{ "pdf_info": [ { "preproc_blocks": [ { "type": "title", "bbox": [ 119, 97, 492, 117 ], "lines": [ { "bbox": [ 117, 97, 493, 120 ], "spans": [ { "bbox": [ 117, 97, 493, 120 ], "score": 1.0, "content": "Locating and Editing Factual Associations in GPT", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 126, 159, 182, 181 ], "lines": [ { "bbox": [ 123, 157, 183, 173 ], "spans": [ { "bbox": [ 123, 157, 183, 173 ], "score": 1.0, "content": "Kevin Meng⇤", "type": "text" } ], "index": 1 }, { "bbox": [ 124, 169, 179, 182 ], "spans": [ { "bbox": [ 124, 169, 179, 182 ], "score": 1.0, "content": "MIT CSAIL", "type": "text" } ], "index": 2 } ], "index": 1.5 }, { "type": "text", "bbox": [ 201, 159, 298, 181 ], "lines": [ { "bbox": [ 223, 159, 275, 171 ], "spans": [ { "bbox": [ 223, 159, 275, 171 ], "score": 1.0, "content": "David Bau⇤", "type": "text" } ], "index": 3 }, { "bbox": [ 200, 169, 298, 183 ], "spans": [ { "bbox": [ 200, 169, 298, 183 ], "score": 1.0, "content": "Northeastern University", "type": "text" } ], "index": 4 } ], "index": 3.5 }, { "type": "text", "bbox": [ 321, 160, 387, 181 ], "lines": [ { "bbox": [ 320, 158, 388, 172 ], "spans": [ { "bbox": [ 320, 158, 388, 172 ], "score": 1.0, "content": "Alex Andonian", "type": "text" } ], "index": 5 }, { "bbox": [ 327, 170, 381, 181 ], "spans": [ { "bbox": [ 327, 170, 381, 181 ], "score": 1.0, "content": "MIT CSAIL", "type": "text" } ], "index": 6 } ], "index": 5.5 }, { "type": "text", "bbox": [ 409, 159, 488, 181 ], "lines": [ { "bbox": [ 408, 159, 490, 171 ], "spans": [ { "bbox": [ 408, 159, 490, 171 ], "score": 1.0, "content": "Yonatan Belinkov†", "type": "text" } ], "index": 7 }, { "bbox": [ 416, 170, 479, 182 ], "spans": [ { "bbox": [ 416, 170, 479, 182 ], "score": 1.0, "content": "Technion – IIT", "type": "text" } ], "index": 8 } ], "index": 7.5 }, { "type": "title", "bbox": [ 283, 210, 328, 223 ], "lines": [ { "bbox": [ 281, 209, 330, 225 ], "spans": [ { "bbox": [ 281, 209, 330, 225 ], "score": 1.0, "content": "Abstract", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "text", "bbox": [ 143, 234, 469, 409 ], "lines": [ { "bbox": [ 142, 234, 471, 247 ], "spans": [ { "bbox": [ 142, 234, 471, 247 ], "score": 1.0, "content": "We analyze the storage and recall of factual associations in autoregressive trans-", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 245, 469, 257 ], "spans": [ { "bbox": [ 141, 245, 469, 257 ], "score": 1.0, "content": "former language models, finding evidence that these associations correspond to", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 255, 469, 268 ], "spans": [ { "bbox": [ 141, 255, 469, 268 ], "score": 1.0, "content": "localized, directly-editable computations. We first develop a causal intervention", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 266, 471, 279 ], "spans": [ { "bbox": [ 141, 266, 471, 279 ], "score": 1.0, "content": "for identifying neuron activations that are decisive in a model’s factual predictions.", "type": "text" } ], "index": 13 }, { "bbox": [ 141, 277, 471, 290 ], "spans": [ { "bbox": [ 141, 277, 471, 290 ], "score": 1.0, "content": "This reveals a distinct set of steps in middle-layer feed-forward modules that me-", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 289, 470, 301 ], "spans": [ { "bbox": [ 141, 289, 470, 301 ], "score": 1.0, "content": "diate factual predictions while processing subject tokens. To test our hypothesis", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 300, 470, 312 ], "spans": [ { "bbox": [ 141, 300, 470, 312 ], "score": 1.0, "content": "that these computations correspond to factual association recall, we modify feed-", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 311, 470, 323 ], "spans": [ { "bbox": [ 141, 311, 470, 323 ], "score": 1.0, "content": "forward weights to update specific factual associations using Rank-One Model", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 322, 469, 333 ], "spans": [ { "bbox": [ 141, 322, 469, 333 ], "score": 1.0, "content": "Editing (ROME). We find that ROME is effective on a standard zero-shot relation", "type": "text" } ], "index": 18 }, { "bbox": [ 142, 333, 469, 344 ], "spans": [ { "bbox": [ 142, 333, 469, 344 ], "score": 1.0, "content": "extraction (zsRE) model-editing task. We also evaluate ROME on a new dataset", "type": "text" } ], "index": 19 }, { "bbox": [ 142, 343, 469, 355 ], "spans": [ { "bbox": [ 142, 343, 469, 355 ], "score": 1.0, "content": "of difficult counterfactual assertions, on which it simultaneously maintains both", "type": "text" } ], "index": 20 }, { "bbox": [ 141, 354, 470, 366 ], "spans": [ { "bbox": [ 141, 354, 470, 366 ], "score": 1.0, "content": "specificity and generalization, whereas other methods sacrifice one or another. Our", "type": "text" } ], "index": 21 }, { "bbox": [ 142, 366, 470, 377 ], "spans": [ { "bbox": [ 142, 366, 470, 377 ], "score": 1.0, "content": "results confirm an important role for mid-layer feed-forward modules in storing fac-", "type": "text" } ], "index": 22 }, { "bbox": [ 142, 376, 469, 388 ], "spans": [ { "bbox": [ 142, 376, 469, 388 ], "score": 1.0, "content": "tual associations and suggest that direct manipulation of computational mechanisms", "type": "text" } ], "index": 23 }, { "bbox": [ 141, 388, 469, 399 ], "spans": [ { "bbox": [ 141, 388, 469, 399 ], "score": 1.0, "content": "may be a feasible approach for model editing. The code, dataset, visualizations, and", "type": "text" } ], "index": 24 }, { "bbox": [ 141, 398, 460, 410 ], "spans": [ { "bbox": [ 141, 398, 460, 410 ], "score": 1.0, "content": "an interactive demo notebook are available at https://rome.baulab.info/.", "type": "text" } ], "index": 25 } ], "index": 17.5 }, { "type": "title", "bbox": [ 107, 429, 190, 442 ], "lines": [ { "bbox": [ 105, 428, 192, 445 ], "spans": [ { "bbox": [ 105, 428, 192, 445 ], "score": 1.0, "content": "1 Introduction", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 456, 504, 478 ], "lines": [ { "bbox": [ 106, 455, 505, 468 ], "spans": [ { "bbox": [ 106, 455, 505, 468 ], "score": 1.0, "content": "Where does a large language model store its facts? In this paper, we report evidence that factual", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 466, 451, 479 ], "spans": [ { "bbox": [ 105, 466, 451, 479 ], "score": 1.0, "content": "associations in GPT correspond to a localized computation that can be directly edited.", "type": "text" } ], "index": 28 } ], "index": 27.5 }, { "type": "text", "bbox": [ 107, 483, 505, 538 ], "lines": [ { "bbox": [ 105, 482, 506, 496 ], "spans": [ { "bbox": [ 105, 482, 506, 496 ], "score": 1.0, "content": "Large language models can predict factual statements about the world (Petroni et al., 2019; Jiang", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 494, 505, 506 ], "spans": [ { "bbox": [ 105, 494, 505, 506 ], "score": 1.0, "content": "et al., 2020; Roberts et al., 2020). For example, given the prefix “The Space Needle is located in the", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 505, 505, 517 ], "spans": [ { "bbox": [ 105, 505, 505, 517 ], "score": 1.0, "content": "city of,” GPT will reliably predict the true answer: “Seattle” (Figure 1a). Factual knowledge has been", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 515, 505, 529 ], "spans": [ { "bbox": [ 105, 515, 505, 529 ], "score": 1.0, "content": "observed to emerge in both autoregressive GPT models (Radford et al., 2019; Brown et al., 2020) and", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 527, 282, 538 ], "spans": [ { "bbox": [ 105, 527, 282, 538 ], "score": 1.0, "content": "masked BERT models (Devlin et al., 2019).", "type": "text" } ], "index": 33 } ], "index": 31 }, { "type": "text", "bbox": [ 107, 543, 505, 609 ], "lines": [ { "bbox": [ 105, 542, 505, 556 ], "spans": [ { "bbox": [ 105, 542, 505, 556 ], "score": 1.0, "content": "In this paper, we investigate how such factual associations are stored within GPT-like autoregressive", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 554, 506, 567 ], "spans": [ { "bbox": [ 105, 554, 506, 567 ], "score": 1.0, "content": "transformer models. Although many of the largest neural networks in use today are autoregressive,", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 566, 505, 577 ], "spans": [ { "bbox": [ 106, 566, 505, 577 ], "score": 1.0, "content": "the way that they store knowledge remains under-explored. Some research has been done for masked", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 576, 505, 588 ], "spans": [ { "bbox": [ 106, 576, 505, 588 ], "score": 1.0, "content": "models (Petroni et al., 2019; Jiang et al., 2020; Elazar et al., 2021a; Geva et al., 2021; Dai et al.,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 587, 506, 600 ], "spans": [ { "bbox": [ 105, 587, 506, 600 ], "score": 1.0, "content": "2022; De Cao et al., 2021), but GPT has architectural differences such as unidirectional attention and", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 598, 379, 610 ], "spans": [ { "bbox": [ 105, 598, 379, 610 ], "score": 1.0, "content": "generation capabilities that provide an opportunity for new insights.", "type": "text" } ], "index": 39 } ], "index": 36.5 }, { "type": "text", "bbox": [ 107, 614, 505, 658 ], "lines": [ { "bbox": [ 105, 613, 506, 627 ], "spans": [ { "bbox": [ 105, 613, 506, 627 ], "score": 1.0, "content": "We use two approaches. First, we trace the causal effects of hidden state activations within GPT using", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 625, 505, 637 ], "spans": [ { "bbox": [ 105, 625, 505, 637 ], "score": 1.0, "content": "causal mediation analysis (Pearl, 2001; Vig et al., 2020b) to identify the specific modules that mediate", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 637, 505, 648 ], "spans": [ { "bbox": [ 106, 637, 505, 648 ], "score": 1.0, "content": "recall of a fact about a subject (Figure 1). Our analysis reveals that feedforward MLPs at a range of", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 647, 491, 659 ], "spans": [ { "bbox": [ 105, 647, 491, 659 ], "score": 1.0, "content": "middle layers are decisive when processing the last token of the subject name (Figures 1b,2b,3).", "type": "text" } ], "index": 43 } ], "index": 41.5 }, { "type": "text", "bbox": [ 107, 663, 505, 686 ], "lines": [ { "bbox": [ 106, 663, 505, 676 ], "spans": [ { "bbox": [ 106, 663, 505, 676 ], "score": 1.0, "content": "Second, we test this finding in model weights by introducing a Rank-One Model Editing method", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 676, 506, 687 ], "spans": [ { "bbox": [ 106, 676, 506, 687 ], "score": 1.0, "content": "(ROME) to alter the parameters that determine a feedfoward layer’s behavior at the decisive token.", "type": "text" } ], "index": 45 } ], "index": 44.5 } ], "page_idx": 0, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 118, 694, 456, 714 ], "lines": [ { "bbox": [ 118, 692, 452, 705 ], "spans": [ { "bbox": [ 118, 692, 452, 705 ], "score": 1.0, "content": "⇤Equal contribution. Correspondence to mengk@mit.edu, davidbau@northeastern.edu.", "type": "text" } ] }, { "bbox": [ 118, 703, 458, 716 ], "spans": [ { "bbox": [ 118, 703, 458, 716 ], "score": 1.0, "content": "†Supported by the Viterbi Fellowship in the Center for Computer Engineering at the Technion.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 106, 731, 385, 742 ], "lines": [ { "bbox": [ 106, 730, 385, 743 ], "spans": [ { "bbox": [ 106, 730, 385, 743 ], "score": 1.0, "content": "36th Conference on Neural Information Processing Systems (NeurIPS 2022).", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 119, 97, 492, 117 ], "lines": [ { "bbox": [ 117, 97, 493, 120 ], "spans": [ { "bbox": [ 117, 97, 493, 120 ], "score": 1.0, "content": "Locating and Editing Factual Associations in GPT", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 126, 159, 182, 181 ], "lines": [ { "bbox": [ 123, 157, 183, 173 ], "spans": [ { "bbox": [ 123, 157, 183, 173 ], "score": 1.0, "content": "Kevin Meng⇤", "type": "text" } ], "index": 1 }, { "bbox": [ 124, 169, 179, 182 ], "spans": [ { "bbox": [ 124, 169, 179, 182 ], "score": 1.0, "content": "MIT CSAIL", "type": "text" } ], "index": 2 } ], "index": 1.5, "bbox_fs": [ 123, 157, 183, 182 ] }, { "type": "text", "bbox": [ 201, 159, 298, 181 ], "lines": [ { "bbox": [ 223, 159, 275, 171 ], "spans": [ { "bbox": [ 223, 159, 275, 171 ], "score": 1.0, "content": "David Bau⇤", "type": "text" } ], "index": 3 }, { "bbox": [ 200, 169, 298, 183 ], "spans": [ { "bbox": [ 200, 169, 298, 183 ], "score": 1.0, "content": "Northeastern University", "type": "text" } ], "index": 4 } ], "index": 3.5, "bbox_fs": [ 200, 159, 298, 183 ] }, { "type": "text", "bbox": [ 321, 160, 387, 181 ], "lines": [ { "bbox": [ 320, 158, 388, 172 ], "spans": [ { "bbox": [ 320, 158, 388, 172 ], "score": 1.0, "content": "Alex Andonian", "type": "text" } ], "index": 5 }, { "bbox": [ 327, 170, 381, 181 ], "spans": [ { "bbox": [ 327, 170, 381, 181 ], "score": 1.0, "content": "MIT CSAIL", "type": "text" } ], "index": 6 } ], "index": 5.5, "bbox_fs": [ 320, 158, 388, 181 ] }, { "type": "text", "bbox": [ 409, 159, 488, 181 ], "lines": [ { "bbox": [ 408, 159, 490, 171 ], "spans": [ { "bbox": [ 408, 159, 490, 171 ], "score": 1.0, "content": "Yonatan Belinkov†", "type": "text" } ], "index": 7 }, { "bbox": [ 416, 170, 479, 182 ], "spans": [ { "bbox": [ 416, 170, 479, 182 ], "score": 1.0, "content": "Technion – IIT", "type": "text" } ], "index": 8 } ], "index": 7.5, "bbox_fs": [ 408, 159, 490, 182 ] }, { "type": "title", "bbox": [ 283, 210, 328, 223 ], "lines": [ { "bbox": [ 281, 209, 330, 225 ], "spans": [ { "bbox": [ 281, 209, 330, 225 ], "score": 1.0, "content": "Abstract", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "text", "bbox": [ 143, 234, 469, 409 ], "lines": [ { "bbox": [ 142, 234, 471, 247 ], "spans": [ { "bbox": [ 142, 234, 471, 247 ], "score": 1.0, "content": "We analyze the storage and recall of factual associations in autoregressive trans-", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 245, 469, 257 ], "spans": [ { "bbox": [ 141, 245, 469, 257 ], "score": 1.0, "content": "former language models, finding evidence that these associations correspond to", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 255, 469, 268 ], "spans": [ { "bbox": [ 141, 255, 469, 268 ], "score": 1.0, "content": "localized, directly-editable computations. We first develop a causal intervention", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 266, 471, 279 ], "spans": [ { "bbox": [ 141, 266, 471, 279 ], "score": 1.0, "content": "for identifying neuron activations that are decisive in a model’s factual predictions.", "type": "text" } ], "index": 13 }, { "bbox": [ 141, 277, 471, 290 ], "spans": [ { "bbox": [ 141, 277, 471, 290 ], "score": 1.0, "content": "This reveals a distinct set of steps in middle-layer feed-forward modules that me-", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 289, 470, 301 ], "spans": [ { "bbox": [ 141, 289, 470, 301 ], "score": 1.0, "content": "diate factual predictions while processing subject tokens. To test our hypothesis", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 300, 470, 312 ], "spans": [ { "bbox": [ 141, 300, 470, 312 ], "score": 1.0, "content": "that these computations correspond to factual association recall, we modify feed-", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 311, 470, 323 ], "spans": [ { "bbox": [ 141, 311, 470, 323 ], "score": 1.0, "content": "forward weights to update specific factual associations using Rank-One Model", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 322, 469, 333 ], "spans": [ { "bbox": [ 141, 322, 469, 333 ], "score": 1.0, "content": "Editing (ROME). We find that ROME is effective on a standard zero-shot relation", "type": "text" } ], "index": 18 }, { "bbox": [ 142, 333, 469, 344 ], "spans": [ { "bbox": [ 142, 333, 469, 344 ], "score": 1.0, "content": "extraction (zsRE) model-editing task. We also evaluate ROME on a new dataset", "type": "text" } ], "index": 19 }, { "bbox": [ 142, 343, 469, 355 ], "spans": [ { "bbox": [ 142, 343, 469, 355 ], "score": 1.0, "content": "of difficult counterfactual assertions, on which it simultaneously maintains both", "type": "text" } ], "index": 20 }, { "bbox": [ 141, 354, 470, 366 ], "spans": [ { "bbox": [ 141, 354, 470, 366 ], "score": 1.0, "content": "specificity and generalization, whereas other methods sacrifice one or another. Our", "type": "text" } ], "index": 21 }, { "bbox": [ 142, 366, 470, 377 ], "spans": [ { "bbox": [ 142, 366, 470, 377 ], "score": 1.0, "content": "results confirm an important role for mid-layer feed-forward modules in storing fac-", "type": "text" } ], "index": 