": 1.0, "content": "2022.", "type": "text" } ], "index": 3, "is_list_end_line": true }, { "bbox": [ 110, 129, 506, 141 ], "spans": [ { "bbox": [ 110, 129, 506, 141 ], "score": 1.0, "content": "[99] William Saunders, Catherine Yeh, Jeff Wu, Steven Bills, Long Ouyang, Jonathan Ward, and Jan Leike.", "type": "text" } ], "index": 4 }, { "bbox": [ 129, 138, 474, 153 ], "spans": [ { "bbox": [ 129, 138, 474, 153 ], "score": 1.0, "content": "Self-critiquing models for assisting human evaluators. arXiv preprint arXiv:2206.05802, 2022.", "type": "text" } ], "index": 5, "is_list_end_line": true }, { "bbox": [ 106, 156, 506, 168 ], "spans": [ { "bbox": [ 106, 156, 506, 168 ], "score": 1.0, "content": "[100] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer,", "type": "text" } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 129, 167, 506, 179 ], "spans": [ { "bbox": [ 129, 167, 506, 179 ], "score": 1.0, "content": "Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools.", "type": "text" } ], "index": 7 }, { "bbox": [ 128, 176, 277, 189 ], "spans": [ { "bbox": [ 128, 176, 277, 189 ], "score": 1.0, "content": "arXiv preprint arXiv:2302.04761, 2023.", "type": "text" } ], "index": 8, "is_list_end_line": true }, { "bbox": [ 105, 195, 506, 207 ], "spans": [ { "bbox": [ 105, 195, 506, 207 ], "score": 1.0, "content": "[101] Timo Schick and Hinrich Schütze. Generating datasets with pretrained language models. In Proceedings", "type": "text" } ], "index": 9, "is_list_start_line": true }, { "bbox": [ 129, 204, 506, 218 ], "spans": [ { "bbox": [ 129, 204, 506, 218 ], "score": 1.0, "content": "of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943–6951, Online", "type": "text" } ], "index": 10 }, { "bbox": [ 130, 215, 491, 227 ], "spans": [ { "bbox": [ 130, 215, 491, 227 ], "score": 1.0, "content": "and Punta Cana, Dominican Republic, November 2021. Association for Computational Linguistics.", "type": "text" } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 106, 232, 506, 244 ], "spans": [ { "bbox": [ 106, 232, 506, 244 ], "score": 1.0, "content": "[102] Junjie Sheng, Xiangfeng Wang, Bo Jin, Junchi Yan, Wenhao Li, Tsung-Hui Chang, Jun Wang, and", "type": "text" } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 130, 243, 506, 255 ], "spans": [ { "bbox": [ 130, 243, 506, 255 ], "score": 1.0, "content": "Hongyuan Zha. Learning structured communication for multi-agent reinforcement learning. Autonomous", "type": "text" } ], "index": 13 }, { "bbox": [ 128, 253, 309, 264 ], "spans": [ { "bbox": [ 128, 253, 309, 264 ], "score": 1.0, "content": "Agents and Multi-Agent Systems, 36(2):50, 2022.", "type": "text" } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 106, 270, 506, 282 ], "spans": [ { "bbox": [ 106, 270, 506, 282 ], "score": 1.0, "content": "[103] Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won", "type": "text" } ], "index": 15, "is_list_start_line": true }, { "bbox": [ 130, 279, 506, 293 ], "spans": [ { "bbox": [ 130, 279, 506, 293 ], "score": 1.0, "content": "Chung, Yi Tay, Sebastian Ruder, Denny Zhou, et al. Language models are multilingual chain-of-thought", "type": "text" } ], "index": 16 }, { "bbox": [ 129, 290, 226, 302 ], "spans": [ { "bbox": [ 129, 290, 226, 302 ], "score": 1.0, "content": "reasoners. In ICLR, 2023.", "type": "text" } ], "index": 17, "is_list_end_line": true }, { "bbox": [ 103, 306, 507, 321 ], "spans": [ { "bbox": [ 103, 306, 507, 321 ], "score": 1.0, "content": "[104] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt:", "type": "text" } ], "index": 18, "is_list_start_line": true }, { "bbox": [ 130, 318, 505, 330 ], "spans": [ { "bbox": [ 130, 318, 505, 330 ], "score": 1.0, "content": "Eliciting knowledge from language models with automatically generated prompts. arXiv preprint", "type": "text" } ], "index": 19 }, { "bbox": [ 129, 328, 224, 339 ], "spans": [ { "bbox": [ 129, 328, 224, 339 ], "score": 1.0, "content": "arXiv:2010.15980, 2020.", "type": "text" } ], "index": 20, "is_list_end_line": true }, { "bbox": [ 104, 344, 506, 359 ], "spans": [ { "bbox": [ 104, 344, 506, 359 ], "score": 1.0, "content": "[105] Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflexion: an autonomous agent with dynamic memory", "type": "text" } ], "index": 21, "is_list_start_line": true }, { "bbox": [ 130, 356, 348, 368 ], "spans": [ { "bbox": [ 130, 356, 348, 368 ], "score": 1.0, "content": "and self-reflection. arXiv preprint arXiv:2303.11366, 2023.", "type": "text" } ], "index": 22, "is_list_end_line": true }, { "bbox": [ 106, 373, 506, 385 ], "spans": [ { "bbox": [ 106, 373, 506, 385 ], "score": 1.0, "content": "[106] Kurt Shuster, Jing Xu, Mojtaba Komeili, Da Ju, Eric Michael Smith, Stephen Roller, Megan Ung, Moya", "type": "text" } ], "index": 23, "is_list_start_line": true }, { "bbox": [ 130, 384, 505, 396 ], "spans": [ { "bbox": [ 130, 