22 }, { "bbox": [ 142, 376, 469, 388 ], "spans": [ { "bbox": [ 142, 376, 469, 388 ], "score": 1.0, "content": "tual associations and suggest that direct manipulation of computational mechanisms", "type": "text" } ], "index": 23 }, { "bbox": [ 141, 388, 469, 399 ], "spans": [ { "bbox": [ 141, 388, 469, 399 ], "score": 1.0, "content": "may be a feasible approach for model editing. The code, dataset, visualizations, and", "type": "text" } ], "index": 24 }, { "bbox": [ 141, 398, 460, 410 ], "spans": [ { "bbox": [ 141, 398, 460, 410 ], "score": 1.0, "content": "an interactive demo notebook are available at https://rome.baulab.info/.", "type": "text" } ], "index": 25 } ], "index": 17.5, "bbox_fs": [ 141, 234, 471, 410 ] }, { "type": "title", "bbox": [ 107, 429, 190, 442 ], "lines": [ { "bbox": [ 105, 428, 192, 445 ], "spans": [ { "bbox": [ 105, 428, 192, 445 ], "score": 1.0, "content": "1 Introduction", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 456, 504, 478 ], "lines": [ { "bbox": [ 106, 455, 505, 468 ], "spans": [ { "bbox": [ 106, 455, 505, 468 ], "score": 1.0, "content": "Where does a large language model store its facts? In this paper, we report evidence that factual", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 466, 451, 479 ], "spans": [ { "bbox": [ 105, 466, 451, 479 ], "score": 1.0, "content": "associations in GPT correspond to a localized computation that can be directly edited.", "type": "text" } ], "index": 28 } ], "index": 27.5, "bbox_fs": [ 105, 455, 505, 479 ] }, { "type": "text", "bbox": [ 107, 483, 505, 538 ], "lines": [ { "bbox": [ 105, 482, 506, 496 ], "spans": [ { "bbox": [ 105, 482, 506, 496 ], "score": 1.0, "content": "Large language models can predict factual statements about the world (Petroni et al., 2019; Jiang", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 494, 505, 506 ], "spans": [ { "bbox": [ 105, 494, 505, 506 ], "score": 1.0, "content": "et al., 2020; Roberts et al., 2020). For example, given the prefix “The Space Needle is located in the", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 505, 505, 517 ], "spans": [ { "bbox": [ 105, 505, 505, 517 ], "score": 1.0, "content": "city of,” GPT will reliably predict the true answer: “Seattle” (Figure 1a). Factual knowledge has been", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 515, 505, 529 ], "spans": [ { "bbox": [ 105, 515, 505, 529 ], "score": 1.0, "content": "observed to emerge in both autoregressive GPT models (Radford et al., 2019; Brown et al., 2020) and", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 527, 282, 538 ], "spans": [ { "bbox": [ 105, 527, 282, 538 ], "score": 1.0, "content": "masked BERT models (Devlin et al., 2019).", "type": "text" } ], "index": 33 } ], "index": 31, "bbox_fs": [ 105, 482, 506, 538 ] }, { "type": "text", "bbox": [ 107, 543, 505, 609 ], "lines": [ { "bbox": [ 105, 542, 505, 556 ], "spans": [ { "bbox": [ 105, 542, 505, 556 ], "score": 1.0, "content": "In this paper, we investigate how such factual associations are stored within GPT-like autoregressive", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 554, 506, 567 ], "spans": [ { "bbox": [ 105, 554, 506, 567 ], "score": 1.0, "content": "transformer models. Although many of the largest neural networks in use today are autoregressive,", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 566, 505, 577 ], "spans": [ { "bbox": [ 106, 566, 505, 577 ], "score": 1.0, "content": "the way that they store knowledge remains under-explored. Some research has been done for masked", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 576, 505, 588 ], "spans": [ { "bbox": [ 106, 576, 505, 588 ], "score": 1.0, "content": "models (Petroni et al., 2019; Jiang et al., 2020; Elazar et al., 2021a; Geva et al., 2021; Dai et al.,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 587, 506, 600 ], "spans": [ { "bbox": [ 105, 587, 506, 600 ], "score": 1.0, "content": "2022; De Cao et al., 2021), but GPT has architectural differences such as unidirectional attention and", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 598, 379, 610 ], "spans": [ { "bbox": [ 105, 598, 379, 610 ], "score": 1.0, "content": "generation capabilities that provide an opportunity for new insights.", "type": "text" } ], "index": 39 } ], "index": 36.5, "bbox_fs": [ 105, 542, 506, 610 ] }, { "type": "text", "bbox": [ 107, 614, 505, 658 ], "lines": [ { "bbox": [ 105, 613, 506, 627 ], "spans": [ { "bbox": [ 105, 613, 506, 627 ], "score": 1.0, "content": "We use two approaches. First, we trace the causal effects of hidden state activations within GPT using", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 625, 505, 637 ], "spans": [ { "bbox": [ 105, 625, 505, 637 ], "score": 1.0, "content": "causal mediation analysis (Pearl, 2001; Vig et al., 2020b) to identify the specific modules that mediate", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 637, 505, 648 ], "spans": [ { "bbox": [ 106, 637, 505, 648 ], "score": 1.0, "content": "recall of a fact about a subject (Figure 1). Our analysis reveals that feedforward MLPs at a range of", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 647, 491, 659 ], "spans": [ { "bbox": [ 105, 647, 491, 659 ], "score": 1.0, "content": "middle layers are decisive when processing the last token of the subject name (Figures 1b,2b,3).", "type": "text" } ], "index": 43 } ], "index": 41.5, "bbox_fs": [ 105, 613, 506, 659 ] }, { "type": "text", "bbox": [ 107, 663, 505, 686 ], "lines": [ { "bbox": [ 106, 663, 505, 676 ], "spans": [ { "bbox": [ 106, 663, 505, 676 ], "score": 1.0, "content": "Second, we test this finding in model weights by introducing a Rank-One Model Editing method", "type": "text" } ], "index": 44 }, { "bbox": [ 106, 676, 506, 687 ], "spans": [ { "bbox": [ 106, 676, 506, 687 ], "score": 1.0, "content": "(ROME) to alter the parameters that determine a feedfoward layer’s behavior at the decisive token.", "type": "text" } ], "index": 45 } ], "index": 44.5, "bbox_fs": [ 106, 663, 506, 687 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 107, 69, 505, 251 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 69, 505, 251 ], "group_id": 0, "lines": [ { "bbox": [ 107, 69, 505, 251 ], "spans": [ { "bbox": [ 107, 69, 505, 251 ], "score": 0.956, "type": "image", "image_path": "7c0dd76794a726970bdc2ce46cc2c22679725ee4074a597abf7a6e36dff2b24d.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 69, 505, 129.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 107, 129.66666666666666, 505, 190.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 107, 190.33333333333331, 505, 250.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 256, 505, 307 ], "group_id": 0, "lines": [ { "bbox": [ 105, 254, 506, 268 ], "spans": [ { "bbox": [ 105, 254, 506, 268 ], "score": 1.0, "content": "Figure 1: Causal Traces compute the causal effect of neuron activations by running the network twice: (a)", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 266, 506, 277 ], "spans": [ { "bbox": [ 106, 266, 506, 277 ], "score": 1.0, "content": "once normally, and (b) once where we corrupt the subject token and then (c) restore selected internal activations", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 276, 505, 287 ], "spans": [ { "bbox": [ 106, 276, 505, 287 ], "score": 1.0, "content": "to their clean value. (d) Some sets of activations cause the output to return to the original prediction; the light", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 285, 505, 298 ], "spans": [ { "bbox": [ 105, 285, 505, 298 ], "score": 1.0, "content": "blue path shows an example of information flow. The causal impact on output probability is mapped for the", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 296, 496, 307 ], "spans": [ { "bbox": [ 106, 296, 496, 307 ], "score": 1.0, "content": "effect of (e) each hidden state on the prediction, (f) only MLP activations, and (g) only attention activations.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 106, 329, 504, 352 ], "lines": [ { "bbox": [ 106, 329, 506, 342 ], "spans": [ { "bbox": [ 106, 329, 506, 342 ], "score": 1.0, "content": "Despite the simplicity of the intervention, we find that ROME is similarly effective to other model-", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 341, 458, 354 ], "spans": [ { "bbox": [ 106, 341, 458, 354 ], "score": 1.0, "content": "editing approaches on a standard zero-shot relation extraction benchmark (Section 3.2).", "type": "text" } ], "index": 9 } ], "index": 8.5 }, { "type": "text", "bbox": [ 106, 357, 505, 434 ], "lines": [ { "bbox": [ 105, 356, 505, 369 ], "spans": [ { "bbox": [ 105, 356, 505, 369 ], "score": 1.0, "content": "To evaluate ROME’s impact on more difficult cases, we introduce a dataset of counterfactual assertions", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 368, 505, 381 ], "spans": [ { "bbox": [ 106, 368, 505, 381 ], "score": 1.0, "content": "(Section 3.3) that would not have been observed in pretraining. Our evaluations (Section 3.4) confirm", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 379, 505, 391 ], "spans": [ { "bbox": [ 105, 379, 505, 391 ], "score": 1.0, "content": "that midlayer MLP modules can store factual associations that generalize beyond specific surface", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 390, 506, 402 ], "spans": [ { "bbox": [ 106, 390, 506, 402 ], "score": 1.0, "content": "forms, while remaining specific to the subject. Compared to previous fine-tuning (Zhu et al., 2020),", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 401, 505, 413 ], "spans": [ { "bbox": [ 106, 401, 505, 413 ], "score": 1.0, "content": "interpretability-based (Dai et al., 2022), and meta-learning (Mitchell et al., 2021; De Cao et al., 2021)", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 411, 505, 425 ], "spans": [ { "bbox": [ 105, 411, 505, 425 ], "score": 1.0, "content": "methods, ROME achieves good generalization and specificity simultaneously, whereas previous", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 423, 257, 435 ], "spans": [ { "bbox": [ 106, 423, 257, 435 ], "score": 1.0, "content": "approaches sacrifice one or the other.", "type": "text" } ], "index": 16 } ], "index": 13 }, { "type": "title", "bbox": [ 106, 457, 424, 471 ], "lines": [ { "bbox": [ 104, 456, 425, 474 ], "spans": [ { "bbox": [ 104, 456, 425, 474 ], "score": 1.0, "content": "2 Interventions on Activations for Tracing Information Flow", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 106, 488, 505, 544 ], "lines": [ { "bbox": [ 105, 487, 505, 501 ], "spans": [ { "bbox": [ 105, 487, 505, 501 ], "score": 1.0, "content": "To locate facts within the parameters of a large pretrained autoregressive transformer, we begin by", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 500, 505, 511 ], "spans": [ { "bbox": [ 106, 500, 505, 511 ], "score": 1.0, "content": "analyzing and identifying the specific hidden states that have the strongest causal effect on predictions", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 510, 505, 523 ], "spans": [ { "bbox": [ 106, 510, 365, 523 ], "score": 1.0, "content": "of individual facts. We represent each fact as a knowledge tuple", "type": "text" }, { "bbox": [ 365, 510, 413, 522 ], "score": 0.92, "content": "t = ( s , r , o )", "type": "inline_equation" }, { "bbox": [ 413, 510, 505, 523 ], "score": 1.0, "content": "containing the subject", "type": "text" } ], "index": 20 }, { "bbox": [ 107, 520, 506, 533 ], "spans": [ { "bbox": [ 107, 524, 112, 531 ], "score": 0.6, "content": "s", "type": "inline_equation" }, { "bbox": [ 113, 520, 144, 533 ], "score": 1.0, "content": ", object", "type": "text" }, { "bbox": [ 144, 523, 150, 531 ], "score": 0.7, "content": "o", "type": "inline_equation" }, { "bbox": [ 150, 520, 204, 533 ], "score": 1.0, "content": ", and relation", "type": "text" }, { "bbox": [ 205, 523, 211, 531 ], "score": 0.73, "content": "r", "type": "inline_equation" }, { "bbox": [ 211, 520, 506, 533 ], "score": 1.0, "content": "connecting the two. Then to elicit the fact in GPT, we provide a natural", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 532, 417, 545 ], "spans": [ { "bbox": [ 105, 532, 176, 545 ], "score": 1.0, "content": "language prompt", "type": "text" }, { "bbox": [ 176, 534, 183, 544 ], "score": 0.78, "content": "p", "type": "inline_equation" }, { "bbox": [ 183, 532, 228, 545 ], "score": 1.0, "content": "describing", "type": "text" }, { "bbox": [ 228, 532, 250, 544 ], "score": 0.92, "content": "( s , r )", "type": "inline_equation" }, { "bbox": [ 251, 532, 407, 545 ], "score": 1.0, "content": "and examine the model’s prediction of", "type": "text" }, { "bbox": [ 407, 534, 413, 542 ], "score": 0.69, "content": "o", "type": "inline_equation" }, { "bbox": [ 414, 532, 417, 545 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 22 } ], "index": 20 }, { "type": "text", "bbox": [ 106, 548, 505, 603 ], "lines": [ { "bbox": [ 106, 547, 505, 561 ], "spans": [ { "bbox": [ 106, 547, 289, 561 ], "score": 1.0, "content": "An autoregressive transformer language model", "type": "text" }, { "bbox": [ 290, 549, 339, 559 ], "score": 0.91, "content": "G : \\mathcal { X } \\mathcal { Y }", "type": "inline_equation" }, { "bbox": [ 339, 547, 403, 561 ], "score": 1.0, "content": "over vocabulary", "type": "text" }, { "bbox": [ 404, 549, 413, 558 ], "score": 0.71, "content": "V", "type": "inline_equation" }, { "bbox": [ 413, 547, 505, 561 ], "score": 1.0, "content": "maps a token sequence", "type": "text" } ], "index": 23 }, { "bbox": [ 107, 558, 506, 572 ], "spans": [ { "bbox": [ 107, 559, 194, 571 ], "score": 0.91, "content": "x = [ x _ { 1 } , . . . , x _ { T } ] \\in \\mathcal { X }", "type": "inline_equation" }, { "bbox": [ 195, 558, 198, 572 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 199, 559, 230, 570 ], "score": 0.88, "content": "x _ { i } \\in V", "type": "inline_equation" }, { "bbox": [ 231, 558, 346, 572 ], "score": 1.0, "content": "to a probability distribution", "type": "text" }, { "bbox": [ 347, 558, 405, 570 ], "score": 0.92, "content": "y \\in \\mathcal { y } \\subset \\mathbb { R } ^ { | \\check { V } | }", "type": "inline_equation" }, { "bbox": [ 405, 558, 506, 572 ], "score": 1.0, "content": "that predicts next-token", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 569, 506, 583 ], "spans": [ { "bbox": [ 105, 569, 176, 583 ], "score": 1.0, "content": "continuations of", "type": "text" }, { "bbox": [ 176, 573, 182, 580 ], "score": 0.74, "content": "x", "type": "inline_equation" }, { "bbox": [ 183, 569, 506, 583 ], "score": 1.0, "content": ". Within the transformer, the ith token is embedded as a series of hidden state", "type": "text" } ], "index": 25 }, { "bbox": [ 104, 577, 507, 596 ], "spans": [ { "bbox": [ 104, 577, 137, 596 ], "score": 1.0, "content": "vectors", "type": "text" }, { "bbox": [ 137, 579, 153, 594 ], "score": 0.91, "content": "h _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 153, 577, 219, 596 ], "score": 1.0, "content": ", beginning with", "type": "text" }, { "bbox": [ 219, 579, 350, 594 ], "score": 0.93, "content": "h _ { i } ^ { ( 0 ) } = \\mathrm { e m b } ( x _ { i } ) + \\mathrm { p o s } ( i ) \\in \\mathbb { R } ^ { H }", "type": "inline_equation" }, { "bbox": [ 350, 579, 420, 595 ], "score": 1.0, "content": ". The final output", "type": "text" }, { "bbox": [ 420, 579, 494, 594 ], "score": 0.91, "content": "y = \\operatorname* { d e c o d e } ( h _ { T } ^ { ( L ) } )", "type": "inline_equation" }, { "bbox": [ 495, 579, 507, 595 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 591, 231, 604 ], "spans": [ { "bbox": [ 105, 591, 231, 604 ], "score": 1.0, "content": "read from the last hidden state.", "type": "text" } ], "index": 27 } ], "index": 25 }, { "type": "text", "bbox": [ 108, 608, 505, 653 ], "lines": [ { "bbox": [ 102, 605, 505, 624 ], "spans": [ { "bbox": [ 102, 605, 271, 624 ], "score": 1.0, "content": "We visualize the internal computation of", "type": "text" }, { "bbox": [ 271, 609, 280, 618 ], "score": 0.83, "content": "G", "type": "inline_equation" }, { "bbox": [ 280, 605, 429, 624 ], "score": 1.0, "content": "as a grid (Figure 1a) of hidden states", "type": "text" }, { "bbox": [ 430, 606, 446, 621 ], "score": 0.92, "content": "h _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 446, 605, 505, 624 ], "score": 1.0, "content": "in which each", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 617, 505, 633 ], "spans": [ { "bbox": [ 105, 619, 129, 633 ], "score": 1.0, "content": "layer", "type": "text" }, { "bbox": [ 129, 620, 134, 629 ], "score": 0.68, "content": "l", "type": "inline_equation" }, { "bbox": [ 137, 619, 190, 631 ], "score": 0.47, "content": "( \\mathrm { l e f t } \\to \\mathrm { r i g h t } )", "type": "inline_equation" }, { "bbox": [ 190, 619, 281, 633 ], "score": 1.0, "content": ") adds global attention", "type": "text" }, { "bbox": [ 282, 618, 297, 631 ], "score": 0.9, "content": "a _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 298, 619, 363, 633 ], "score": 1.0, "content": "and local MLP", "type": "text" }, { "bbox": [ 363, 617, 382, 632 ], "score": 0.9, "content": "m _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 382, 620, 505, 631 ], "score": 1.0, "content": "contributions computed from", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 630, 507, 642 ], "spans": [ { "bbox": [ 105, 630, 259, 642 ], "score": 1.0, "content": "previous layers, and where each token", "type": "text" }, { "bbox": [ 260, 631, 264, 640 ], "score": 0.71, "content": "i", "type": "inline_equation" }, { "bbox": [ 264, 630, 283, 642 ], "score": 1.0, "content": "(top", "type": "text" }, { "bbox": [ 284, 632, 296, 640 ], "score": 0.74, "content": "", "type": "inline_equation" }, { "bbox": [ 296, 630, 507, 642 ], "score": 1.0, "content": "bottom) attends to previous states from other tokens.", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 641, 486, 654 ], "spans": [ { "bbox": [ 106, 641, 486, 654 ], "score": 1.0, "content": "Recall that, in the autoregressive case, tokens only draw information from past (above) tokens:", "type": "text" } ], "index": 31 } ], "index": 29.5 }, { "type": "interline_equation", "bbox": [ 185, 659, 425, 722 ], "lines": [ { "bbox": [ 185, 659, 425, 722 ], "spans": [ { "bbox": [ 185, 659, 425, 722 ], "score": 0.93, "content": "\\begin{array} { r l } & { h _ { i } ^ { ( l ) } = h _ { i } ^ { ( l - 1 ) } + a _ { i } ^ { ( l ) } + m _ { i } ^ { ( l ) } } \\\\ & { ~ a _ { i } ^ { ( l ) } = \\mathrm { a t t n } ^ { ( l ) } \\left( h _ { 1 } ^ { ( l - 1 ) } , h _ { 2 } ^ { ( l - 1 ) } , \\ldots , h _ { i } ^ { ( l - 1 ) } \\right) } \\\\ & { ~ m _ { i } ^ { ( l ) } = W _ { p r o j } ^ { ( l ) } \\sigma \\left( W _ { f c } ^ { ( l ) } \\gamma \\left( a _ { i } ^ { ( l ) } + h _ { i } ^ { ( l - 1 ) } \\right) \\right) . } \\end{array}", "type": "interline_equation", "image_path": "ea9fb193121cb3d049c5a8aa9dc432a5bd1263c33768a13b1bc4cc8c5c2928bc.jpg" } ] } ], "index": 33.5, "virtual_lines": [ { "bbox": [ 185, 659, 425, 674.75 ], "spans": [], "index": 32 }, { "bbox": [ 185, 674.75, 425, 690.5 ], "spans": [], "index": 33 }, { "bbox": [ 185, 690.5, 425, 706.25 ], "spans": [], "index": 34 }, { "bbox": [ 185, 706.25, 425, 722.0 ], "spans": [], "index": 35 } ] } ], "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": "image", "bbox": [ 107, 69, 505, 251 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 69, 505, 251 ], "group_id": 0, "lines": [ { "bbox": [ 107, 69, 505, 251 ], "spans": [ { "bbox": [ 107, 69, 505, 251 ], "score": 0.956, "type": "image", "image_path": "7c0dd76794a726970bdc2ce46cc2c22679725ee4074a597abf7a6e36dff2b24d.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 69, 505, 129.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 107, 129.66666666666666, 505, 190.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 107, 190.33333333333331, 505, 250.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 256, 505, 307 ], "group_id": 0, "lines": [ { "bbox": [ 105, 254, 506, 268 ], "spans": [ { "bbox": [ 105, 254, 506, 268 ], "score": 1.0, "content": "Figure 1: Causal Traces compute the causal effect of neuron activations by running the network twice: (a)", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 266, 506, 277 ], "spans": [ { "bbox": [ 106, 266, 506, 277 ], "score": 1.0, "content": "once normally, and (b) once where we corrupt the subject token and then (c) restore selected internal activations", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 276, 505, 287 ], "spans": [ { "bbox": [ 106, 276, 505, 287 ], "score": 1.0, "content": "to their clean value. (d) Some sets of activations cause the output to return to the original prediction; the light", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 285, 505, 298 ], "spans": [ { "bbox": [ 105, 285, 505, 298 ], "score": 1.0, "content": "blue path shows an example of information flow. The causal impact on output probability is mapped for the", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 296, 496, 307 ], "spans": [ { "bbox": [ 106, 296, 496, 307 ], "score": 1.0, "content": "effect of (e) each hidden state on the prediction, (f) only MLP activations, and (g) only attention activations.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 106, 329, 504, 352 ], "lines": [ { "bbox": [ 106, 329, 506, 342 ], "spans": [ { "bbox": [ 106, 329, 506, 342 ], "score": 1.0, "content": "Despite the simplicity of the intervention, we find that ROME is similarly effective to other model-", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 341, 458, 354 ], "spans": [ { "bbox": [ 106, 341, 458, 354 ], "score": 1.0, "content": "editing approaches on a standard zero-shot relation extraction benchmark (Section 3.2).", "type": "text" } ], "index": 9 } ], "index": 8.5, "bbox_fs": [ 106, 329, 506, 354 ] }, { "type": "text", "bbox": [ 106, 357, 505, 434 ], "lines": [ { "bbox": [ 105, 356, 505, 369 ], "spans": [ { "bbox": [ 105, 356, 505, 369 ], "score": 1.0, "content": "To evaluate ROME’s impact on more difficult cases, we introduce a dataset of counterfactual assertions", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 368, 505, 381 ], "spans": [ { "bbox": [ 106, 368, 505, 381 ], "score": 1.0, "content": "(Section 3.3) that would not have been observed in pretraining. Our evaluations (Section 3.4) confirm", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 379, 505, 391 ], "spans": [ { "bbox": [ 105, 379, 505, 391 ], "score": 1.0, "content": "that midlayer MLP modules can store factual associations that generalize beyond specific surface", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 390, 506, 402 ], "spans": [ { "bbox": [ 106, 390, 506, 402 ], "score": 1.0, "content": "forms, while remaining specific to the subject. Compared to previous fine-tuning (Zhu et al., 2020),", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 401, 505, 413 ], "spans": [ { "bbox": [ 106, 401, 505, 413 ], "score": 1.0, "content": "interpretability-based (Dai et al., 2022), and meta-learning (Mitchell et al., 2021; De Cao et al., 2021)", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 411, 505, 425 ], "spans": [ { "bbox": [ 105, 411, 505, 425 ], "score": 1.0, "content": "methods, ROME achieves good generalization and specificity simultaneously, whereas previous", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 423, 257, 435 ], "spans": [ { "bbox": [ 106, 423, 257, 435 ], "score": 1.0, "content": "approaches sacrifice one or the other.", "type": "text" } ], "index": 16 } ], "index": 13, "bbox_fs": [ 105, 356, 506, 435 ] }, { "type": "title", "bbox": [ 106, 457, 424, 471 ], "lines": [ { "bbox": [ 104, 456, 425, 474 ], "spans": [ { "bbox": [ 104, 456, 425, 474 ], "score": 1.0, "content": "2 Interventions on Activations for Tracing Information Flow", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 106, 488, 505, 544 ], "lines": [ { "bbox": [ 105, 487, 505, 501 ], "spans": [ { "bbox": [ 105, 487, 505, 501 ], "score": 1.0, "content": "To locate facts within the parameters of a large pretrained autoregressive transformer, we begin by", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 500, 505, 511 ], "spans": [ { "bbox": [ 106, 500, 505, 511 ], "score": 1.0, "content": "analyzing and identifying the specific hidden states that have the strongest causal effect on predictions", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 510, 505, 523 ], "spans": [ { "bbox": [ 106, 510, 365, 523 ], "score": 1.0, "content": "of individual facts. We represent each fact as a knowledge tuple", "type": "text" }, { "bbox": [ 365, 510, 413, 522 ], "score": 0.92, "content": "t = ( s , r , o )", "type": "inline_equation" }, { "bbox": [ 413, 510, 505, 523 ], "score": 1.0, "content": "containing the subject", "type": "text" } ], "index": 20 }, { "bbox": [ 107, 520, 506, 533 ], "spans": [ { "bbox": [ 107, 524, 112, 531 ], "score": 0.6, "content": "s", "type": "inline_equation" }, { "bbox": [ 113, 520, 144, 533 ], "score": 1.0, "content": ", object", "type": "text" }, { "bbox": [ 144, 523, 150, 531 ], "score": 0.7, "content": "o", "type": "inline_equation" }, { "bbox": [ 150, 520, 204, 533 ], "score": 1.0, "content": ", and relation", "type": "text" }, { "bbox": [ 205, 523, 211, 531 ], "score": 0.73, "content": "r", "type": "inline_equation" }, { "bbox": [ 211, 520, 506, 533 ], "score": 1.0, "content": "connecting the two. Then to elicit the fact in GPT, we provide a natural", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 532, 417, 545 ], "spans": [ { "bbox": [ 105, 532, 176, 545 ], "score": 1.0, "content": "language prompt", "type": "text" }, { "bbox": [ 176, 534, 183, 544 ], "score": 0.78, "content": "p", "type": "inline_equation" }, { "bbox": [ 183, 532, 228, 545 ], "score": 1.0, "content": "describing", "type": "text" }, { "bbox": [ 228, 532, 250, 544 ], "score": 0.92, "content": "( s , r )", "type": "inline_equation" }, { "bbox": [ 251, 532, 407, 545 ], "score": 1.0, "content": "and examine the model’s prediction of", "type": "text" }, { "bbox": [ 407, 534, 413, 542 ], "score": 0.69, "content": "o", "type": "inline_equation" }, { "bbox": [ 414, 532, 417, 545 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 22 } ], "index": 20, "bbox_fs": [ 105, 487, 506, 545 ] }, { "type": "text", "bbox": [ 106, 548, 505, 603 ], "lines": [ { "bbox": [ 106, 547, 505, 561 ], "spans": [ { "bbox": [ 106, 547, 289, 561 ], "score": 1.0, "content": "An autoregressive transformer language model", "type": "text" }, { "bbox": [ 290, 549, 339, 559 ], "score": 0.91, "content": "G : \\mathcal { X } \\mathcal { Y }", "type": "inline_equation" }, { "bbox": [ 339, 547, 403, 561 ], "score": 1.0, "content": "over vocabulary", "type": "text" }, { "bbox": [ 404, 549, 413, 558 ], "score": 0.71, "content": "V", "type": "inline_equation" }, { "bbox": [ 413, 547, 505, 561 ], "score": 1.0, "content": "maps a token sequence", "type": "text" } ], "index": 23 }, { "bbox": [ 107, 558, 506, 572 ], "spans": [ { "bbox": [ 107, 559, 194, 571 ], "score": 0.91, "content": "x = [ x _ { 1 } , . . . , x _ { T } ] \\in \\mathcal { X }", "type": "inline_equation" }, { "bbox": [ 195, 558, 198, 572 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 199, 559, 230, 570 ], "score": 0.88, "content": "x _ { i } \\in V", "type": "inline_equation" }, { "bbox": [ 231, 558, 346, 572 ], "score": 1.0, "content": "to a probability distribution", "type": "text" }, { "bbox": [ 347, 558, 405, 570 ], "score": 0.92, "content": "y \\in \\mathcal { y } \\subset \\mathbb { R } ^ { | \\check { V } | }", "type": "inline_equation" }, { "bbox": [ 405, 558, 506, 572 ], "score": 1.0, "content": "that predicts next-token", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 569, 506, 583 ], "spans": [ { "bbox": [ 105, 569, 176, 583 ], "score": 1.0, "content": "continuations of", "type": "text" }, { "bbox": [ 176, 573, 182, 580 ], "score": 0.74, "content": "x", "type": "inline_equation" }, { "bbox": [ 183, 569, 506, 583 ], "score": 1.0, "content": ". Within the transformer, the ith token is embedded as a series of hidden state", "type": "text" } ], "index": 25 }, { "bbox": [ 104, 577, 507, 596 ], "spans": [ { "bbox": [ 104, 577, 137, 596 ], "score": 1.0, "content": "vectors", "type": "text" }, { "bbox": [ 137, 579, 153, 594 ], "score": 0.91, "content": "h _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 153, 577, 219, 596 ], "score": 1.0, "content": ", beginning with", "type": "text" }, { "bbox": [ 219, 579, 350, 594 ], "score": 0.93, "content": "h _ { i } ^ { ( 0 ) } = \\mathrm { e m b } ( x _ { i } ) + \\mathrm { p o s } ( i ) \\in \\mathbb { R } ^ { H }", "type": "inline_equation" }, { "bbox": [ 350, 579, 420, 595 ], "score": 1.0, "content": ". The final output", "type": "text" }, { "bbox": [ 420, 579, 494, 594 ], "score": 0.91, "content": "y = \\operatorname* { d e c o d e } ( h _ { T } ^ { ( L ) } )", "type": "inline_equation" }, { "bbox": [ 495, 579, 507, 595 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 591, 231, 604 ], "spans": [ { "bbox": [ 105, 591, 231, 604 ], "score": 1.0, "content": "read from the last hidden state.", "type": "text" } ], "index": 27 } ], "index": 25, "bbox_fs": [ 104, 547, 507, 604 ] }, { "type": "text", "bbox": [ 108, 608, 505, 653 ], "lines": [ { "bbox": [ 102, 605, 505, 624 ], "spans": [ { "bbox": [ 102, 605, 271, 624 ], "score": 1.0, "content": "We visualize the internal computation of", "type": "text" }, { "bbox": [ 271, 609, 280, 618 ], "score": 0.83, "content": "G", "type": "inline_equation" }, { "bbox": [ 280, 605, 429, 624 ], "score": 1.0, "content": "as a grid (Figure 1a) of hidden states", "type": "text" }, { "bbox": [ 430, 606, 446, 621 ], "score": 0.92, "content": "h _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 446, 605, 505, 624 ], "score": 1.0, "content": "in which each", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 617, 505, 633 ], "spans": [ { "bbox": [ 105, 619, 129, 633 ], "score": 1.0, "content": "layer", "type": "text" }, { "bbox": [ 129, 620, 134, 629 ], "score": 0.68, "content": "l", "type": "inline_equation" }, { "bbox": [ 137, 619, 190, 631 ], "score": 0.47, "content": "( \\mathrm { l e f t } \\to \\mathrm { r i g h t } )", "type": "inline_equation" }, { "bbox": [ 190, 619, 281, 633 ], "score": 1.0, "content": ") adds global attention", "type": "text" }, { "bbox": [ 282, 618, 297, 631 ], "score": 0.9, "content": "a _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 298, 619, 363, 633 ], "score": 1.0, "content": "and local MLP", "type": "text" }, { "bbox": [ 363, 617, 382, 632 ], "score": 0.9, "content": "m _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 382, 620, 505, 631 ], "score": 1.0, "content": "contributions computed from", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 630, 507, 642 ], "spans": [ { "bbox": [ 105, 630, 259, 642 ], "score": 1.0, "content": "previous layers, and where each token", "type": "text" }, { "bbox": [ 260, 631, 264, 640 ], "score": 0.71, "content": "i", "type": "inline_equation" }, { "bbox": [ 264, 630, 283, 642 ], "score": 1.0, "content": "(top", "type": "text" }, { "bbox": [ 284, 632, 296, 640 ], "score": 0.74, "content": "", "type": "inline_equation" }, { "bbox": [ 296, 630, 507, 642 ], "score": 1.0, "content": "bottom) attends to previous states from other tokens.", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 641, 486, 654 ], "spans": [ { "bbox": [ 106, 641, 486, 654 ], "score": 1.0, "content": "Recall that, in the autoregressive case, tokens only draw information from past (above) tokens:", "type": "text" } ], "index": 31 } ], "index": 29.5, "bbox_fs": [ 102, 605, 507, 654 ] }, { "type": "interline_equation", "bbox": [ 185, 659, 425, 722 ], "lines": [ { "bbox": [ 185, 659, 425, 722 ], "spans": [ { "bbox": [ 185, 659, 425, 722 ], "score": 0.93, "content": "\\begin{array} { r l } & { h _ { i } ^ { ( l ) } = h _ { i } ^ { ( l - 1 ) } + a _ { i } ^ { ( l ) } + m _ { i } ^ { ( l ) } } \\\\ & { ~ a _ { i } ^ { ( l ) } = \\mathrm { a t t n } ^ { ( l ) } \\left( h _ { 1 } ^ { ( l - 1 ) } , h _ { 2 } ^ { ( l - 1 ) } , \\ldots , h _ { i } ^ { ( l - 1 ) } \\right) } \\\\ & { ~ m _ { i } ^ { ( l ) } = W _ { p r o j } ^ { ( l ) } \\sigma \\left( W _ { f c } ^ { ( l ) } \\gamma \\left( a _ { i } ^ { ( l ) } + h _ { i } ^ { ( l - 1 ) } \\right) \\right) . } \\end{array}", "type": "interline_equation", "image_path": "ea9fb193121cb3d049c5a8aa9dc432a5bd1263c33768a13b1bc4cc8c5c2928bc.jpg" } ] } ], "index": 33.5, "virtual_lines": [ { "bbox": [ 185, 659, 425, 674.75 ], "spans": [], "index": 32 }, { "bbox": [ 185, 674.75, 425, 690.5 ], "spans": [], "index": 33 }, { "bbox": [ 185, 690.5, 425, 706.25 ], "spans": [], "index": 34 }, { "bbox": [ 185, 706.25, 425, 722.0 ], "spans": [], "index": 35 } ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 106, 70, 506, 154 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 70, 506, 154 ], "group_id": 0, "lines": [ { "bbox": [ 106, 70, 506, 154 ], "spans": [ { "bbox": [ 106, 70, 506, 154 ], "score": 0.615, "type": "image", "image_path": "ff44c5a930650e71d33fcf1fed31a47d79ac39dabf784d6ac96d862851815a6c.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 70, 506, 98.0 ], "spans": [], "index": 0 }, { "bbox": [ 106, 98.0, 506, 126.0 ], "spans": [], "index": 1 }, { "bbox": [ 106, 126.0, 506, 154.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 160, 505, 210 ], "group_id": 0, "lines": [ { "bbox": [ 106, 160, 505, 172 ], "spans": [ { "bbox": [ 106, 160, 505, 172 ], "score": 1.0, "content": "Figure 2: Average Indirect Effect of individual model components over a sample of 1000 factual statements", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 169, 506, 183 ], "spans": [ { "bbox": [ 105, 169, 506, 183 ], "score": 1.0, "content": "reveals two important sites. (a) Strong causality at a ‘late site’ in the last layers at the last token is unsurprising,", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 181, 505, 192 ], "spans": [ { "bbox": [ 106, 181, 505, 192 ], "score": 1.0, "content": "but strongly causal states at an ‘early site’ in middle layers at the last subject token is a new discovery. (b) MLP", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 190, 505, 203 ], "spans": [ { "bbox": [ 105, 190, 505, 203 ], "score": 1.0, "content": "contributions dominate the early site. (c) Attention is important at the late site. Appendix B, Figure 7 shows", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 200, 321, 212 ], "spans": [ { "bbox": [ 106, 200, 226, 212 ], "score": 1.0, "content": "these heatmaps as line plots with", "type": "text" }, { "bbox": [ 226, 200, 244, 210 ], "score": 0.87, "content": "9 5 \\%", "type": "inline_equation" }, { "bbox": [ 245, 200, 321, 212 ], "score": 1.0, "content": "confidence intervals.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 107, 232, 505, 266 ], "lines": [ { "bbox": [ 103, 229, 505, 249 ], "spans": [ { "bbox": [ 103, 229, 439, 249 ], "score": 1.0, "content": "Each layer’s MLP is a two-layer neural network parameterized by matrices W (l)proj", "type": "text" }, { "bbox": [ 441, 232, 459, 247 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 459, 231, 480, 247 ], "score": 0.92, "content": "W _ { f c } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 480, 234, 505, 244 ], "score": 1.0, "content": ", with", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 243, 507, 258 ], "spans": [ { "bbox": [ 105, 243, 200, 258 ], "score": 1.0, "content": "rectifying nonlinearity", "type": "text" }, { "bbox": [ 200, 247, 207, 254 ], "score": 0.76, "content": "\\sigma", "type": "inline_equation" }, { "bbox": [ 208, 243, 329, 258 ], "score": 1.0, "content": "and normalizing nonlinearity", "type": "text" }, { "bbox": [ 329, 246, 336, 256 ], "score": 0.8, "content": "\\gamma", "type": "inline_equation" }, { "bbox": [ 336, 243, 507, 258 ], "score": 1.0, "content": ". For further background on transformers,", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 255, 245, 267 ], "spans": [ { "bbox": [ 105, 255, 245, 267 ], "score": 1.0, "content": "we refer to Vaswani et al. (2017).3", "type": "text" } ], "index": 10 } ], "index": 9 }, { "type": "title", "bbox": [ 108, 280, 295, 292 ], "lines": [ { "bbox": [ 105, 280, 297, 294 ], "spans": [ { "bbox": [ 105, 280, 297, 294 ], "score": 1.0, "content": "2.1 Causal Tracing of Factual Associations", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 301, 505, 345 ], "lines": [ { "bbox": [ 105, 300, 505, 313 ], "spans": [ { "bbox": [ 105, 300, 505, 313 ], "score": 1.0, "content": "The grid of states (Figure 1) forms a causal graph (Pearl, 2009) describing dependencies between the", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 311, 505, 324 ], "spans": [ { "bbox": [ 105, 311, 505, 324 ], "score": 1.0, "content": "hidden variables. This graph contains many paths from inputs on the left to the output (next-word", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 323, 505, 335 ], "spans": [ { "bbox": [ 105, 323, 505, 335 ], "score": 1.0, "content": "prediction) at the lower-right, and we wish to understand if there are specific hidden state variables", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 334, 336, 345 ], "spans": [ { "bbox": [ 106, 334, 336, 345 ], "score": 1.0, "content": "that are more important than others when recalling a fact.", "type": "text" } ], "index": 15 } ], "index": 13.5 }, { "type": "text", "bbox": [ 106, 350, 505, 406 ], "lines": [ { "bbox": [ 106, 351, 505, 361 ], "spans": [ { "bbox": [ 106, 351, 505, 361 ], "score": 1.0, "content": "As Vig et al. (2020b) have shown, this is a natural case for causal mediation analysis, which quantifies", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 361, 505, 373 ], "spans": [ { "bbox": [ 106, 361, 505, 373 ], "score": 1.0, "content": "the contribution of intermediate variables in causal graphs (Pearl, 2001). To calculate each state’s", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 370, 505, 386 ], "spans": [ { "bbox": [ 105, 370, 379, 386 ], "score": 1.0, "content": "contribution towards a correct factual prediction, we observe all of", "type": "text" }, { "bbox": [ 379, 372, 388, 382 ], "score": 0.83, "content": "G", "type": "inline_equation" }, { "bbox": [ 389, 370, 505, 386 ], "score": 1.0, "content": "’s internal activations during", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 383, 505, 395 ], "spans": [ { "bbox": [ 106, 383, 505, 395 ], "score": 1.0, "content": "three runs: a clean run that predicts the fact, a corrupted run where the prediction is damaged, and a", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 395, 483, 406 ], "spans": [ { "bbox": [ 106, 395, 483, 406 ], "score": 1.0, "content": "corrupted-with-restoration run that tests the ability of a single state to restore the prediction.", "type": "text" } ], "index": 20 } ], "index": 18 }, { "type": "text", "bbox": [ 106, 410, 506, 559 ], "lines": [ { "bbox": [ 104, 409, 506, 423 ], "spans": [ { "bbox": [ 104, 409, 314, 423 ], "score": 1.0, "content": "• In the clean run, we pass a factual prompt", "type": "text" }, { "bbox": [ 314, 412, 321, 420 ], "score": 0.75, "content": "x", "type": "inline_equation" }, { "bbox": [ 322, 409, 345, 423 ], "score": 1.0, "content": "into", "type": "text" }, { "bbox": [ 346, 411, 355, 420 ], "score": 0.81, "content": "G", "type": "inline_equation" }, { "bbox": [ 356, 409, 506, 423 ], "score": 1.0, "content": "and collect all hidden activations", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 420, 506, 434 ], "spans": [ { "bbox": [ 115, 420, 228, 433 ], "score": 0.91, "content": "\\{ h _ { i } ^ { ( l ) } \\ | \\ i \\in [ 1 , T ] , l \\in [ 1 , \\dot { L } ] \\}", "type": "inline_equation" }, { "bbox": [ 229, 420, 506, 434 ], "score": 1.0, "content": ". Figure 1a provides an example illustration with the prompt: “The", "type": "text" } ], "index": 22 }, { "bbox": [ 114, 431, 483, 445 ], "spans": [ { "bbox": [ 114, 432, 237, 444 ], "score": 1.0, "content": "Space Needle is in downtown", "type": "text" }, { "bbox": [ 262, 431, 423, 445 ], "score": 1.0, "content": "”, for which the expected completion is", "type": "text" }, { "bbox": [ 423, 433, 478, 443 ], "score": 0.42, "content": "o = { } ^ { \\mathrm { * } } \\mathrm { S e a t t l e } ^ { \\mathrm { * } }", "type": "inline_equation" }, { "bbox": [ 478, 431, 483, 445 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 23 }, { "bbox": [ 108, 446, 506, 458 ], "spans": [ { "bbox": [ 108, 446, 370, 458 ], "score": 1.0, "content": "• In the baseline corrupted run, the subject is obfuscated from", "type": "text" }, { "bbox": [ 370, 446, 379, 456 ], "score": 0.73, "content": "G", "type": "inline_equation" }, { "bbox": [ 379, 446, 506, 458 ], "score": 1.0, "content": "before the network runs. Con-", "type": "text" } ], "index": 24 }, { "bbox": [ 111, 454, 505, 471 ], "spans": [ { "bbox": [ 111, 456, 221, 471 ], "score": 1.0, "content": "cretely, immediately after", "type": "text" }, { "bbox": [ 222, 459, 229, 467 ], "score": 0.76, "content": "x", "type": "inline_equation" }, { "bbox": [ 229, 456, 295, 471 ], "score": 1.0, "content": "is embedded as", "type": "text" }, { "bbox": [ 295, 455, 378, 469 ], "score": 0.89, "content": "[ h _ { 1 } ^ { ( 0 ) } , h _ { 2 } ^ { ( 0 ) } , . . . , h _ { T } ^ { ( 0 ) } ]", "type": "inline_equation" }, { "bbox": [ 378, 454, 410, 471 ], "score": 1.0, "content": ", we set", "type": "text" }, { "bbox": [ 410, 455, 477, 469 ], "score": 0.92, "content": "h _ { i } ^ { ( 0 ) } : = h _ { i } ^ { ( 0 ) } + \\epsilon", "type": "inline_equation" }, { "bbox": [ 477, 454, 505, 471 ], "score": 1.0, "content": "for all", "type": "text" } ], "index": 25 }, { "bbox": [ 113, 466, 506, 481 ], "spans": [ { "bbox": [ 113, 466, 145, 481 ], "score": 1.0, "content": "indices", "type": "text" }, { "bbox": [ 145, 469, 150, 478 ], "score": 0.69, "content": "i", "type": "inline_equation" }, { "bbox": [ 150, 466, 323, 481 ], "score": 1.0, "content": "that correspond to the subject entity, where", "type": "text" }, { "bbox": [ 323, 468, 395, 479 ], "score": 0.7, "content": "\\epsilon \\sim \\mathcal { N } ( 0 ; \\bar { \\nu } ) ^ { 4 } ; . \\ : G", "type": "inline_equation" }, { "bbox": [ 396, 466, 506, 481 ], "score": 1.0, "content": "is then allowed to continue", "type": "text" } ], "index": 26 }, { "bbox": [ 110, 473, 506, 498 ], "spans": [ { "bbox": [ 110, 473, 317, 498 ], "score": 1.0, "content": "normally, giving us a set of corrupted activations", "type": "text" }, { "bbox": [ 317, 478, 430, 491 ], "score": 0.9, "content": "\\{ h _ { i * } ^ { ( l ) } \\ | \\ i \\in [ 1 , T ] , l \\in [ 1 , L ] \\}", "type": "inline_equation" }, { "bbox": [ 430, 473, 471, 498 ], "score": 1.0, "content": ". Because", "type": "text" }, { "bbox": [ 472, 479, 481, 488 ], "score": 0.79, "content": "G", "type": "inline_equation" }, { "bbox": [ 481, 473, 506, 498 ], "score": 1.0, "content": "loses", "type": "text" } ], "index": 27 }, { "bbox": [ 114, 490, 468, 502 ], "spans": [ { "bbox": [ 114, 490, 468, 502 ], "score": 1.0, "content": "some information about the subject, it will likely return an incorrect answer (Figure 1b).", "type": "text" } ], "index": 28 }, { "bbox": [ 107, 503, 505, 515 ], "spans": [ { "bbox": [ 107, 503, 286, 515 ], "score": 1.0, "content": "• The corrupted-with-restoration run, lets", "type": "text" }, { "bbox": [ 286, 504, 295, 513 ], "score": 0.76, "content": "G", "type": "inline_equation" }, { "bbox": [ 296, 503, 505, 515 ], "score": 1.0, "content": "run computations on the noisy embeddings as in the", "type": "text" } ], "index": 29 }, { "bbox": [ 114, 513, 505, 527 ], "spans": [ { "bbox": [ 114, 515, 283, 527 ], "score": 1.0, "content": "corrupted baseline, except at some token", "type": "text" }, { "bbox": [ 284, 513, 289, 524 ], "score": 0.73, "content": "\\hat { i }", "type": "inline_equation" }, { "bbox": [ 289, 515, 330, 527 ], "score": 1.0, "content": "and layer", "type": "text" }, { "bbox": [ 331, 514, 335, 524 ], "score": 0.63, "content": "\\hat { l }", "type": "inline_equation" }, { "bbox": [ 336, 515, 407, 527 ], "score": 1.0, "content": ". There, we hook", "type": "text" }, { "bbox": [ 407, 515, 416, 525 ], "score": 0.81, "content": "G", "type": "inline_equation" }, { "bbox": [ 417, 515, 505, 527 ], "score": 1.0, "content": "so that it is forced to", "type": "text" } ], "index": 30 }, { "bbox": [ 113, 523, 506, 539 ], "spans": [ { "bbox": [ 113, 523, 200, 538 ], "score": 1.0, "content": "output the clean state", "type": "text" }, { "bbox": [ 200, 524, 216, 539 ], "score": 0.89, "content": "h _ { \\widehat { i } } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 216, 523, 506, 538 ], "score": 1.0, "content": "; future computations execute without further intervention. Intuitively, the", "type": "text" } ], "index": 31 }, { "bbox": [ 113, 534, 506, 550 ], "spans": [ { "bbox": [ 113, 534, 506, 550 ], "score": 1.0, "content": "i ability of a few clean states to recover the correct fact, despite many other states being corrupted by", "type": "text" } ], "index": 32 }, { "bbox": [ 114, 547, 460, 560 ], "spans": [ { "bbox": [ 114, 547, 460, 560 ], "score": 1.0, "content": "the obfuscated subject, will indicate their causal importance in the computation graph.", "type": "text" } ], "index": 33 } ], "index": 27 }, { "type": "text", "bbox": [ 106, 563, 505, 651 ], "lines": [ { "bbox": [ 104, 562, 507, 578 ], "spans": [ { "bbox": [ 104, 562, 123, 578 ], "score": 1.0, "content": "Let", "type": "text" }, { "bbox": [ 123, 563, 167, 576 ], "score": 0.47, "content": "\\mathbb { P } [ o ] , \\mathbb { P } _ { * } [ o ]", "type": "inline_equation" }, { "bbox": [ 167, 562, 188, 578 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 188, 564, 200, 575 ], "score": 0.82, "content": "\\mathbb { P } _ { * }", "type": "inline_equation" }, { "bbox": [ 200, 562, 219, 578 ], "score": 1.0, "content": ", clean", "type": "text" }, { "bbox": [ 220, 563, 244, 577 ], "score": 0.86, "content": "h _ { i } ^ { ( l ) } \\left[ O \\right]", "type": "inline_equation" }, { "bbox": [ 245, 562, 386, 578 ], "score": 1.0, "content": "denote the probability of emitting", "type": "text" }, { "bbox": [ 387, 566, 393, 574 ], "score": 0.64, "content": "o", "type": "inline_equation" }, { "bbox": [ 393, 562, 507, 578 ], "score": 1.0, "content": "under the clean, corrupted,", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 574, 506, 586 ], "spans": [ { "bbox": [ 105, 574, 400, 586 ], "score": 1.0, "content": "and corrupted-with-restoration runs, respectively; dependence on the input", "type": "text" }, { "bbox": [ 400, 576, 407, 584 ], "score": 0.77, "content": "x", "type": "inline_equation" }, { "bbox": [ 407, 574, 506, 586 ], "score": 1.0, "content": "is omitted for notational", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 585, 506, 598 ], "spans": [ { "bbox": [ 105, 585, 423, 598 ], "score": 1.0, "content": "simplicity. The total effect (TE) is the difference between these quantities:", "type": "text" }, { "bbox": [ 423, 585, 503, 597 ], "score": 0.92, "content": "\\mathrm { T E } = \\mathbb { P } [ o ] - \\mathbb { P } _ { * } [ o ]", "type": "inline_equation" }, { "bbox": [ 503, 585, 506, 598 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 595, 506, 609 ], "spans": [ { "bbox": [ 105, 596, 324, 609 ], "score": 1.0, "content": "The indirect effect (IE) of a specific mediating state", "type": "text" }, { "bbox": [ 324, 595, 340, 608 ], "score": 0.91, "content": "h _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 340, 596, 506, 609 ], "score": 1.0, "content": "is defined as the difference between the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 607, 506, 619 ], "spans": [ { "bbox": [ 105, 607, 165, 619 ], "score": 1.0, "content": "probability of", "type": "text" }, { "bbox": [ 165, 609, 172, 617 ], "score": 0.7, "content": "o", "type": "inline_equation" }, { "bbox": [ 172, 607, 506, 619 ], "score": 1.0, "content": "under the corrupted version and the probability when that state is set to its clean", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 618, 506, 631 ], "spans": [ { "bbox": [ 105, 618, 288, 631 ], "score": 1.0, "content": "version, while the subject remains corrupted:", "type": "text" }, { "bbox": [ 288, 618, 322, 629 ], "score": 0.88, "content": "\\mathrm { I E } = \\mathbb { P } _ { * }", "type": "inline_equation" }, { "bbox": [ 322, 618, 342, 631 ], "score": 1.0, "content": ", clean", "type": "text" }, { "bbox": [ 342, 618, 399, 631 ], "score": 0.86, "content": "h _ { i } ^ { ( l ) } \\left[ O \\right] - \\mathbb { P } _ { * } [ O ]", "type": "inline_equation" }, { "bbox": [ 400, 618, 506, 631 ], "score": 1.0, "content": ". Averaging over a sample", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 629, 505, 641 ], "spans": [ { "bbox": [ 105, 629, 505, 641 ], "score": 1.0, "content": "of statements, we obtain the average total effect (ATE) and average indirect effect (AIE) for each", "type": "text" } ], "index": 40 }, { "bbox": [ 104, 639, 198, 651 ], "spans": [ { "bbox": [ 104, 639, 198, 651 ], "score": 1.0, "content": "hidden state variable.5", "type": "text" } ], "index": 41 } ], "index": 37.5 } ], "page_idx": 2, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 659, 505, 722 ], "lines": [ { "bbox": [ 105, 658, 505, 672 ], "spans": [ { "bbox": [ 105, 658, 505, 672 ], "score": 1.0, "content": "3 Eqn. 1 calculates attention sequentially after the MLP module as in Brown et al. (2020). Our methods also", "type": "text" } ] }, { "bbox": [ 107, 669, 473, 681 ], "spans": [ { "bbox": [ 107, 669, 473, 681 ], "score": 1.0, "content": "apply to GPT variants such as Wang & Komatsuzaki (2021) that put attention in parallel to the MLP.", "type": "text" } ] }, { "bbox": [ 106, 679, 505, 693 ], "spans": [ { "bbox": [ 106, 679, 147, 693 ], "score": 1.0, "content": "4 We select", "type": "text" }, { "bbox": [ 147, 682, 154, 690 ], "score": 0.65, "content": "\\nu", "type": "inline_equation" }, { "bbox": [ 154, 679, 505, 693 ], "score": 1.0, "content": "to be 3 times larger than the empirical standard deviation of embeddings; see Appendix B.1 for", "type": "text" } ] }, { "bbox": [ 108, 689, 366, 703 ], "spans": [ { "bbox": [ 108, 689, 366, 703 ], "score": 1.0, "content": "details, and see Appendix B.4 for an analysis of other corruption rules.", "type": "text" } ] }, { "bbox": [ 105, 700, 505, 713 ], "spans": [ { "bbox": [ 105, 700, 505, 713 ], "score": 1.0, "content": "5 One could also compute the direct effect, which flows through other model components besides the chosen", "type": "text" } ] }, { "bbox": [ 109, 712, 504, 723 ], "spans": [ { "bbox": [ 109, 712, 504, 723 ], "score": 1.0, "content": "mediator. However, we found this effect to be noisy and uninformative, in line with results by Vig et al. (2020b).", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 741, 309, 750 ], "lines": [ { "bbox": [ 301, 740, 310, 752 ], "spans": [ { "bbox": [ 301, 740, 310, 752 ], "score": 1.0, "content": "3", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 106, 70, 506, 154 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 70, 506, 154 ], "group_id": 0, "lines": [ { "bbox": [ 106, 70, 506, 154 ], "spans": [ { "bbox": [ 106, 70, 506, 154 ], "score": 0.615, "type": "image", "image_path": "ff44c5a930650e71d33fcf1fed31a47d79ac39dabf784d6ac96d862851815a6c.