384, 505, 396 ], "score": 1.0, "content": "Chen, Kushal Arora, Joshua Lane, et al. Blenderbot 3: a deployed conversational agent that continually", "type": "text" } ], "index": 24 }, { "bbox": [ 129, 393, 385, 407 ], "spans": [ { "bbox": [ 129, 393, 385, 407 ], "score": 1.0, "content": "learns to responsibly engage. arXiv preprint arXiv:2208.03188, 2022.", "type": "text" } ], "index": 25, "is_list_end_line": true }, { "bbox": [ 105, 411, 506, 423 ], "spans": [ { "bbox": [ 105, 411, 506, 423 ], "score": 1.0, "content": "[107] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", "type": "text" } ], "index": 26, "is_list_start_line": true }, { "bbox": [ 129, 421, 506, 434 ], "spans": [ { "bbox": [ 129, 421, 506, 434 ], "score": 1.0, "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game", "type": "text" } ], "index": 27 }, { "bbox": [ 128, 430, 434, 444 ], "spans": [ { "bbox": [ 128, 430, 434, 444 ], "score": 1.0, "content": "of go with deep neural networks and tree search. nature, 529(7587):484–489, 2016.", "type": "text" } ], "index": 28, "is_list_end_line": true }, { "bbox": [ 103, 447, 507, 462 ], "spans": [ { "bbox": [ 103, 447, 507, 462 ], "score": 1.0, "content": "[108] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez,", "type": "text" } ], "index": 29, "is_list_start_line": true }, { "bbox": [ 128, 457, 506, 473 ], "spans": [ { "bbox": [ 128, 457, 506, 473 ], "score": 1.0, "content": "Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without", "type": "text" } ], "index": 30 }, { "bbox": [ 130, 470, 327, 479 ], "spans": [ { "bbox": [ 130, 470, 327, 479 ], "score": 1.0, "content": "human knowledge. nature, 550(7676):354–359, 2017.", "type": "text" } ], "index": 31, "is_list_end_line": true }, { "bbox": [ 105, 486, 506, 500 ], "spans": [ { "bbox": [ 105, 486, 506, 500 ], "score": 1.0, "content": "[109] Abishek Sridhar, Robert Lo, Frank F. Xu, Hao Zhu, and Shuyan Zhou. Hierarchical prompting assists", "type": "text" } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 130, 497, 351, 510 ], "spans": [ { "bbox": [ 130, 497, 351, 510 ], "score": 1.0, "content": "large language model on web navigation. In ArXiv, preprint.", "type": "text" } ], "index": 33, "is_list_end_line": true }, { "bbox": [ 105, 515, 505, 527 ], "spans": [ { "bbox": [ 105, 515, 505, 527 ], "score": 1.0, "content": "[110] Jonathan Stray. Aligning ai optimization to community well-being. International Journal of Community", "type": "text" } ], "index": 34, "is_list_start_line": true }, { "bbox": [ 130, 525, 245, 537 ], "spans": [ { "bbox": [ 130, 525, 245, 537 ], "score": 1.0, "content": "Well-Being, 3:443 – 463, 2020.", "type": "text" } ], "index": 35, "is_list_end_line": true }, { "bbox": [ 106, 543, 506, 555 ], "spans": [ { "bbox": [ 106, 543, 506, 555 ], "score": 1.0, "content": "[111] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation.", "type": "text" } ], "index": 36, "is_list_start_line": true }, { "bbox": [ 129, 552, 355, 565 ], "spans": [ { "bbox": [ 129, 552, 355, 565 ], "score": 1.0, "content": "Advances in neural information processing systems, 29, 2016.", "type": "text" } ], "index": 37, "is_list_end_line": true }, { "bbox": [ 105, 570, 506, 582 ], "spans": [ { "bbox": [ 105, 570, 506, 582 ], "score": 1.0, "content": "[112] Alex Tamkin, Miles Brundage, Jack Clark, and Deep Ganguli. Understanding the capabilities, limitations,", "type": "text" } ], "index": 38, "is_list_start_line": true }, { "bbox": [ 129, 580, 445, 594 ], "spans": [ { "bbox": [ 129, 580, 445, 594 ], "score": 1.0, "content": "and societal impact of large language models. arXiv preprint arXiv:2102.02503, 2021.", "type": "text" } ], "index": 39, "is_list_end_line": true }, { "bbox": [ 104, 597, 506, 611 ], "spans": [ { "bbox": [ 104, 597, 506, 611 ], "score": 1.0, "content": "[113] Ming Tan. Multi-agent reinforcement learning: Independent versus cooperative agents. In International", "type": "text" } ], "index": 40, "is_list_start_line": true }, { "bbox": [ 130, 608, 279, 621 ], "spans": [ { "bbox": [ 130, 608, 279, 621 ], "score": 1.0, "content": "Conference on Machine Learning, 1997.", "type": "text" } ], "index": 41, "is_list_end_line": true }, { "bbox": [ 104, 624, 507, 639 ], "spans": [ { "bbox": [ 104, 624, 507, 639 ], "score": 1.0, "content": "[114] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang,", "type": "text" } ], "index": 42, "is_list_start_line": true }, { "bbox": [ 128, 634, 506, 649 ], "spans": [ { "bbox": [ 128, 634, 506, 649 ], "score": 1.0, "content": "and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.", "type": "text" } ], "index": 43 }, { "bbox": [ 130, 646, 296, 658 ], "spans": [ { "bbox": [ 130, 646, 296, 658 ], "score": 1.0, "content": "com/tatsu-lab/stanford_alpaca, 2023.", "type": "text" } ], "index": 44, "is_list_end_line": true }, { "bbox": [ 104, 663, 507, 