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 70, 506, 98.0 ], "spans": [], "index": 0 }, { "bbox": [ 106, 98.0, 506, 126.0 ], "spans": [], "index": 1 }, { "bbox": [ 106, 126.0, 506, 154.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 160, 505, 210 ], "group_id": 0, "lines": [ { "bbox": [ 106, 160, 505, 172 ], "spans": [ { "bbox": [ 106, 160, 505, 172 ], "score": 1.0, "content": "Figure 2: Average Indirect Effect of individual model components over a sample of 1000 factual statements", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 169, 506, 183 ], "spans": [ { "bbox": [ 105, 169, 506, 183 ], "score": 1.0, "content": "reveals two important sites. (a) Strong causality at a ‘late site’ in the last layers at the last token is unsurprising,", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 181, 505, 192 ], "spans": [ { "bbox": [ 106, 181, 505, 192 ], "score": 1.0, "content": "but strongly causal states at an ‘early site’ in middle layers at the last subject token is a new discovery. (b) MLP", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 190, 505, 203 ], "spans": [ { "bbox": [ 105, 190, 505, 203 ], "score": 1.0, "content": "contributions dominate the early site. (c) Attention is important at the late site. Appendix B, Figure 7 shows", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 200, 321, 212 ], "spans": [ { "bbox": [ 106, 200, 226, 212 ], "score": 1.0, "content": "these heatmaps as line plots with", "type": "text" }, { "bbox": [ 226, 200, 244, 210 ], "score": 0.87, "content": "9 5 \\%", "type": "inline_equation" }, { "bbox": [ 245, 200, 321, 212 ], "score": 1.0, "content": "confidence intervals.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 107, 232, 505, 266 ], "lines": [ { "bbox": [ 103, 229, 505, 249 ], "spans": [ { "bbox": [ 103, 229, 439, 249 ], "score": 1.0, "content": "Each layer’s MLP is a two-layer neural network parameterized by matrices W (l)proj", "type": "text" }, { "bbox": [ 441, 232, 459, 247 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 459, 231, 480, 247 ], "score": 0.92, "content": "W _ { f c } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 480, 234, 505, 244 ], "score": 1.0, "content": ", with", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 243, 507, 258 ], "spans": [ { "bbox": [ 105, 243, 200, 258 ], "score": 1.0, "content": "rectifying nonlinearity", "type": "text" }, { "bbox": [ 200, 247, 207, 254 ], "score": 0.76, "content": "\\sigma", "type": "inline_equation" }, { "bbox": [ 208, 243, 329, 258 ], "score": 1.0, "content": "and normalizing nonlinearity", "type": "text" }, { "bbox": [ 329, 246, 336, 256 ], "score": 0.8, "content": "\\gamma", "type": "inline_equation" }, { "bbox": [ 336, 243, 507, 258 ], "score": 1.0, "content": ". For further background on transformers,", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 255, 245, 267 ], "spans": [ { "bbox": [ 105, 255, 245, 267 ], "score": 1.0, "content": "we refer to Vaswani et al. (2017).3", "type": "text" } ], "index": 10 } ], "index": 9, "bbox_fs": [ 103, 229, 507, 267 ] }, { "type": "title", "bbox": [ 108, 280, 295, 292 ], "lines": [ { "bbox": [ 105, 280, 297, 294 ], "spans": [ { "bbox": [ 105, 280, 297, 294 ], "score": 1.0, "content": "2.1 Causal Tracing of Factual Associations", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 301, 505, 345 ], "lines": [ { "bbox": [ 105, 300, 505, 313 ], "spans": [ { "bbox": [ 105, 300, 505, 313 ], "score": 1.0, "content": "The grid of states (Figure 1) forms a causal graph (Pearl, 2009) describing dependencies between the", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 311, 505, 324 ], "spans": [ { "bbox": [ 105, 311, 505, 324 ], "score": 1.0, "content": "hidden variables. This graph contains many paths from inputs on the left to the output (next-word", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 323, 505, 335 ], "spans": [ { "bbox": [ 105, 323, 505, 335 ], "score": 1.0, "content": "prediction) at the lower-right, and we wish to understand if there are specific hidden state variables", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 334, 336, 345 ], "spans": [ { "bbox": [ 106, 334, 336, 345 ], "score": 1.0, "content": "that are more important than others when recalling a fact.", "type": "text" } ], "index": 15 } ], "index": 13.5, "bbox_fs": [ 105, 300, 505, 345 ] }, { "type": "text", "bbox": [ 106, 350, 505, 406 ], "lines": [ { "bbox": [ 106, 351, 505, 361 ], "spans": [ { "bbox": [ 106, 351, 505, 361 ], "score": 1.0, "content": "As Vig et al. (2020b) have shown, this is a natural case for causal mediation analysis, which quantifies", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 361, 505, 373 ], "spans": [ { "bbox": [ 106, 361, 505, 373 ], "score": 1.0, "content": "the contribution of intermediate variables in causal graphs (Pearl, 2001). To calculate each state’s", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 370, 505, 386 ], "spans": [ { "bbox": [ 105, 370, 379, 386 ], "score": 1.0, "content": "contribution towards a correct factual prediction, we observe all of", "type": "text" }, { "bbox": [ 379, 372, 388, 382 ], "score": 0.83, "content": "G", "type": "inline_equation" }, { "bbox": [ 389, 370, 505, 386 ], "score": 1.0, "content": "’s internal activations during", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 383, 505, 395 ], "spans": [ { "bbox": [ 106, 383, 505, 395 ], "score": 1.0, "content": "three runs: a clean run that predicts the fact, a corrupted run where the prediction is damaged, and a", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 395, 483, 406 ], "spans": [ { "bbox": [ 106, 395, 483, 406 ], "score": 1.0, "content": "corrupted-with-restoration run that tests the ability of a single state to restore the prediction.", "type": "text" } ], "index": 20 } ], "index": 18, "bbox_fs": [ 105, 351, 505, 406 ] }, { "type": "text", "bbox": [ 106, 410, 506, 559 ], "lines": [ { "bbox": [ 104, 409, 506, 423 ], "spans": [ { "bbox": [ 104, 409, 314, 423 ], "score": 1.0, "content": "• In the clean run, we pass a factual prompt", "type": "text" }, { "bbox": [ 314, 412, 321, 420 ], "score": 0.75, "content": "x", "type": "inline_equation" }, { "bbox": [ 322, 409, 345, 423 ], "score": 1.0, "content": "into", "type": "text" }, { "bbox": [ 346, 411, 355, 420 ], "score": 0.81, "content": "G", "type": "inline_equation" }, { "bbox": [ 356, 409, 506, 423 ], "score": 1.0, "content": "and collect all hidden activations", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 420, 506, 434 ], "spans": [ { "bbox": [ 115, 420, 228, 433 ], "score": 0.91, "content": "\\{ h _ { i } ^ { ( l ) } \\ | \\ i \\in [ 1 , T ] , l \\in [ 1 , \\dot { L } ] \\}", "type": "inline_equation" }, { "bbox": [ 229, 420, 506, 434 ], "score": 1.0, "content": ". Figure 1a provides an example illustration with the prompt: “The", "type": "text" } ], "index": 22 }, { "bbox": [ 114, 431, 483, 445 ], "spans": [ { "bbox": [ 114, 432, 237, 444 ], "score": 1.0, "content": "Space Needle is in downtown", "type": "text" }, { "bbox": [ 262, 431, 423, 445 ], "score": 1.0, "content": "”, for which the expected completion is", "type": "text" }, { "bbox": [ 423, 433, 478, 443 ], "score": 0.42, "content": "o = { } ^ { \\mathrm { * } } \\mathrm { S e a t t l e } ^ { \\mathrm { * } }", "type": "inline_equation" }, { "bbox": [ 478, 431, 483, 445 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 23 }, { "bbox": [ 108, 446, 506, 458 ], "spans": [ { "bbox": [ 108, 446, 370, 458 ], "score": 1.0, "content": "• In the baseline corrupted run, the subject is obfuscated from", "type": "text" }, { "bbox": [ 370, 446, 379, 456 ], "score": 0.73, "content": "G", "type": "inline_equation" }, { "bbox": [ 379, 446, 506, 458 ], "score": 1.0, "content": "before the network runs. Con-", "type": "text" } ], "index": 24 }, { "bbox": [ 111, 454, 505, 471 ], "spans": [ { "bbox": [ 111, 456, 221, 471 ], "score": 1.0, "content": "cretely, immediately after", "type": "text" }, { "bbox": [ 222, 459, 229, 467 ], "score": 0.76, "content": "x", "type": "inline_equation" }, { "bbox": [ 229, 456, 295, 471 ], "score": 1.0, "content": "is embedded as", "type": "text" }, { "bbox": [ 295, 455, 378, 469 ], "score": 0.89, "content": "[ h _ { 1 } ^ { ( 0 ) } , h _ { 2 } ^ { ( 0 ) } , . . . , h _ { T } ^ { ( 0 ) } ]", "type": "inline_equation" }, { "bbox": [ 378, 454, 410, 471 ], "score": 1.0, "content": ", we set", "type": "text" }, { "bbox": [ 410, 455, 477, 469 ], "score": 0.92, "content": "h _ { i } ^ { ( 0 ) } : = h _ { i } ^ { ( 0 ) } + \\epsilon", "type": "inline_equation" }, { "bbox": [ 477, 454, 505, 471 ], "score": 1.0, "content": "for all", "type": "text" } ], "index": 25 }, { "bbox": [ 113, 466, 506, 481 ], "spans": [ { "bbox": [ 113, 466, 145, 481 ], "score": 1.0, "content": "indices", "type": "text" }, { "bbox": [ 145, 469, 150, 478 ], "score": 0.69, "content": "i", "type": "inline_equation" }, { "bbox": [ 150, 466, 323, 481 ], "score": 1.0, "content": "that correspond to the subject entity, where", "type": "text" }, { "bbox": [ 323, 468, 395, 479 ], "score": 0.7, "content": "\\epsilon \\sim \\mathcal { N } ( 0 ; \\bar { \\nu } ) ^ { 4 } ; . \\ : G", "type": "inline_equation" }, { "bbox": [ 396, 466, 506, 481 ], "score": 1.0, "content": "is then allowed to continue", "type": "text" } ], "index": 26 }, { "bbox": [ 110, 473, 506, 498 ], "spans": [ { "bbox": [ 110, 473, 317, 498 ], "score": 1.0, "content": "normally, giving us a set of corrupted activations", "type": "text" }, { "bbox": [ 317, 478, 430, 491 ], "score": 0.9, "content": "\\{ h _ { i * } ^ { ( l ) } \\ | \\ i \\in [ 1 , T ] , l \\in [ 1 , L ] \\}", "type": "inline_equation" }, { "bbox": [ 430, 473, 471, 498 ], "score": 1.0, "content": ". Because", "type": "text" }, { "bbox": [ 472, 479, 481, 488 ], "score": 0.79, "content": "G", "type": "inline_equation" }, { "bbox": [ 481, 473, 506, 498 ], "score": 1.0, "content": "loses", "type": "text" } ], "index": 27 }, { "bbox": [ 114, 490, 468, 502 ], "spans": [ { "bbox": [ 114, 490, 468, 502 ], "score": 1.0, "content": "some information about the subject, it will likely return an incorrect answer (Figure 1b).", "type": "text" } ], "index": 28 }, { "bbox": [ 107, 503, 505, 515 ], "spans": [ { "bbox": [ 107, 503, 286, 515 ], "score": 1.0, "content": "• The corrupted-with-restoration run, lets", "type": "text" }, { "bbox": [ 286, 504, 295, 513 ], "score": 0.76, "content": "G", "type": "inline_equation" }, { "bbox": [ 296, 503, 505, 515 ], "score": 1.0, "content": "run computations on the noisy embeddings as in the", "type": "text" } ], "index": 29 }, { "bbox": [ 114, 513, 505, 527 ], "spans": [ { "bbox": [ 114, 515, 283, 527 ], "score": 1.0, "content": "corrupted baseline, except at some token", "type": "text" }, { "bbox": [ 284, 513, 289, 524 ], "score": 0.73, "content": "\\hat { i }", "type": "inline_equation" }, { "bbox": [ 289, 515, 330, 527 ], "score": 1.0, "content": "and layer", "type": "text" }, { "bbox": [ 331, 514, 335, 524 ], "score": 0.63, "content": "\\hat { l }", "type": "inline_equation" }, { "bbox": [ 336, 515, 407, 527 ], "score": 1.0, "content": ". There, we hook", "type": "text" }, { "bbox": [ 407, 515, 416, 525 ], "score": 0.81, "content": "G", "type": "inline_equation" }, { "bbox": [ 417, 515, 505, 527 ], "score": 1.0, "content": "so that it is forced to", "type": "text" } ], "index": 30 }, { "bbox": [ 113, 523, 506, 539 ], "spans": [ { "bbox": [ 113, 523, 200, 538 ], "score": 1.0, "content": "output the clean state", "type": "text" }, { "bbox": [ 200, 524, 216, 539 ], "score": 0.89, "content": "h _ { \\widehat { i } } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 216, 523, 506, 538 ], "score": 1.0, "content": "; future computations execute without further intervention. Intuitively, the", "type": "text" } ], "index": 31 }, { "bbox": [ 113, 534, 506, 550 ], "spans": [ { "bbox": [ 113, 534, 506, 550 ], "score": 1.0, "content": "i ability of a few clean states to recover the correct fact, despite many other states being corrupted by", "type": "text" } ], "index": 32 }, { "bbox": [ 114, 547, 460, 560 ], "spans": [ { "bbox": [ 114, 547, 460, 560 ], "score": 1.0, "content": "the obfuscated subject, will indicate their causal importance in the computation graph.", "type": "text" } ], "index": 33 } ], "index": 27, "bbox_fs": [ 104, 409, 506, 560 ] }, { "type": "text", "bbox": [ 106, 563, 505, 651 ], "lines": [ { "bbox": [ 104, 562, 507, 578 ], "spans": [ { "bbox": [ 104, 562, 123, 578 ], "score": 1.0, "content": "Let", "type": "text" }, { "bbox": [ 123, 563, 167, 576 ], "score": 0.47, "content": "\\mathbb { P } [ o ] , \\mathbb { P } _ { * } [ o ]", "type": "inline_equation" }, { "bbox": [ 167, 562, 188, 578 ], "score": 1.0, "content": ", and", "type": "text" }, { "bbox": [ 188, 564, 200, 575 ], "score": 0.82, "content": "\\mathbb { P } _ { * }", "type": "inline_equation" }, { "bbox": [ 200, 562, 219, 578 ], "score": 1.0, "content": ", clean", "type": "text" }, { "bbox": [ 220, 563, 244, 577 ], "score": 0.86, "content": "h _ { i } ^ { ( l ) } \\left[ O \\right]", "type": "inline_equation" }, { "bbox": [ 245, 562, 386, 578 ], "score": 1.0, "content": "denote the probability of emitting", "type": "text" }, { "bbox": [ 387, 566, 393, 574 ], "score": 0.64, "content": "o", "type": "inline_equation" }, { "bbox": [ 393, 562, 507, 578 ], "score": 1.0, "content": "under the clean, corrupted,", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 574, 506, 586 ], "spans": [ { "bbox": [ 105, 574, 400, 586 ], "score": 1.0, "content": "and corrupted-with-restoration runs, respectively; dependence on the input", "type": "text" }, { "bbox": [ 400, 576, 407, 584 ], "score": 0.77, "content": "x", "type": "inline_equation" }, { "bbox": [ 407, 574, 506, 586 ], "score": 1.0, "content": "is omitted for notational", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 585, 506, 598 ], "spans": [ { "bbox": [ 105, 585, 423, 598 ], "score": 1.0, "content": "simplicity. The total effect (TE) is the difference between these quantities:", "type": "text" }, { "bbox": [ 423, 585, 503, 597 ], "score": 0.92, "content": "\\mathrm { T E } = \\mathbb { P } [ o ] - \\mathbb { P } _ { * } [ o ]", "type": "inline_equation" }, { "bbox": [ 503, 585, 506, 598 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 595, 506, 609 ], "spans": [ { "bbox": [ 105, 596, 324, 609 ], "score": 1.0, "content": "The indirect effect (IE) of a specific mediating state", "type": "text" }, { "bbox": [ 324, 595, 340, 608 ], "score": 0.91, "content": "h _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 340, 596, 506, 609 ], "score": 1.0, "content": "is defined as the difference between the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 607, 506, 619 ], "spans": [ { "bbox": [ 105, 607, 165, 619 ], "score": 1.0, "content": "probability of", "type": "text" }, { "bbox": [ 165, 609, 172, 617 ], "score": 0.7, "content": "o", "type": "inline_equation" }, { "bbox": [ 172, 607, 506, 619 ], "score": 1.0, "content": "under the corrupted version and the probability when that state is set to its clean", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 618, 506, 631 ], "spans": [ { "bbox": [ 105, 618, 288, 631 ], "score": 1.0, "content": "version, while the subject remains corrupted:", "type": "text" }, { "bbox": [ 288, 618, 322, 629 ], "score": 0.88, "content": "\\mathrm { I E } = \\mathbb { P } _ { * }", "type": "inline_equation" }, { "bbox": [ 322, 618, 342, 631 ], "score": 1.0, "content": ", clean", "type": "text" }, { "bbox": [ 342, 618, 399, 631 ], "score": 0.86, "content": "h _ { i } ^ { ( l ) } \\left[ O \\right] - \\mathbb { P } _ { * } [ O ]", "type": "inline_equation" }, { "bbox": [ 400, 618, 506, 631 ], "score": 1.0, "content": ". Averaging over a sample", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 629, 505, 641 ], "spans": [ { "bbox": [ 105, 629, 505, 641 ], "score": 1.0, "content": "of statements, we obtain the average total effect (ATE) and average indirect effect (AIE) for each", "type": "text" } ], "index": 40 }, { "bbox": [ 104, 639, 198, 651 ], "spans": [ { "bbox": [ 104, 639, 198, 651 ], "score": 1.0, "content": "hidden state variable.5", "type": "text" } ], "index": 41 } ], "index": 37.5, "bbox_fs": [ 104, 562, 507, 651 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 106, 70, 501, 169 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 70, 501, 169 ], "group_id": 0, "lines": [ { "bbox": [ 106, 70, 501, 169 ], "spans": [ { "bbox": [ 106, 70, 501, 169 ], "score": 0.93, "type": "image", "image_path": "998ebba8dd8219e05022e67ab6cd31734ad78ac85b806ef3b88bf374737872c6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 70, 501, 103.0 ], "spans": [], "index": 0 }, { "bbox": [ 106, 103.0, 501, 136.0 ], "spans": [], "index": 1 }, { "bbox": [ 106, 136.0, 501, 169.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 175, 505, 216 ], "group_id": 0, "lines": [ { "bbox": [ 106, 174, 506, 188 ], "spans": [ { "bbox": [ 106, 174, 506, 188 ], "score": 1.0, "content": "Figure 3: Causal effects with a modified computation graph. (a,b) To isolate the effects of MLP modules", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 185, 505, 197 ], "spans": [ { "bbox": [ 106, 185, 505, 197 ], "score": 1.0, "content": "when measuring causal effects, the computation graph is modified. (c) Comparing Average Indirect Effects with", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 195, 505, 207 ], "spans": [ { "bbox": [ 106, 195, 505, 207 ], "score": 1.0, "content": "and without severing MLP implicates the computation of (e) midlayer MLP modules in the causal effects. No", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 205, 304, 216 ], "spans": [ { "bbox": [ 106, 205, 304, 216 ], "score": 1.0, "content": "similar gap is seen when attention is similarly severed.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "title", "bbox": [ 107, 234, 229, 245 ], "lines": [ { "bbox": [ 105, 232, 230, 248 ], "spans": [ { "bbox": [ 105, 232, 230, 248 ], "score": 1.0, "content": "2.2 Causal Tracing Results", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 107, 254, 505, 331 ], "lines": [ { "bbox": [ 106, 255, 506, 267 ], "spans": [ { "bbox": [ 106, 255, 506, 267 ], "score": 1.0, "content": "We compute the average indirect effect (AIE) over 1000 factual statements (details in Appendix B.1),", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 265, 506, 279 ], "spans": [ { "bbox": [ 105, 265, 506, 279 ], "score": 1.0, "content": "varying the mediator over different positions in the sentence and different model components including", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 277, 505, 289 ], "spans": [ { "bbox": [ 105, 277, 505, 289 ], "score": 1.0, "content": "individual states, MLP layers, and attention layers. Figure 2 plots the AIE of the internal components", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 286, 505, 301 ], "spans": [ { "bbox": [ 105, 286, 372, 301 ], "score": 1.0, "content": "of GPT-2 XL (1.5B parameters). The ATE of this experiment is", "type": "text" }, { "bbox": [ 372, 288, 399, 298 ], "score": 0.86, "content": "1 8 . 6 \\%", "type": "inline_equation" }, { "bbox": [ 400, 286, 505, 301 ], "score": 1.0, "content": ", and we note that a large", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 298, 505, 311 ], "spans": [ { "bbox": [ 105, 298, 384, 311 ], "score": 1.0, "content": "portion of the effect is mediated by strongly causal individual states", "type": "text" }, { "bbox": [ 384, 299, 431, 309 ], "score": 0.86, "content": "( \\mathrm { A I E { = } } 8 . 7 \\%", "type": "inline_equation" }, { "bbox": [ 431, 298, 505, 311 ], "score": 1.0, "content": "at layer 15) at the", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 309, 505, 322 ], "spans": [ { "bbox": [ 105, 309, 505, 322 ], "score": 1.0, "content": "last subject token. The presence of strong causal states at a late site immediately before the prediction", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 319, 507, 334 ], "spans": [ { "bbox": [ 105, 319, 507, 334 ], "score": 1.0, "content": "is unsurprising, but their emergence at an early site at the last token of the subject is a new discovery.", "type": "text" } ], "index": 14 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 336, 505, 381 ], "lines": [ { "bbox": [ 105, 336, 506, 349 ], "spans": [ { "bbox": [ 105, 336, 506, 349 ], "score": 1.0, "content": "Decomposing the causal effects of contributions of MLP and attention modules (Figure 1fg and", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 347, 506, 360 ], "spans": [ { "bbox": [ 105, 347, 506, 360 ], "score": 1.0, "content": "Figure 2bc) suggests a decisive role for MLP modules at the early site: MLP contributions peak at", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 358, 506, 371 ], "spans": [ { "bbox": [ 105, 358, 125, 371 ], "score": 1.0, "content": "AIE", "type": "text" }, { "bbox": [ 125, 358, 147, 369 ], "score": 0.86, "content": "6 . 6 \\%", "type": "inline_equation" }, { "bbox": [ 147, 358, 358, 371 ], "score": 1.0, "content": ", while attention at the last subject token is only AIE", "type": "text" }, { "bbox": [ 358, 359, 380, 369 ], "score": 0.87, "content": "1 . 6 \\%", "type": "inline_equation" }, { "bbox": [ 380, 358, 506, 371 ], "score": 1.0, "content": "; attention is more important at", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 370, 430, 381 ], "spans": [ { "bbox": [ 105, 370, 430, 381 ], "score": 1.0, "content": "the last token of the prompt. Appendix B.2 further discusses this decomposition.", "type": "text" } ], "index": 18 } ], "index": 16.5 }, { "type": "text", "bbox": [ 107, 385, 505, 506 ], "lines": [ { "bbox": [ 105, 385, 505, 398 ], "spans": [ { "bbox": [ 105, 385, 505, 398 ], "score": 1.0, "content": "Finally, to gain a clearer picture of the special role of MLP layers at the early site, we analyze indirect", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 397, 505, 408 ], "spans": [ { "bbox": [ 106, 397, 505, 408 ], "score": 1.0, "content": "effects with a modified causal graph (Figure 3). (a) First, we collect each MLP module contribution", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 407, 505, 420 ], "spans": [ { "bbox": [ 105, 407, 505, 420 ], "score": 1.0, "content": "in the baseline condition with corrupted input. (b) Then, to isolate the effects of MLP modules when", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 419, 504, 431 ], "spans": [ { "bbox": [ 105, 419, 499, 431 ], "score": 1.0, "content": "measuring causal effects, we modify the computation graph to sever MLP computations at token", "type": "text" }, { "bbox": [ 499, 419, 504, 428 ], "score": 0.45, "content": "i", "type": "inline_equation" } ], "index": 22 }, { "bbox": [ 105, 429, 505, 442 ], "spans": [ { "bbox": [ 105, 429, 505, 442 ], "score": 1.0, "content": "and freeze them in the baseline corrupted state so that they are unaffected by the insertion of clean", "type": "text" } ], "index": 23 }, { "bbox": [ 103, 437, 507, 455 ], "spans": [ { "bbox": [ 103, 437, 141, 455 ], "score": 1.0, "content": "state for", "type": "text" }, { "bbox": [ 141, 438, 157, 452 ], "score": 0.91, "content": "h _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 158, 437, 507, 455 ], "score": 1.0, "content": ". This modification is a way of probing path-specific effects (Pearl, 2001) for paths that", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 451, 505, 464 ], "spans": [ { "bbox": [ 105, 451, 505, 464 ], "score": 1.0, "content": "avoid MLP computations. (c) Comparing Average Indirect Effects in the modified graph to the those", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 462, 505, 474 ], "spans": [ { "bbox": [ 105, 462, 505, 474 ], "score": 1.0, "content": "in the original graph, we observe (d) the lowest layers lose their causal effect without the activity", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 473, 505, 486 ], "spans": [ { "bbox": [ 105, 473, 505, 486 ], "score": 1.0, "content": "of future MLP modules, while (f) higher layer states’ effects depend little on the MLP activity. No", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 484, 505, 497 ], "spans": [ { "bbox": [ 105, 484, 505, 497 ], "score": 1.0, "content": "such transition is seen when the comparison is carried out severing the attention modules. This result", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 495, 500, 507 ], "spans": [ { "bbox": [ 105, 495, 500, 507 ], "score": 1.0, "content": "confirms an essential role for (e) MLP module computation at middle layers when recalling a fact.", "type": "text" } ], "index": 29 } ], "index": 24 }, { "type": "text", "bbox": [ 108, 511, 505, 544 ], "lines": [ { "bbox": [ 106, 511, 505, 524 ], "spans": [ { "bbox": [ 106, 511, 505, 524 ], "score": 1.0, "content": "Appendix B has results on other autoregressive models and experimental settings. In particular, we", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 522, 505, 534 ], "spans": [ { "bbox": [ 105, 522, 505, 534 ], "score": 1.0, "content": "find that Causal Tracing is more informative than gradient-based salience methods such as integrated", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 533, 502, 546 ], "spans": [ { "bbox": [ 105, 533, 502, 546 ], "score": 1.0, "content": "gradients (Sundararajan et al., 2017) (Figure 16) and is robust under different noise configurations.", "type": "text" } ], "index": 32 } ], "index": 31 }, { "type": "text", "bbox": [ 108, 549, 504, 561 ], "lines": [ { "bbox": [ 106, 548, 505, 562 ], "spans": [ { "bbox": [ 106, 548, 505, 562 ], "score": 1.0, "content": "We hypothesize that this localized midlayer MLP key–value mapping recalls facts about the subject.", "type": "text" } ], "index": 33 } ], "index": 33 }, { "type": "title", "bbox": [ 107, 575, 326, 587 ], "lines": [ { "bbox": [ 105, 573, 327, 589 ], "spans": [ { "bbox": [ 105, 573, 327, 589 ], "score": 1.0, "content": "2.3 The Localized Factual Association Hypothesis", "type": "text" } ], "index": 34 } ], "index": 34 }, { "type": "text", "bbox": [ 107, 596, 505, 640 ], "lines": [ { "bbox": [ 105, 596, 505, 609 ], "spans": [ { "bbox": [ 105, 596, 505, 609 ], "score": 1.0, "content": "Based on causal traces, we posit a specific mechanism for storage of factual associations: each", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 607, 505, 619 ], "spans": [ { "bbox": [ 106, 607, 505, 619 ], "score": 1.0, "content": "midlayer MLP module accepts inputs that encode a subject, then produces outputs that recall", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 618, 505, 630 ], "spans": [ { "bbox": [ 106, 618, 505, 630 ], "score": 1.0, "content": "memorized properties about that subject. Middle layer MLP outputs accumulate information, then", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 628, 419, 642 ], "spans": [ { "bbox": [ 105, 628, 419, 642 ], "score": 1.0, "content": "the summed information is copied to the last token by attention at high layers.", "type": "text" } ], "index": 38 } ], "index": 36.5 }, { "type": "text", "bbox": [ 107, 645, 505, 722 ], "lines": [ { "bbox": [ 105, 644, 505, 658 ], "spans": [ { "bbox": [ 105, 644, 505, 658 ], "score": 1.0, "content": "This hypothesis localizes factual association along three dimensions, placing it (i) in the MLP modules", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 656, 505, 668 ], "spans": [ { "bbox": [ 105, 656, 505, 668 ], "score": 1.0, "content": "(ii) at specific middle layers (iii) and specifically at the processing of the subject’s last token. It is", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 666, 506, 680 ], "spans": [ { "bbox": [ 105, 666, 506, 680 ], "score": 1.0, "content": "consistent with the Geva et al. (2021) view that MLP layers store knowledge, and the Elhage et al.", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 678, 505, 691 ], "spans": [ { "bbox": [ 105, 678, 505, 691 ], "score": 1.0, "content": "(2021) study showing an information-copying role for self-attention. Furthermore, informed by the", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 688, 505, 701 ], "spans": [ { "bbox": [ 105, 688, 505, 701 ], "score": 1.0, "content": "Zhao et al. (2021) finding that transformer layer order can be exchanged with minimal change in", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 700, 505, 712 ], "spans": [ { "bbox": [ 105, 700, 505, 712 ], "score": 1.0, "content": "behavior, we propose that this picture is complete. That is, there is no further special role for the", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 711, 505, 723 ], "spans": [ { "bbox": [ 105, 711, 505, 723 ], "score": 1.0, "content": "particular choice or arrangement of individual layers in the middle range. We conjecture that any fact", "type": "text" } ], "index": 45 } ], "index": 42 } ], "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": [ 106, 70, 501, 169 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 70, 501, 169 ], "group_id": 0, "lines": [ { "bbox": [ 106, 70, 501, 169 ], "spans": [ { "bbox": [ 106, 70, 501, 169 ], "score": 0.93, "type": "image", "image_path": "998ebba8dd8219e05022e67ab6cd31734ad78ac85b806ef3b88bf374737872c6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 70, 501, 103.0 ], "spans": [], "index": 0 }, { "bbox": [ 106, 103.0, 501, 136.0 ], "spans": [], "index": 1 }, { "bbox": [ 106, 136.0, 501, 169.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 175, 505, 216 ], "group_id": 0, "lines": [ { "bbox": [ 106, 174, 506, 188 ], "spans": [ { "bbox": [ 106, 174, 506, 188 ], "score": 1.0, "content": "Figure 3: Causal effects with a modified computation graph. (a,b) To isolate the effects of MLP modules", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 185, 505, 197 ], "spans": [ { "bbox": [ 106, 185, 505, 197 ], "score": 1.0, "content": "when measuring causal effects, the computation graph is modified. (c) Comparing Average Indirect Effects with", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 195, 505, 207 ], "spans": [ { "bbox": [ 106, 195, 505, 207 ], "score": 1.0, "content": "and without severing MLP implicates the computation of (e) midlayer MLP modules in the causal effects. No", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 205, 304, 216 ], "spans": [ { "bbox": [ 106, 205, 304, 216 ], "score": 1.0, "content": "similar gap is seen when attention is similarly severed.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "title", "bbox": [ 107, 234, 229, 245 ], "lines": [ { "bbox": [ 105, 232, 230, 248 ], "spans": [ { "bbox": [ 105, 232, 230, 248 ], "score": 1.0, "content": "2.2 Causal Tracing Results", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 107, 254, 505, 331 ], "lines": [ { "bbox": [ 106, 255, 506, 267 ], "spans": [ { "bbox": [ 106, 255, 506, 267 ], "score": 1.0, "content": "We compute the average indirect effect (AIE) over 1000 factual statements (details in Appendix B.1),", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 265, 506, 279 ], "spans": [ { "bbox": [ 105, 265, 506, 279 ], "score": 1.0, "content": "varying the mediator over different positions in the sentence and different model components including", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 277, 505, 289 ], "spans": [ { "bbox": [ 105, 277, 505, 289 ], "score": 1.0, "content": "individual states, MLP layers, and attention layers. Figure 2 plots the AIE of the internal components", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 286, 505, 301 ], "spans": [ { "bbox": [ 105, 286, 372, 301 ], "score": 1.0, "content": "of GPT-2 XL (1.5B parameters). The ATE of this experiment is", "type": "text" }, { "bbox": [ 372, 288, 399, 298 ], "score": 0.86, "content": "1 8 . 6 \\%", "type": "inline_equation" }, { "bbox": [ 400, 286, 505, 301 ], "score": 1.0, "content": ", and we note that a large", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 298, 505, 311 ], "spans": [ { "bbox": [ 105, 298, 384, 311 ], "score": 1.0, "content": "portion of the effect is mediated by strongly causal individual states", "type": "text" }, { "bbox": [ 384, 299, 431, 309 ], "score": 0.86, "content": "( \\mathrm { A I E { = } } 8 . 