677 ], "spans": [ { "bbox": [ 104, 663, 507, 677 ], "score": 1.0, "content": "[115] Gerald Tesauro et al. Temporal difference learning and td-gammon. Communications of the ACM,", "type": "text" } ], "index": 45, "is_list_start_line": true }, { "bbox": [ 130, 673, 203, 685 ], "spans": [ { "bbox": [ 130, 673, 203, 685 ], "score": 1.0, "content": "38(3):58–68, 1995.", "type": "text" } ], "index": 46, "is_list_end_line": true }, { "bbox": [ 102, 688, 508, 706 ], "spans": [ { "bbox": [ 102, 688, 508, 706 ], "score": 1.0, "content": "[116] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng,", "type": "text" } ], "index": 47, "is_list_start_line": true }, { "bbox": [ 129, 699, 507, 714 ], "spans": [ { "bbox": [ 129, 699, 507, 714 ], "score": 1.0, "content": "Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications.", "type": "text" } ], "index": 48 }, { "bbox": [ 130, 712, 277, 723 ], "spans": [ { "bbox": [ 130, 712, 277, 723 ], "score": 1.0, "content": "arXiv preprint arXiv:2201.08239, 2022.", "type": "text" } ], "index": 49, "is_list_end_line": true } ], "index": 24.5, "bbox_fs": [ 102, 73, 508, 723 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 73, 504, 104 ], "lines": [ { "bbox": [ 106, 73, 506, 86 ], "spans": [ { "bbox": [ 106, 73, 506, 86 ], "score": 1.0, "content": "[117] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix,", "type": "text" } ], "index": 0 }, { "bbox": [ 129, 83, 505, 95 ], "spans": [ { "bbox": [ 129, 83, 505, 95 ], "score": 1.0, "content": "Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation", "type": "text" } ], "index": 1 }, { "bbox": [ 129, 93, 343, 104 ], "spans": [ { "bbox": [ 129, 93, 343, 104 ], "score": 1.0, "content": "language models. arXiv preprint arXiv:2302.13971, 2023.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "text", "bbox": [ 105, 110, 504, 132 ], "lines": [ { "bbox": [ 105, 109, 506, 124 ], "spans": [ { "bbox": [ 105, 109, 506, 124 ], "score": 1.0, "content": "[118] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. Self-consistency", "type": "text" } ], "index": 3 }, { "bbox": [ 130, 120, 396, 132 ], "spans": [ { "bbox": [ 130, 120, 396, 132 ], "score": 1.0, "content": "improves chain of thought reasoning in language models. In ICLR, 2023.", "type": "text" } ], "index": 4 } ], "index": 3.5 }, { "type": "text", "bbox": [ 106, 138, 504, 169 ], "lines": [ { "bbox": [ 105, 138, 506, 150 ], "spans": [ { "bbox": [ 105, 138, 506, 150 ], "score": 1.0, "content": "[119] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and", "type": "text" } ], "index": 5 }, { "bbox": [ 130, 149, 505, 160 ], "spans": [ { "bbox": [ 130, 149, 505, 160 ], "score": 1.0, "content": "Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv", "type": "text" } ], "index": 6 }, { "bbox": [ 129, 159, 255, 169 ], "spans": [ { "bbox": [ 129, 159, 255, 169 ], "score": 1.0, "content": "preprint arXiv:2212.10560, 2022.", "type": "text" } ], "index": 7 } ], "index": 6 }, { "type": "text", "bbox": [ 105, 176, 504, 207 ], "lines": [ { "bbox": [ 106, 176, 506, 188 ], "spans": [ { "bbox": [ 106, 176, 506, 188 ], "score": 1.0, "content": "[120] Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, An-", "type": "text" } ], "index": 8 }, { "bbox": [ 129, 186, 506, 199 ], "spans": [ { "bbox": [ 129, 186, 506, 199 ], "score": 1.0, "content": "jana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al. Super-", "type": "text" } ], "index": 9 }, { "bbox": [ 129, 196, 478, 207 ], "spans": [ { "bbox": [ 129, 196, 363, 207 ], "score": 1.0, "content": "naturalinstructions:generalization via declarative instructions on", "type": "text" }, { "bbox": [ 363, 197, 388, 206 ], "score": 0.84, "content": "1 6 0 0 +", "type": "inline_equation" }, { "bbox": [ 388, 196, 478, 207 ], "score": 1.0, "content": "tasks. In EMNLP, 2022.", "type": "text" } ], "index": 10 } ], "index": 9 }, { "type": "text", "bbox": [ 105, 213, 504, 245 ], "lines": [ { "bbox": [ 105, 213, 506, 226 ], "spans": [ { "bbox": [ 105, 213, 506, 226 ], "score": 1.0, "content": "[121] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du,", "type": "text" } ], "index": 11 }, { "bbox": [ 129, 223, 505, 236 ], "spans": [ { "bbox": [ 129, 223, 505, 236 ], "score": 1.0, "content": "Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint", "type": "text" } ], "index": 12 }, { "bbox": [ 130, 234, 224, 244 ], "spans": [ { "bbox": [ 130, 234, 224, 244 ], "score": 1.0, "content": "arXiv:2109.01652, 2021.