7 \\%", "type": "inline_equation" }, { "bbox": [ 431, 298, 505, 311 ], "score": 1.0, "content": "at layer 15) at the", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 309, 505, 322 ], "spans": [ { "bbox": [ 105, 309, 505, 322 ], "score": 1.0, "content": "last subject token. The presence of strong causal states at a late site immediately before the prediction", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 319, 507, 334 ], "spans": [ { "bbox": [ 105, 319, 507, 334 ], "score": 1.0, "content": "is unsurprising, but their emergence at an early site at the last token of the subject is a new discovery.", "type": "text" } ], "index": 14 } ], "index": 11, "bbox_fs": [ 105, 255, 507, 334 ] }, { "type": "text", "bbox": [ 107, 336, 505, 381 ], "lines": [ { "bbox": [ 105, 336, 506, 349 ], "spans": [ { "bbox": [ 105, 336, 506, 349 ], "score": 1.0, "content": "Decomposing the causal effects of contributions of MLP and attention modules (Figure 1fg and", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 347, 506, 360 ], "spans": [ { "bbox": [ 105, 347, 506, 360 ], "score": 1.0, "content": "Figure 2bc) suggests a decisive role for MLP modules at the early site: MLP contributions peak at", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 358, 506, 371 ], "spans": [ { "bbox": [ 105, 358, 125, 371 ], "score": 1.0, "content": "AIE", "type": "text" }, { "bbox": [ 125, 358, 147, 369 ], "score": 0.86, "content": "6 . 6 \\%", "type": "inline_equation" }, { "bbox": [ 147, 358, 358, 371 ], "score": 1.0, "content": ", while attention at the last subject token is only AIE", "type": "text" }, { "bbox": [ 358, 359, 380, 369 ], "score": 0.87, "content": "1 . 6 \\%", "type": "inline_equation" }, { "bbox": [ 380, 358, 506, 371 ], "score": 1.0, "content": "; attention is more important at", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 370, 430, 381 ], "spans": [ { "bbox": [ 105, 370, 430, 381 ], "score": 1.0, "content": "the last token of the prompt. Appendix B.2 further discusses this decomposition.", "type": "text" } ], "index": 18 } ], "index": 16.5, "bbox_fs": [ 105, 336, 506, 381 ] }, { "type": "text", "bbox": [ 107, 385, 505, 506 ], "lines": [ { "bbox": [ 105, 385, 505, 398 ], "spans": [ { "bbox": [ 105, 385, 505, 398 ], "score": 1.0, "content": "Finally, to gain a clearer picture of the special role of MLP layers at the early site, we analyze indirect", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 397, 505, 408 ], "spans": [ { "bbox": [ 106, 397, 505, 408 ], "score": 1.0, "content": "effects with a modified causal graph (Figure 3). (a) First, we collect each MLP module contribution", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 407, 505, 420 ], "spans": [ { "bbox": [ 105, 407, 505, 420 ], "score": 1.0, "content": "in the baseline condition with corrupted input. (b) Then, to isolate the effects of MLP modules when", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 419, 504, 431 ], "spans": [ { "bbox": [ 105, 419, 499, 431 ], "score": 1.0, "content": "measuring causal effects, we modify the computation graph to sever MLP computations at token", "type": "text" }, { "bbox": [ 499, 419, 504, 428 ], "score": 0.45, "content": "i", "type": "inline_equation" } ], "index": 22 }, { "bbox": [ 105, 429, 505, 442 ], "spans": [ { "bbox": [ 105, 429, 505, 442 ], "score": 1.0, "content": "and freeze them in the baseline corrupted state so that they are unaffected by the insertion of clean", "type": "text" } ], "index": 23 }, { "bbox": [ 103, 437, 507, 455 ], "spans": [ { "bbox": [ 103, 437, 141, 455 ], "score": 1.0, "content": "state for", "type": "text" }, { "bbox": [ 141, 438, 157, 452 ], "score": 0.91, "content": "h _ { i } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 158, 437, 507, 455 ], "score": 1.0, "content": ". This modification is a way of probing path-specific effects (Pearl, 2001) for paths that", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 451, 505, 464 ], "spans": [ { "bbox": [ 105, 451, 505, 464 ], "score": 1.0, "content": "avoid MLP computations. (c) Comparing Average Indirect Effects in the modified graph to the those", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 462, 505, 474 ], "spans": [ { "bbox": [ 105, 462, 505, 474 ], "score": 1.0, "content": "in the original graph, we observe (d) the lowest layers lose their causal effect without the activity", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 473, 505, 486 ], "spans": [ { "bbox": [ 105, 473, 505, 486 ], "score": 1.0, "content": "of future MLP modules, while (f) higher layer states’ effects depend little on the MLP activity. No", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 484, 505, 497 ], "spans": [ { "bbox": [ 105, 484, 505, 497 ], "score": 1.0, "content": "such transition is seen when the comparison is carried out severing the attention modules. This result", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 495, 500, 507 ], "spans": [ { "bbox": [ 105, 495, 500, 507 ], "score": 1.0, "content": "confirms an essential role for (e) MLP module computation at middle layers when recalling a fact.", "type": "text" } ], "index": 29 } ], "index": 24, "bbox_fs": [ 103, 385, 507, 507 ] }, { "type": "text", "bbox": [ 108, 511, 505, 544 ], "lines": [ { "bbox": [ 106, 511, 505, 524 ], "spans": [ { "bbox": [ 106, 511, 505, 524 ], "score": 1.0, "content": "Appendix B has results on other autoregressive models and experimental settings. In particular, we", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 522, 505, 534 ], "spans": [ { "bbox": [ 105, 522, 505, 534 ], "score": 1.0, "content": "find that Causal Tracing is more informative than gradient-based salience methods such as integrated", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 533, 502, 546 ], "spans": [ { "bbox": [ 105, 533, 502, 546 ], "score": 1.0, "content": "gradients (Sundararajan et al., 2017) (Figure 16) and is robust under different noise configurations.", "type": "text" } ], "index": 32 } ], "index": 31, "bbox_fs": [ 105, 511, 505, 546 ] }, { "type": "text", "bbox": [ 108, 549, 504, 561 ], "lines": [ { "bbox": [ 106, 548, 505, 562 ], "spans": [ { "bbox": [ 106, 548, 505, 562 ], "score": 1.0, "content": "We hypothesize that this localized midlayer MLP key–value mapping recalls facts about the subject.", "type": "text" } ], "index": 33 } ], "index": 33, "bbox_fs": [ 106, 548, 505, 562 ] }, { "type": "title", "bbox": [ 107, 575, 326, 587 ], "lines": [ { "bbox": [ 105, 573, 327, 589 ], "spans": [ { "bbox": [ 105, 573, 327, 589 ], "score": 1.0, "content": "2.3 The Localized Factual Association Hypothesis", "type": "text" } ], "index": 34 } ], "index": 34 }, { "type": "text", "bbox": [ 107, 596, 505, 640 ], "lines": [ { "bbox": [ 105, 596, 505, 609 ], "spans": [ { "bbox": [ 105, 596, 505, 609 ], "score": 1.0, "content": "Based on causal traces, we posit a specific mechanism for storage of factual associations: each", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 607, 505, 619 ], "spans": [ { "bbox": [ 106, 607, 505, 619 ], "score": 1.0, "content": "midlayer MLP module accepts inputs that encode a subject, then produces outputs that recall", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 618, 505, 630 ], "spans": [ { "bbox": [ 106, 618, 505, 630 ], "score": 1.0, "content": "memorized properties about that subject. Middle layer MLP outputs accumulate information, then", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 628, 419, 642 ], "spans": [ { "bbox": [ 105, 628, 419, 642 ], "score": 1.0, "content": "the summed information is copied to the last token by attention at high layers.", "type": "text" } ], "index": 38 } ], "index": 36.5, "bbox_fs": [ 105, 596, 505, 642 ] }, { "type": "text", "bbox": [ 107, 645, 505, 722 ], "lines": [ { "bbox": [ 105, 644, 505, 658 ], "spans": [ { "bbox": [ 105, 644, 505, 658 ], "score": 1.0, "content": "This hypothesis localizes factual association along three dimensions, placing it (i) in the MLP modules", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 656, 505, 668 ], "spans": [ { "bbox": [ 105, 656, 505, 668 ], "score": 1.0, "content": "(ii) at specific middle layers (iii) and specifically at the processing of the subject’s last token. It is", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 666, 506, 680 ], "spans": [ { "bbox": [ 105, 666, 506, 680 ], "score": 1.0, "content": "consistent with the Geva et al. (2021) view that MLP layers store knowledge, and the Elhage et al.", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 678, 505, 691 ], "spans": [ { "bbox": [ 105, 678, 505, 691 ], "score": 1.0, "content": "(2021) study showing an information-copying role for self-attention. Furthermore, informed by the", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 688, 505, 701 ], "spans": [ { "bbox": [ 105, 688, 505, 701 ], "score": 1.0, "content": "Zhao et al. (2021) finding that transformer layer order can be exchanged with minimal change in", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 700, 505, 712 ], "spans": [ { "bbox": [ 105, 700, 505, 712 ], "score": 1.0, "content": "behavior, we propose that this picture is complete. That is, there is no further special role for the", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 711, 505, 723 ], "spans": [ { "bbox": [ 105, 711, 505, 723 ], "score": 1.0, "content": "particular choice or arrangement of individual layers in the middle range. We conjecture that any fact", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 246, 505, 260 ], "spans": [ { "bbox": [ 105, 246, 505, 260 ], "score": 1.0, "content": "could be equivalently stored in any one of the middle MLP layers. To test our hypothesis, we narrow", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 105, 258, 505, 270 ], "spans": [ { "bbox": [ 105, 258, 347, 270 ], "score": 1.0, "content": "our attention to a single MLP module at a mid-range layer", "type": "text", "cross_page": true }, { "bbox": [ 347, 259, 356, 268 ], "score": 0.83, "content": "l ^ { * }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 356, 258, 505, 270 ], "score": 1.0, "content": ", and ask whether its weights can be", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 105, 269, 285, 282 ], "spans": [ { "bbox": [ 105, 269, 285, 282 ], "score": 1.0, "content": "explicitly modified to store an arbitrary fact.", "type": "text", "cross_page": true } ], "index": 10 } ], "index": 42, "bbox_fs": [ 105, 644, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 105, 70, 506, 167 ], "blocks": [ { "type": "image_body", "bbox": [ 105, 70, 506, 167 ], "group_id": 0, "lines": [ { "bbox": [ 105, 70, 506, 167 ], "spans": [ { "bbox": [ 105, 70, 506, 167 ], "score": 0.966, "type": "image", "image_path": "eeb4c5f59c324d7747ed46f32e3dadf49334273e55fa708df9ff0730a7f294f8.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 105, 70, 506, 102.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 105, 102.33333333333334, 506, 134.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 105, 134.66666666666669, 506, 167.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 173, 505, 224 ], "group_id": 0, "lines": [ { "bbox": [ 106, 173, 505, 184 ], "spans": [ { "bbox": [ 106, 173, 505, 184 ], "score": 1.0, "content": "Figure 4: Editing one MLP layer with ROME. To associate Space Needle with Paris, the ROME method", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 182, 504, 195 ], "spans": [ { "bbox": [ 105, 182, 156, 195 ], "score": 1.0, "content": "inserts a new", "type": "text" }, { "bbox": [ 156, 183, 186, 194 ], "score": 0.92, "content": "( k _ { * } , v _ { * } )", "type": "inline_equation" }, { "bbox": [ 186, 182, 266, 195 ], "score": 1.0, "content": "association into layer", "type": "text" }, { "bbox": [ 266, 183, 274, 192 ], "score": 0.77, "content": "l ^ { * }", "type": "inline_equation" }, { "bbox": [ 274, 182, 329, 195 ], "score": 1.0, "content": ", where (a) key", "type": "text" }, { "bbox": [ 329, 183, 339, 192 ], "score": 0.87, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 340, 182, 493, 195 ], "score": 1.0, "content": "is determined by the subject and (b) value", "type": "text" }, { "bbox": [ 494, 185, 504, 192 ], "score": 0.82, "content": "v _ { * }", "type": "inline_equation" } ], "index": 4 }, { "bbox": [ 104, 192, 505, 205 ], "spans": [ { "bbox": [ 104, 192, 321, 205 ], "score": 1.0, "content": "is optimized to select the object. (c) Hidden state at layer", "type": "text" }, { "bbox": [ 321, 193, 329, 202 ], "score": 0.84, "content": "l ^ { * }", "type": "inline_equation" }, { "bbox": [ 330, 192, 369, 205 ], "score": 1.0, "content": "and token", "type": "text" }, { "bbox": [ 369, 194, 374, 202 ], "score": 0.64, "content": "_ { i }", "type": "inline_equation" }, { "bbox": [ 374, 192, 505, 205 ], "score": 1.0, "content": "is expanded to produce (d) the key", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 202, 506, 215 ], "spans": [ { "bbox": [ 105, 202, 131, 215 ], "score": 1.0, "content": "vector", "type": "text" }, { "bbox": [ 132, 204, 141, 213 ], "score": 0.86, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 142, 202, 312, 215 ], "score": 1.0, "content": "for the subject. (e) To write new value vector", "type": "text" }, { "bbox": [ 312, 204, 322, 213 ], "score": 0.86, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 322, 202, 506, 215 ], "score": 1.0, "content": "into the layer, (f) we calculate a rank-one update", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 209, 500, 228 ], "spans": [ { "bbox": [ 106, 213, 152, 224 ], "score": 0.92, "content": "\\Lambda ( C ^ { - 1 } k _ { * } ) ^ { T }", "type": "inline_equation" }, { "bbox": [ 153, 209, 185, 228 ], "score": 1.0, "content": "to cause", "type": "text" }, { "bbox": [ 186, 212, 240, 225 ], "score": 0.92, "content": "\\hat { W } _ { p r o j } ^ { ( l ) } \\hat { k } _ { * } = v _ { * }", "type": "inline_equation" }, { "bbox": [ 240, 209, 500, 228 ], "score": 1.0, "content": "⇤ while minimizing interference with other memories stored in the layer.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 107, 246, 505, 280 ], "lines": [ { "bbox": [ 105, 246, 505, 260 ], "spans": [ { "bbox": [ 105, 246, 505, 260 ], "score": 1.0, "content": "could be equivalently stored in any one of the middle MLP layers. To test our hypothesis, we narrow", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 258, 505, 270 ], "spans": [ { "bbox": [ 105, 258, 347, 270 ], "score": 1.0, "content": "our attention to a single MLP module at a mid-range layer", "type": "text" }, { "bbox": [ 347, 259, 356, 268 ], "score": 0.83, "content": "l ^ { * }", "type": "inline_equation" }, { "bbox": [ 356, 258, 505, 270 ], "score": 1.0, "content": ", and ask whether its weights can be", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 269, 285, 282 ], "spans": [ { "bbox": [ 105, 269, 285, 282 ], "score": 1.0, "content": "explicitly modified to store an arbitrary fact.", "type": "text" } ], "index": 10 } ], "index": 9 }, { "type": "title", "bbox": [ 105, 303, 496, 318 ], "lines": [ { "bbox": [ 103, 302, 498, 320 ], "spans": [ { "bbox": [ 103, 302, 498, 320 ], "score": 1.0, "content": "3 Interventions on Weights for Understanding Factual Association Storage", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 335, 505, 390 ], "lines": [ { "bbox": [ 106, 335, 506, 347 ], "spans": [ { "bbox": [ 106, 335, 506, 347 ], "score": 1.0, "content": "While Causal Tracing has implicated MLP modules in recalling factual associations, we also wish to", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 346, 506, 358 ], "spans": [ { "bbox": [ 106, 346, 506, 358 ], "score": 1.0, "content": "understand how facts are stored in weights. Geva et al. (2021) observed that MLP layers (Figure 4cde)", "type": "text" } ], "index": 13 }, { "bbox": [ 103, 353, 509, 375 ], "spans": [ { "bbox": [ 103, 353, 434, 375 ], "score": 1.0, "content": "can act as two-layer key–value memories,6 where the neurons of the first layer (l)", "type": "text" }, { "bbox": [ 434, 356, 455, 371 ], "score": 0.92, "content": "\\mathbf { \\overline { { \\it W } } } _ { f c } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 456, 353, 509, 375 ], "score": 1.0, "content": "form a key,", "type": "text" } ], "index": 14 }, { "bbox": [ 104, 366, 507, 385 ], "spans": [ { "bbox": [ 104, 366, 220, 385 ], "score": 1.0, "content": "with which the second layer", "type": "text" }, { "bbox": [ 221, 367, 247, 382 ], "score": 0.93, "content": "W _ { p r o j } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 247, 366, 507, 385 ], "score": 1.0, "content": "retrieves an associated value. We hypothesize that MLPs can be", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 378, 496, 392 ], "spans": [ { "bbox": [ 105, 378, 496, 392 ], "score": 1.0, "content": "modeled as a linear associative memory; note that this differs