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "text", "bbox": [ 106, 252, 507, 292 ], "lines": [ { "bbox": [ 106, 252, 506, 263 ], "spans": [ { "bbox": [ 106, 252, 506, 263 ], "score": 1.0, "content": "[122] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama,", "type": "text" } ], "index": 14 }, { "bbox": [ 129, 260, 505, 275 ], "spans": [ { "bbox": [ 129, 260, 505, 275 ], "score": 1.0, "content": "Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy", "type": "text" } ], "index": 15 }, { "bbox": [ 129, 271, 505, 283 ], "spans": [ { "bbox": [ 129, 271, 505, 283 ], "score": 1.0, "content": "Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. Transactions on", "type": "text" } ], "index": 16 }, { "bbox": [ 129, 281, 337, 294 ], "spans": [ { "bbox": [ 129, 281, 337, 294 ], "score": 1.0, "content": "Machine Learning Research, 2022. Survey Certification.", "type": "text" } ], "index": 17 } ], "index": 15.5 }, { "type": "text", "bbox": [ 103, 299, 505, 321 ], "lines": [ { "bbox": [ 106, 299, 505, 311 ], "spans": [ { "bbox": [ 106, 299, 505, 311 ], "score": 1.0, "content": "[123] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain", "type": "text" } ], "index": 18 }, { "bbox": [ 131, 309, 506, 321 ], "spans": [ { "bbox": [ 131, 309, 506, 321 ], "score": 1.0, "content": "of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903, 2022.", "type": "text" } ], "index": 19 } ], "index": 18.5 }, { "type": "text", "bbox": [ 107, 326, 505, 388 ], "lines": [ { "bbox": [ 106, 327, 506, 339 ], "spans": [ { "bbox": [ 106, 327, 506, 339 ], "score": 1.0, "content": "[124] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric", "type": "text" } ], "index": 20 }, { "bbox": [ 129, 336, 507, 349 ], "spans": [ { "bbox": [ 129, 336, 507, 349 ], "score": 1.0, "content": "Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen,", "type": "text" } ], "index": 21 }, { "bbox": [ 129, 346, 507, 359 ], "spans": [ { "bbox": [ 129, 346, 507, 359 ], "score": 1.0, "content": "Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame,", "type": "text" } ], "index": 22 }, { "bbox": [ 131, 357, 505, 368 ], "spans": [ { "bbox": [ 131, 357, 505, 368 ], "score": 1.0, "content": "Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the-art natural language processing. In", "type": "text" } ], "index": 23 }, { "bbox": [ 129, 367, 506, 379 ], "spans": [ { "bbox": [ 129, 367, 506, 379 ], "score": 1.0, "content": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System", "type": "text" } ], "index": 24 }, { "bbox": [ 129, 376, 482, 389 ], "spans": [ { "bbox": [ 129, 376, 482, 389 ], "score": 1.0, "content": "Demonstrations, pages 38–45, Online, October 2020. Association for Computational Linguistics.", "type": "text" } ], "index": 25 } ], "index": 22.5 }, { "type": "text", "bbox": [ 106, 394, 444, 406 ], "lines": [ { "bbox": [ 105, 393, 445, 407 ], "spans": [ { "bbox": [ 105, 393, 445, 407 ], "score": 1.0, "content": "[125] Michael Wooldridge. An introduction to multiagent systems. John wiley & sons, 2009.", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 412, 504, 443 ], "lines": [ { "bbox": [ 105, 411, 505, 424 ], "spans": [ { "bbox": [ 105, 411, 505, 424 ], "score": 1.0, "content": "[126] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao.", "type": "text" } ], "index": 27 }, { "bbox": [ 129, 421, 505, 435 ], "spans": [ { "bbox": [ 129, 421, 505, 435 ], "score": 1.0, "content": "ReAct: Synergizing reasoning and acting in language models. In International Conference on Learning", "type": "text" } ], "index": 28 }, { "bbox": [ 130, 433, 243, 443 ], "spans": [ { "bbox": [ 130, 433, 243, 443 ], "score": 1.0, "content": "Representations (ICLR), 2023.", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "text", "bbox": [ 105, 450, 503, 471 ], "lines": [ { "bbox": [ 106, 450, 505, 462 ], "spans": [ { "bbox": [ 106, 450, 505, 462 ], "score": 1.0, "content": "[127] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance", "type": "text" } ], "index": 30 }, { "bbox": [ 130, 460, 477, 471 ], "spans": [ { "bbox": [ 130, 460, 477, 471 ], "score": 1.0, "content": "of convolutional neural networks via attention transfer. arXiv preprint arXiv:1612.03928, 2016.", "type": "text" } ], "index": 31 } ], "index": 30.5 }, { "type": "text", "bbox": [ 107, 477, 502, 498 ], "lines": [ { "bbox": [ 106, 477, 504, 489 ], "spans": [ { "bbox": [ 106, 477, 504, 489 ], "score": 1.0, "content": "[128] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. Star: Bootstrapping reasoning with", "type": "text" } ], "index": 32 }, { "bbox": [ 129, 487, 193, 499 ], "spans": [ { "bbox": [ 129, 487, 193, 499 ], "score": 1.0, "content": "reasoning, 2022.", "type": "text" } ], "index": 33 } ], "index": 32.5 }, { "type": "text", "bbox": [ 107, 505, 504, 526 ], "lines": [ { "bbox": [ 106, 505, 505, 517 ], "spans": [ { "bbox": [ 106, 505, 505, 517 ], "score": 1.0, "content": "[129] Houyu Zhang, Zhenghao Liu, Chenyan Xiong, and Zhiyuan Liu. Grounded conversation generation as", "type": "text" } ], "index": 34 }, { "bbox": [ 130, 515, 377, 526 ], "spans": [ { "bbox": [ 130, 515, 377, 526 ], "score": 1.0, "content": "guided traverses in commonsense knowledge graphs. In ACL, 2020.", "type": "text" } ], "index": 35 } ], "index": 34.5 }, { "type": "text", "bbox": [ 106, 533, 504, 554 ], "lines": [ { "bbox": [ 105, 533, 505, 546 ], "spans": [ { "bbox": [ 105, 533, 505, 546 ], "score": 1.0, "content": "[130] Rongsheng Zhang, Yinhe Zheng, Jianzhi Shao, Xiao-Xi Mao, Yadong Xi, and Minlie Huang. Dialogue", "type": "text" } ], "index": 36 }, { "bbox": [ 130, 542, 496, 555 ], "spans": [ { "bbox": [ 130, 542, 496, 555 ], "score": 1.0, "content": "distillation: Open-domain dialogue augmentation using unpaired data. ArXiv, abs/2009.09427, 2020.", "type": "text" } ], "index": 37 } ], "index": 36.5 }, { "type": "text", "bbox": [ 106, 560, 504, 591 ], "lines": [ { "bbox": [ 105, 560, 505, 572 ], "spans": [ { "bbox": [ 105, 560, 505, 572 ], "score": 1.0, "content": "[131] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher", "type": "text" } ], "index": 38 }, { "bbox": [ 130, 570, 506, 582 ], "spans": [ { "bbox": [ 130, 570, 506, 582 ], "score": 1.0, "content": "Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models.", "type": "text" } ], "index": 39 }, { "bbox": [ 130, 581, 277, 592 ], "spans": [ { "bbox": [ 130, 581, 277, 592 ], "score": 1.0, "content": "arXiv preprint arXiv:2205.01068, 2022.", "type": "text" } ], "index": 40 } ], "index": 39 }, { "type": "text", "bbox": [ 106, 598, 504, 619 ], "lines": [ { "bbox": [ 105, 596, 506, 611 ], "spans": [ { "bbox": [ 105, 596, 506, 611 ], "score": 1.0, "content": "[132] Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large", "type": "text" } ], "index": 41 }, { "bbox": [ 130, 609, 252, 619 ], "spans": [ { "bbox": [ 130, 609, 252, 619 ], "score": 1.0, "content": "language models. In ICLR, 2023.", "type": "text" } ], "index": 42 } ], "index": 41.5 }, { "type": "text", "bbox": [ 106, 626, 504, 647 ], "lines": [ { "bbox": [ 105, 626, 505, 637 ], "spans": [ { "bbox": [ 105, 626, 505, 637 ], "score": 1.0, "content": "[133] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal", "type": "text" } ], "index": 43 }, { "bbox": [ 129, 636, 452, 648 ], "spans": [ { "bbox": [ 129, 636, 452, 648 ], "score": 1.0, "content": "chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023.", "type": "text" } ], "index": 44 } ], "index": 43.5 }, { "type": "text", "bbox": [ 106, 654, 504, 685 ], "lines": [ { "bbox": [ 105, 652, 506, 666 ], "spans": [ { "bbox": [ 105, 652, 506, 666 ], "score": 1.0, "content": "[134] Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans,", "type": "text" } ], "index": 45 }, { "bbox": [ 130, 663, 505, 676 ], "spans": [ { "bbox": [ 130, 663, 505, 676 ], "score": 1.0, "content": "Olivier Bousquet, Quoc Le, and Ed Chi. Least-to-most prompting enables complex reasoning in large", "type": "text" } ], "index": 46 }, { "bbox": [ 130, 674, 343, 685 ], "spans": [ { "bbox": [ 130, 674, 343, 685 ], "score": 1.0, "content": "language models. arXiv preprint arXiv:2205.10625, 2022.", "type": "text" } ], "index": 47 } ], "index": 46 }, { "type": "text", "bbox": [ 105, 691, 503, 722 ], "lines": [ { "bbox": [ 105, 690, 505, 703 ], "spans": [ { "bbox": [ 105, 690, 505, 703 ], "score": 1.0, "content": "[135] Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Yonatan", "type": "text" } ], "index": 48 }, { "bbox": [ 129, 700, 505, 713 ], "spans": [ { "bbox": [ 129, 700, 505, 713 ], "score": 1.0, "content": "Bisk, Daniel Fried, Uri Alon, et al. Webarena: A realistic web environment for building autonomous", "type": "text" } ], "index": 49 }, { "bbox": [ 130, 712, 306, 722 ], "spans": [ { "bbox": [ 130, 712, 306, 722 ], "score": 1.0, "content": "agents. arXiv preprint arXiv:2307.13854, 2023.", "type": "text" } ], "index": 50 } ], "index": 49 } ], "page_idx": 16, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 312, 754 ], "spans": [ { "bbox": [ 299, 740, 312, 754 ], "score": 1.0, "content": "17", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 105, 73, 504, 104 ], "lines": [ { "bbox": [ 106, 73, 506, 86 ], "spans": [ { "bbox": [ 106, 73, 506, 86 ], "score": 1.0, "content": "[117] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix,", "type": "text" } ], "index": 0 }, { "bbox": [ 129, 83, 505, 95 ], "spans": [ { "bbox": [ 129, 83, 505, 95 ], "score": 1.0, "content": "Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation", "type": "text" } ], "index": 1 }, { "bbox": [ 129, 93, 343, 104 ], "spans": [ { "bbox": [ 129, 93, 343, 104 ], "score": 1.0, "content": "language models. arXiv preprint arXiv:2302.13971, 2023.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 106, 73, 506, 104 ] }, { "type": "text", "bbox": [ 105, 110, 504, 132 ], "lines": [ { "bbox": [ 105, 109, 506, 124 ], "spans": [ { "bbox": [ 105, 109, 506, 124 ], "score": 1.0, "content": "[118] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. Self-consistency", "type": "text" } ], "index": 3 }, { "bbox": [ 130, 120, 396, 132 ], "spans": [ { "bbox": [ 130, 120, 396, 132 ], "score": 1.0, "content": "improves chain of thought reasoning in language models. In ICLR, 2023.", "type": "text" } ], "index": 4 } ], "index": 3.5, "bbox_fs": [ 105, 109, 506, 132 ] }, { "type": "text", "bbox": [ 106, 138, 504, 169 ], "lines": [ { "bbox": [ 105, 138, 506, 150 ], "spans": [ { "bbox": [ 105, 138, 506, 150 ], "score": 1.0, "content": "[119] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and", "type": "text" } ], "index": 5 }, { "bbox": [ 130, 149, 505, 160 ], "spans": [ { "bbox": [ 130, 149, 505, 160 ], "score": 1.0, "content": "Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv", "type": "text" } ], "index": 6 }, { "bbox": [ 129, 159, 255, 169 ], "spans": [ { "bbox": [ 129, 159, 255, 169 ], "score": 1.0, "content": "preprint arXiv:2212.10560, 2022.", "type": "text" } ], "index": 7 } ], "index": 6, "bbox_fs": [ 105, 138, 506, 169 ] }, { "type": "text", "bbox": [ 105, 176, 504, 207 ], "lines": [ { "bbox": [ 106, 176, 506, 188 ], "spans": [ { "bbox": [ 106, 176, 506, 188 ], "score": 1.0, "content": "[120] Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, An-", "type": "text" } ], "index": 8 }, { "bbox": [ 129, 186, 506, 199 ], "spans": [ { "bbox": [ 129, 186, 506, 199 ], "score": 1.0, "content": "jana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al. Super-", "type": "text" } ], "index": 9 }, { "bbox": [ 129, 196, 478, 207 ], "spans": [ { "bbox": [ 129, 196, 363, 207 ], "score": 1.0, "content": "naturalinstructions:generalization via declarative instructions on", "type": "text" }, { "bbox": [ 363, 197, 388, 206 ], "score": 0.84, "content": "1 6 0 0 +", "type": "inline_equation" }, { "bbox": [ 388, 196, 478, 207 ], "score": 1.0, "content": "tasks. In EMNLP, 2022.", "type": "text" } ], "index": 10 } ], "index": 9, "bbox_fs": [ 106, 176, 506, 207 ] }, { "type": "text", "bbox": [ 105, 213, 504, 245 ], "lines": [ { "bbox": [ 105, 213, 506, 226 ], "spans": [ { "bbox": [ 105, 213, 506, 226 ], "score": 1.0, "content": "[121] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du,", "type": "text" } ], "index": 11 }, { "bbox": [ 129, 223, 505, 236 ], "spans": [ { "bbox": [ 129, 223, 505, 236 ], "score": 1.0, "content": "Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint", "type": "text" } ], "index": 12 }, { "bbox": [ 130, 234, 224, 244 ], "spans": [ { "bbox": [ 130, 234, 224, 244 ], "score": 1.0, "content": "arXiv:2109.01652, 2021.", "type": "text" } ], "index": 13 } ], "index": 12, "bbox_fs": [ 105, 213, 506, 244 ] }, { "type": "text", "bbox": [ 106, 252, 507, 292 ], "lines": [ { "bbox": [ 106, 252, 506, 263 ], "spans": [ { "bbox": [ 106, 252, 506, 263 ], "score": 1.0, "content": "[122] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama,", "type": "text" } ], "index": 14 }, { "bbox": [ 129, 260, 505, 275 ], "spans": [ { "bbox": [ 129, 260, 505, 275 ], "score": 1.0, "content": "Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy", "type": "text" } ], "index": 15 }, { "bbox": [ 129, 271, 505, 283 ], "spans": [ { "bbox": [ 129, 271, 505, 283 ], "score": 1.0, "content": "Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. Transactions on", "type": "text" } ], "index": 16 }, { "bbox": [ 129, 281, 337, 294 ], "spans": [ { "bbox": [ 129, 281, 337, 294 ], "score": 1.0, "content": "Machine Learning Research, 2022. Survey Certification.", "type": "text" } ], "index": 17 } ], "index": 15.5, "bbox_fs": [ 106, 252, 506, 294 ] }, { "type": "text", "bbox": [ 103, 299, 505, 321 ], "lines": [ { "bbox": [ 106, 299, 505, 311 ], "spans": [ { "bbox": [ 106, 299, 505, 311 ], "score": 1.0, "content": "[123] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain", "type": "text" } ], "index": 18 }, { "bbox": [ 131, 309, 506, 321 ], "spans": [ { "bbox": [ 131, 309, 506, 321 ], "score": 1.0, "content": "of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903, 2022.", "type": "text" } ], "index": 19 } ], "index": 18.5, "bbox_fs": [ 106, 299, 506, 321 ] }, { "type": "text", "bbox": [ 107, 326, 505, 388 ], "lines": [ { "bbox": [ 106, 327, 506, 339 ], "spans": [ { "bbox": [ 106, 327, 506, 339 ], "score": 1.0, "content": "[124] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric", "type": "text" } ], "index": 20 }, { "bbox": [ 129, 336, 507, 349 ], "spans": [ { "bbox": [ 129, 336, 507, 349 ], "score": 1.0, "content": "Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen,", "type": "text" } ], "index": 21 }, { "bbox": [ 129, 346, 507, 359 ], "spans": [ { "bbox": [ 129, 346, 507, 359 ], "score": 1.0, "content": "Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame,", "type": "text" } ], "index": 22 }, { "bbox": [ 131, 357, 505, 368 ], "spans": [ { "bbox": [ 131, 357, 505, 368 ], "score": 1.0, "content": "Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the-art natural language processing. In", "type": "text" } ], "index": 23 }, { "bbox": [ 129, 367, 506, 379 ], "spans": [ { "bbox": [ 129, 367, 506, 379 ], "score": 1.0, "content": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System", "type": "text" } ], "index": 24 }, { "bbox": [ 129, 376, 482, 389 ], "spans": [ { "bbox": [ 129, 376, 482, 389 ], "score": 1.0, "content": "Demonstrations, pages 38–45, Online, October 2020. Association for Computational Linguistics.", "type": "text" } ], "index": 25 } ], "index": 22.5, "bbox_fs": [ 106, 327, 507, 389 ] }, { "type": "text", "bbox": [ 106, 394, 444, 406 ], "lines": [ { "bbox": [ 105, 393, 445, 407 ], "spans": [ { "bbox": [ 105, 393, 445, 407 ], "score": 1.0, "content": "[125] Michael Wooldridge. An introduction to multiagent systems. John wiley & sons, 2009.", "type": "text" } ], "index": 26 } ], "index": 26, "bbox_fs": [ 105, 393, 445, 407 ] }, { "type": "text", "bbox": [ 107, 412, 504, 443 ], "lines": [ { "bbox": [ 105, 411, 505, 424 ], "spans": [ { "bbox": [ 105, 411, 505, 424 ], "score": 1.0, "content": "[126] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao.", "type": "text" } ], "index": 27 }, { "bbox": [ 129, 421, 505, 435 ], "spans": [ { "bbox": [ 129, 421, 505, 435 ], "score": 1.0, "content": "ReAct: Synergizing reasoning and acting in language models. In International Conference on Learning", "type": "text" } ], "index": 28 }, { "bbox": [ 130, 433, 243, 443 ], "spans": [ { "bbox": [ 130, 433, 243, 443 ], "score": 1.0, "content": "Representations (ICLR), 2023.", "type": "text" } ], "index": 29 } ], "index": 28, "bbox_fs": [ 105, 411, 505, 443 ] }, { "type": "text", "bbox": [ 105, 450, 503, 471 ], "lines": [ { "bbox": [ 106, 450, 505, 462 ], "spans": [ { "bbox": [ 106, 450, 505, 462 ], "score": 1.0, "content": "[127] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance", "type": "text" } ], "index": 30 }, { "bbox": [ 130, 460, 477, 471 ], "spans": [ { "bbox": [ 130, 460, 477, 471 ], "score": 1.0, "content": "of convolutional neural networks via attention transfer. arXiv preprint arXiv:1612.03928, 2016.", "type": "text" } ], "index": 31 } ], "index": 30.5, "bbox_fs": [ 106, 450, 505, 471 ] }, { "type": "text", "bbox": [ 107, 477, 502, 498 ], "lines": [ { "bbox": [ 106, 477, 504, 489 ], "spans": [ { "bbox": [ 106, 477, 504, 489 ], "score": 1.0, "content": "[128] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. Star: Bootstrapping reasoning with", "type": "text" } ], "index": 32 }, { "bbox": [ 129, 487, 193, 499 ], "spans": [ { "bbox": [ 129, 487, 193, 499 ], "score": 1.0, "content": "reasoning, 2022.", "type": "text" } ], "index": 33 } ], "index": 32.5, "bbox_fs": [ 106, 477, 504, 499 ] }, { "type": "text", "bbox": [ 107, 505, 504, 526 ], "lines": [ { "bbox": [ 106, 505, 505, 517 ], "spans": [ { "bbox": [ 106, 505, 505, 517 ], "score": 1.0, "content": "[129] Houyu Zhang, Zhenghao Liu, Chenyan Xiong, and Zhiyuan Liu. Grounded conversation generation as", "type": "text" } ], "index": 34 }, { "bbox": [ 130, 515, 377, 526 ], "spans": [ { "bbox": [ 130, 515, 377, 526 ], "score": 1.0, "content": "guided traverses in commonsense knowledge graphs. In ACL, 2020.", "type": "text" } ], "index": 35 } ], "index": 34.5, "bbox_fs": [ 106, 505, 505, 526 ] }, { "type": "text", "bbox": [ 106, 533, 504, 554 ], "lines": [ { "bbox": [ 105, 533, 505, 546 ], "spans": [ { "bbox": [ 105, 533, 505, 546 ], "score": 1.0, "content": "[130] Rongsheng Zhang, Yinhe Zheng, Jianzhi Shao, Xiao-Xi Mao, Yadong Xi, and Minlie Huang. Dialogue", "type": "text" } ], "index": 36 }, { "bbox": [ 130, 542, 496, 555 ], "spans": [ { "bbox": [ 130, 542, 496, 555 ], "score": 1.0, "content": "distillation: Open-domain dialogue augmentation using unpaired data. ArXiv, abs/2009.09427, 2020.", "type": "text" } ], "index": 37 } ], "index": 36.5, "bbox_fs": [ 105, 533, 505, 555 ] }, { "type": "text", "bbox": [ 106, 560, 504, 591 ], "lines": [ { "bbox": [ 105, 560, 505, 572 ], "spans": [ { "bbox": [ 105, 560, 505, 572 ], "score": 1.0, "content": "[131] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher", "type": "text" } ], "index": 38 }, { "bbox": [ 130, 570, 506, 582 ], "spans": [ { "bbox": [ 130, 570, 506, 582 ], "score": 1.0, "content": "Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models.", "type": "text" } ], "index": 39 }, { "bbox": [ 130, 581, 277, 592 ], "spans": [ { "bbox": [ 130, 581, 277, 592 ], "score": 1.0, "content": "arXiv preprint arXiv:2205.01068, 2022.", "type": "text" } ], "index": 40 } ], "index": 39, "bbox_fs": [ 105, 560, 506, 592 ] }, { "type": "text", "bbox": [ 106, 598, 504, 619 ], "lines": [ { "bbox": [ 105, 596, 506, 611 ], "spans": [ { "bbox": [ 105, 596, 506, 611 ], "score": 1.0, "content": "[132] Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large", "type": "text" } ], "index": 41 }, { "bbox": [ 130, 609, 252, 619 ], "spans": [ { "bbox": [ 130, 609, 252, 619 ], "score": 1.0, "content": "language models. In ICLR, 2023.", "type": "text" } ], "index": 42 } ], "index": 41.5, "bbox_fs": [ 105, 596, 506, 619 ] }, { "type": "text", "bbox": [ 106, 626, 504, 647 ], "lines": [ { "bbox": [ 105, 626, 505, 637 ], "spans": [ { "bbox": [ 105, 626, 505, 637 ], "score": 1.0, "content": "[133] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal", "type": "text" } ], "index": 43 }, { "bbox": [ 129, 636, 452, 648 ], "spans": [ { "bbox": [ 129, 636, 452, 648 ], "score": 1.0, "content": "chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023.", "type": "text" } ], "index": 44 } ], "index": 43.5, "bbox_fs": [ 105, 626, 505, 648 ] }, { "type": "text", "bbox": [ 106, 654, 504, 685 ], "lines": [ { "bbox": [ 105, 652, 506, 666 ], "spans": [ { "bbox": [ 105, 652, 506, 666 ], "score": 1.0, "content": "[134] Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans,", "type": "text" } ], "index": 45 }, { "bbox": [ 130, 663, 505, 676 ], "spans": [ { "bbox": [ 130, 663, 505, 676 ], "score": 1.0, "content": "Olivier Bousquet, Quoc Le, and Ed Chi. Least-to-most prompting enables complex reasoning in large", "type": "text" } ], "index": 46 }, { "bbox": [ 130, 674, 343, 685 ], "spans": [ { "bbox": [ 130, 674, 343, 685 ], "score": 1.0, "content": "language models. arXiv preprint arXiv:2205.10625, 2022.", "type": "text" } ], "index": 47 } ], "index": 46, "bbox_fs": [ 105, 652, 506, 685 ] }, { "type": "text", "bbox": [ 105, 691, 503, 722 ], "lines": [ { "bbox": [ 105, 690, 505, 703 ], "spans": [ { "bbox": [ 105, 690, 505, 703 ], "score": 1.0, "content": "[135] Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Yonatan", "type": "text" } ], "index": 48 }, { "bbox": [ 129, 700, 505, 713 ], "spans": [ { "bbox": [ 129, 700, 505, 713 ], "score": 1.0, "content": "Bisk, Daniel Fried, Uri Alon, et al. Webarena: A realistic web environment for building autonomous", "type": "text" } ], "index": 49 }, { "bbox": [ 130, 712, 306, 722 ], "spans": [ { "bbox": [ 130, 712, 306, 722 ], "score": 1.0, "content": "agents. arXiv preprint arXiv:2307.13854, 2023.", "type": "text" } ], "index": 50 } ], "index": 49, "bbox_fs": [ 105, 690, 505, 722 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 73, 506, 132 ], "lines": [ { "bbox": [ 106, 73, 505, 85 ], "spans": [ { "bbox": [ 106, 73, 505, 85 ], "score": 1.0, "content": "[136] Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy", "type": "text" } ], "index": 0 }, { "bbox": [ 131, 83, 504, 94 ], "spans": [ { "bbox": [ 131, 83, 504, 94 ], "score": 1.0, "content": "Ba. Large language models are human-level prompt engineers. In The Eleventh International Conference", "type": "text" } ], "index": 1 }, { "bbox": [ 131, 93, 261, 104 ], "spans": [ { "bbox": [ 131, 93, 261, 104 ], "score": 1.0, "content": "on Learning Representations, 2023.", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 110, 506, 124 ], "spans": [ { "bbox": [ 105, 110, 506, 124 ], "score": 1.0, "content": "[137] Deyao Zhu, Jun Chen, Kilichbek Haydarov, Xiaoqian Shen, Wenxuan Zhang, and Mohamed Elhoseiny.", "type": "text" } ], "index": 3 }, { "bbox": [ 131, 121, 480, 133 ], "spans": [ { "bbox": [ 131, 121, 480, 133 ], "score": 1.0, "content": "Chatgpt asks, blip-2 answers: Automatic questioning towards enriched visual descriptions, 2023.", "type": "text" } ], "index": 4 } ], "index": 2 } ], "page_idx": 17, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 742, 311, 750 ], "lines": [ { "bbox": [ 299, 740, 313, 754 ], "spans": [ { "bbox": [ 299, 740, 313, 754 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 105, 73, 506, 132 ], "lines": [ { "bbox": [ 106, 73, 505, 85 ], "spans": [ { "bbox": [ 106, 73, 505, 85 ], "score": 1.0, "content": "[136] Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy", "type": "text" } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 131, 83, 504, 94 ], "spans": [ { "bbox": [ 131, 83, 504, 94 ], "score": 1.0, "content": "Ba. Large language models are human-level prompt engineers. In The Eleventh International Conference", "type": "text" } ], "index": 1 }, { "bbox": [ 131, 93, 261, 104 ], "spans": [ { "bbox": [ 131, 93, 261, 104 ], "score": 1.0, "content": "on Learning Representations, 2023.", "type": "text" } ], "index": 2, "is_list_end_line": true }, { "bbox": [ 105, 110, 506, 124 ], "spans": [ { "bbox": [ 105, 110, 506, 124 ], "score": 1.0, "content": "[137] Deyao Zhu, Jun Chen, Kilichbek Haydarov, Xiaoqian Shen, Wenxuan Zhang, and Mohamed Elhoseiny.", "type": "text" } ], "index": 3, "is_list_start_line": true }, { "bbox": [ 131, 121, 480, 133 ], "spans": [ { "bbox": [ 131, 121, 480, 133 ], "score": 1.0, "content": "Chatgpt asks, blip-2 answers: Automatic questioning towards enriched visual descriptions, 2023.", "type": "text" } ], "index": 4, "is_list_end_line": true } ], "index": 2, "bbox_fs": [ 105, 73, 506, 133 ] } ] } ], "_backend": "pipeline", "_version_name": "2.2.2" }