from Geva et al.’s per-neuron view.", "type": "text" } ], "index": 16 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 395, 505, 440 ], "lines": [ { "bbox": [ 105, 395, 505, 408 ], "spans": [ { "bbox": [ 105, 395, 505, 408 ], "score": 1.0, "content": "We test this hypothesis by conducting a new type of intervention: modifying factual associations with", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 406, 505, 419 ], "spans": [ { "bbox": [ 105, 406, 434, 419 ], "score": 1.0, "content": "Rank-One Model Editing (ROME). Being able to insert a new knowledge tuple", "type": "text" }, { "bbox": [ 434, 406, 493, 418 ], "score": 0.93, "content": "t ^ { * } = ( s , r , o ^ { * } )", "type": "inline_equation" }, { "bbox": [ 493, 406, 505, 419 ], "score": 1.0, "content": "in", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 417, 506, 430 ], "spans": [ { "bbox": [ 105, 417, 209, 430 ], "score": 1.0, "content": "place of the current tuple", "type": "text" }, { "bbox": [ 209, 417, 265, 429 ], "score": 0.93, "content": "t ^ { c } = \\left( s , r , o ^ { c } \\right)", "type": "inline_equation" }, { "bbox": [ 265, 417, 506, 430 ], "score": 1.0, "content": "with both generalization and specificity would demonstrate", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 429, 374, 441 ], "spans": [ { "bbox": [ 106, 429, 374, 441 ], "score": 1.0, "content": "fine-grained understanding of the association-storage mechanisms.", "type": "text" } ], "index": 20 } ], "index": 18.5 }, { "type": "title", "bbox": [ 106, 454, 489, 466 ], "lines": [ { "bbox": [ 104, 451, 491, 470 ], "spans": [ { "bbox": [ 104, 451, 491, 470 ], "score": 1.0, "content": "3.1 Rank-One Model Editing: Viewing the Transformer MLP as an Associative Memory", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 106, 473, 506, 552 ], "lines": [ { "bbox": [ 103, 471, 506, 489 ], "spans": [ { "bbox": [ 103, 471, 168, 489 ], "score": 1.0, "content": "We view W (l)proj", "type": "text" }, { "bbox": [ 170, 475, 506, 489 ], "score": 1.0, "content": "as a linear associative memory (Kohonen, 1972; Anderson, 1972). This perspective", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 485, 506, 498 ], "spans": [ { "bbox": [ 105, 485, 251, 498 ], "score": 1.0, "content": "observes that any linear operation", "type": "text" }, { "bbox": [ 251, 486, 263, 496 ], "score": 0.65, "content": "W", "type": "inline_equation" }, { "bbox": [ 263, 485, 506, 498 ], "score": 1.0, "content": "can operate as a key–value store for a set of vector keys", "type": "text" } ], "index": 23 }, { "bbox": [ 107, 496, 506, 510 ], "spans": [ { "bbox": [ 107, 497, 190, 509 ], "score": 0.92, "content": "K = [ k _ { 1 } \\ | \\ k _ { 2 } \\ | \\ . \\ . \\ . ]", "type": "inline_equation" }, { "bbox": [ 190, 496, 326, 510 ], "score": 1.0, "content": "and corresponding vector values", "type": "text" }, { "bbox": [ 327, 497, 408, 509 ], "score": 0.91, "content": "V = \\left[ v _ { 1 } \\mid v _ { 2 } \\mid \\ldots \\right]", "type": "inline_equation" }, { "bbox": [ 408, 496, 459, 510 ], "score": 1.0, "content": ", by solving", "type": "text" }, { "bbox": [ 459, 497, 502, 507 ], "score": 0.89, "content": "W K \\approx V", "type": "inline_equation" }, { "bbox": [ 503, 496, 506, 510 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 508, 506, 520 ], "spans": [ { "bbox": [ 105, 508, 412, 520 ], "score": 1.0, "content": "whose squared error is minimized using the Moore-Penrose pseudoinverse:", "type": "text" }, { "bbox": [ 413, 508, 461, 518 ], "score": 0.91, "content": "\\dot { W } = V { \\bar { K ^ { + } } }", "type": "inline_equation" }, { "bbox": [ 461, 508, 506, 520 ], "score": 1.0, "content": ". Bau et al.", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 518, 506, 532 ], "spans": [ { "bbox": [ 105, 518, 283, 532 ], "score": 1.0, "content": "(2020) observed that a new key–value pair", "type": "text" }, { "bbox": [ 283, 519, 315, 531 ], "score": 0.93, "content": "( k _ { * } , v _ { * } )", "type": "inline_equation" }, { "bbox": [ 315, 518, 506, 532 ], "score": 1.0, "content": "can be inserted optimally into the memory by", "type": "text" } ], "index": 26 }, { "bbox": [ 104, 528, 506, 544 ], "spans": [ { "bbox": [ 104, 528, 506, 544 ], "score": 1.0, "content": "solving a constrained least-squares problem. In a convolutional network, Bau et al. solve this using", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 541, 447, 552 ], "spans": [ { "bbox": [ 105, 541, 447, 552 ], "score": 1.0, "content": "an optimization, but in a fully-connected layer, we can derive a closed form solution:", "type": "text" } ], "index": 28 } ], "index": 25 }, { "type": "interline_equation", "bbox": [ 175, 559, 469, 575 ], "lines": [ { "bbox": [ 175, 559, 469, 575 ], "spans": [ { "bbox": [ 175, 559, 469, 575 ], "score": 0.66, "content": "\\mathrm { ~ e ~ } \\Vert \\hat { W } K - V \\Vert \\mathrm { ~ s u c h ~ t h a t ~ } \\hat { W } k _ { * } = v _ { * } \\quad \\mathrm { b y ~ s e t t i n g ~ } \\hat { W } = W + \\Lambda ( C ^ { - 1 } k _ { * } ) ^ { T } .", "type": "interline_equation", "image_path": "812a0ecccfad6534321ab6f067c0a9e82bc6dd7692683887a95ab0e4a7320554.jpg" } ] } ], "index": 29, "virtual_lines": [ { "bbox": [ 175, 559, 469, 575 ], "spans": [], "index": 29 } ] }, { "type": "text", "bbox": [ 106, 583, 504, 639 ], "lines": [ { "bbox": [ 105, 583, 506, 596 ], "spans": [ { "bbox": [ 105, 583, 127, 596 ], "score": 1.0, "content": "Here", "type": "text" }, { "bbox": [ 128, 584, 140, 594 ], "score": 0.57, "content": "W", "type": "inline_equation" }, { "bbox": [ 140, 583, 225, 596 ], "score": 1.0, "content": "is the original matrix,", "type": "text" }, { "bbox": [ 225, 583, 271, 594 ], "score": 0.92, "content": "C = K K ^ { T }", "type": "inline_equation" }, { "bbox": [ 272, 583, 506, 596 ], "score": 1.0, "content": "is a constant that we pre-cache by estimating the uncentered", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 593, 503, 608 ], "spans": [ { "bbox": [ 105, 593, 161, 608 ], "score": 1.0, "content": "covariance of", "type": "text" }, { "bbox": [ 161, 595, 168, 605 ], "score": 0.83, "content": "k", "type": "inline_equation" }, { "bbox": [ 168, 593, 378, 608 ], "score": 1.0, "content": "from a sample of Wikipedia text (Appendix E.5), and", "type": "text" }, { "bbox": [ 378, 594, 503, 607 ], "score": 0.92, "content": "\\Lambda = \\mathbf { \\bar { \\Phi } } ( v _ { * } - W k _ { * } ) / ( C ^ { - 1 } k _ { * } ) ^ { T } k _ { * }", "type": "inline_equation" } ], "index": 31 }, { "bbox": [ 106, 606, 505, 618 ], "spans": [ { "bbox": [ 106, 606, 505, 618 ], "score": 1.0, "content": "is a vector proportional to the residual error of the new key–value pair on the original memory matrix", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 616, 506, 630 ], "spans": [ { "bbox": [ 105, 616, 506, 630 ], "score": 1.0, "content": "(full derivation in Appendix A). Because of this simple algebraic structure, we can insert any fact", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 626, 471, 641 ], "spans": [ { "bbox": [ 105, 626, 161, 641 ], "score": 1.0, "content": "directly once", "type": "text" }, { "bbox": [ 162, 628, 193, 640 ], "score": 0.92, "content": "( k _ { * } , v _ { * } )", "type": "inline_equation" }, { "bbox": [ 193, 626, 427, 641 ], "score": 1.0, "content": "is computed. All that remains is to choose the appropriate", "type": "text" }, { "bbox": [ 427, 628, 438, 639 ], "score": 0.89, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 438, 626, 456, 641 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 457, 629, 467, 639 ], "score": 0.84, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 467, 626, 471, 641 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 34 } ], "index": 32 }, { "type": "text", "bbox": [ 107, 644, 505, 699 ], "lines": [ { "bbox": [ 106, 644, 505, 656 ], "spans": [ { "bbox": [ 106, 644, 183, 656 ], "score": 1.0, "content": "Step 1: Choosing", "type": "text" }, { "bbox": [ 183, 644, 194, 655 ], "score": 0.86, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 194, 644, 505, 656 ], "score": 1.0, "content": "to Select the Subject. Based on the decisive role of MLP inputs at the final", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 655, 505, 667 ], "spans": [ { "bbox": [ 106, 655, 505, 667 ], "score": 1.0, "content": "subject token (Section 2), we shall choose inputs that represent the subject at its last token as the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 664, 506, 680 ], "spans": [ { "bbox": [ 105, 664, 155, 680 ], "score": 1.0, "content": "lookup key", "type": "text" }, { "bbox": [ 155, 666, 165, 677 ], "score": 0.88, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 166, 664, 277, 680 ], "score": 1.0, "content": ". Specifically, we compute", "type": "text" }, { "bbox": [ 277, 666, 288, 677 ], "score": 0.88, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 288, 664, 451, 680 ], "score": 1.0, "content": "by collecting activations: We pass text", "type": "text" }, { "bbox": [ 451, 668, 458, 676 ], "score": 0.75, "content": "x", "type": "inline_equation" }, { "bbox": [ 459, 664, 506, 680 ], "score": 1.0, "content": "containing", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 676, 505, 690 ], "spans": [ { "bbox": [ 105, 676, 153, 690 ], "score": 1.0, "content": "the subject", "type": "text" }, { "bbox": [ 153, 679, 159, 687 ], "score": 0.73, "content": "s", "type": "inline_equation" }, { "bbox": [ 159, 676, 194, 690 ], "score": 1.0, "content": "through", "type": "text" }, { "bbox": [ 195, 677, 203, 687 ], "score": 0.79, "content": "G", "type": "inline_equation" }, { "bbox": [ 204, 676, 260, 690 ], "score": 1.0, "content": "; then at layer", "type": "text" }, { "bbox": [ 261, 677, 269, 687 ], "score": 0.85, "content": "l ^ { * }", "type": "inline_equation" }, { "bbox": [ 270, 676, 386, 690 ], "score": 1.0, "content": "and last subject token index", "type": "text" }, { "bbox": [ 387, 678, 391, 687 ], "score": 0.76, "content": "i", "type": "inline_equation" }, { "bbox": [ 391, 676, 505, 690 ], "score": 1.0, "content": ", we read the value after the", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "non-linearity inside the MLP (Figure 4d). Because the state will vary depending on tokens that", "type": "text" } ], "index": 39 } ], "index": 37 } ], "page_idx": 4, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 711, 275, 722 ], "lines": [ { "bbox": [ 106, 709, 276, 723 ], "spans": [ { "bbox": [ 106, 709, 276, 723 ], "score": 1.0, "content": "6 Unrelated to keys and values in self-attention.", "type": "text" } ] } ] }, { "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": "image", "bbox": [ 105, 70, 506, 167 ], "blocks": [ { "type": "image_body", "bbox": [ 105, 70, 506, 167 ], "group_id": 0, "lines": [ { "bbox": [ 105, 70, 506, 167 ], "spans": [ { "bbox": [ 105, 70, 506, 167 ], "score": 0.966, "type": "image", "image_path": "eeb4c5f59c324d7747ed46f32e3dadf49334273e55fa708df9ff0730a7f294f8.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 105, 70, 506, 102.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 105, 102.33333333333334, 506, 134.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 105, 134.66666666666669, 506, 167.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 173, 505, 224 ], "group_id": 0, "lines": [ { "bbox": [ 106, 173, 505, 184 ], "spans": [ { "bbox": [ 106, 173, 505, 184 ], "score": 1.0, "content": "Figure 4: Editing one MLP layer with ROME. To associate Space Needle with Paris, the ROME method", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 182, 504, 195 ], "spans": [ { "bbox": [ 105, 182, 156, 195 ], "score": 1.0, "content": "inserts a new", "type": "text" }, { "bbox": [ 156, 183, 186, 194 ], "score": 0.92, "content": "( k _ { * } , v _ { * } )", "type": "inline_equation" }, { "bbox": [ 186, 182, 266, 195 ], "score": 1.0, "content": "association into layer", "type": "text" }, { "bbox": [ 266, 183, 274, 192 ], "score": 0.77, "content": "l ^ { * }", "type": "inline_equation" }, { "bbox": [ 274, 182, 329, 195 ], "score": 1.0, "content": ", where (a) key", "type": "text" }, { "bbox": [ 329, 183, 339, 192 ], "score": 0.87, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 340, 182, 493, 195 ], "score": 1.0, "content": "is determined by the subject and (b) value", "type": "text" }, { "bbox": [ 494, 185, 504, 192 ], "score": 0.82, "content": "v _ { * }", "type": "inline_equation" } ], "index": 4 }, { "bbox": [ 104, 192, 505, 205 ], "spans": [ { "bbox": [ 104, 192, 321, 205 ], "score": 1.0, "content": "is optimized to select the object. (c) Hidden state at layer", "type": "text" }, { "bbox": [ 321, 193, 329, 202 ], "score": 0.84, "content": "l ^ { * }", "type": "inline_equation" }, { "bbox": [ 330, 192, 369, 205 ], "score": 1.0, "content": "and token", "type": "text" }, { "bbox": [ 369, 194, 374, 202 ], "score": 0.64, "content": "_ { i }", "type": "inline_equation" }, { "bbox": [ 374, 192, 505, 205 ], "score": 1.0, "content": "is expanded to produce (d) the key", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 202, 506, 215 ], "spans": [ { "bbox": [ 105, 202, 131, 215 ], "score": 1.0, "content": "vector", "type": "text" }, { "bbox": [ 132, 204, 141, 213 ], "score": 0.86, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 142, 202, 312, 215 ], "score": 1.0, "content": "for the subject. (e) To write new value vector", "type": "text" }, { "bbox": [ 312, 204, 322, 213 ], "score": 0.86, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 322, 202, 506, 215 ], "score": 1.0, "content": "into the layer, (f) we calculate a rank-one update", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 209, 500, 228 ], "spans": [ { "bbox": [ 106, 213, 152, 224 ], "score": 0.92, "content": "\\Lambda ( C ^ { - 1 } k _ { * } ) ^ { T }", "type": "inline_equation" }, { "bbox": [ 153, 209, 185, 228 ], "score": 1.0, "content": "to cause", "type": "text" }, { "bbox": [ 186, 212, 240, 225 ], "score": 0.92, "content": "\\hat { W } _ { p r o j } ^ { ( l ) } \\hat { k } _ { * } = v _ { * }", "type": "inline_equation" }, { "bbox": [ 240, 209, 500, 228 ], "score": 1.0, "content": "⇤ while minimizing interference with other memories stored in the layer.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 107, 246, 505, 280 ], "lines": [], "index": 9, "bbox_fs": [ 105, 246, 505, 282 ], "lines_deleted": true }, { "type": "title", "bbox": [ 105, 303, 496, 318 ], "lines": [ { "bbox": [ 103, 302, 498, 320 ], "spans": [ { "bbox": [ 103, 302, 498, 320 ], "score": 1.0, "content": "3 Interventions on Weights for Understanding Factual Association Storage", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 335, 505, 390 ], "lines": [ { "bbox": [ 106, 335, 506, 347 ], "spans": [ { "bbox": [ 106, 335, 506, 347 ], "score": 1.0, "content": "While Causal Tracing has implicated MLP modules in recalling factual associations, we also wish to", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 346, 506, 358 ], "spans": [ { "bbox": [ 106, 346, 506, 358 ], "score": 1.0, "content": "understand how facts are stored in weights. Geva et al. (2021) observed that MLP layers (Figure 4cde)", "type": "text" } ], "index": 13 }, { "bbox": [ 103, 353, 509, 375 ], "spans": [ { "bbox": [ 103, 353, 434, 375 ], "score": 1.0, "content": "can act as two-layer key–value memories,6 where the neurons of the first layer (l)", "type": "text" }, { "bbox": [ 434, 356, 455, 371 ], "score": 0.92, "content": "\\mathbf { \\overline { { \\it W } } } _ { f c } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 456, 353, 509, 375 ], "score": 1.0, "content": "form a key,", "type": "text" } ], "index": 14 }, { "bbox": [ 104, 366, 507, 385 ], "spans": [ { "bbox": [ 104, 366, 220, 385 ], "score": 1.0, "content": "with which the second layer", "type": "text" }, { "bbox": [ 221, 367, 247, 382 ], "score": 0.93, "content": "W _ { p r o j } ^ { ( l ) }", "type": "inline_equation" }, { "bbox": [ 247, 366, 507, 385 ], "score": 1.0, "content": "retrieves an associated value. We hypothesize that MLPs can be", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 378, 496, 392 ], "spans": [ { "bbox": [ 105, 378, 496, 392 ], "score": 1.0, "content": "modeled as a linear associative memory; note that this differs from Geva et al.’s per-neuron view.", "type": "text" } ], "index": 16 } ], "index": 14, "bbox_fs": [ 103, 335, 509, 392 ] }, { "type": "text", "bbox": [ 106, 395, 505, 440 ], "lines": [ { "bbox": [ 105, 395, 505, 408 ], "spans": [ { "bbox": [ 105, 395, 505, 408 ], "score": 1.0, "content": "We test this hypothesis by conducting a new type of intervention: modifying factual associations with", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 406, 505, 419 ], "spans": [ { "bbox": [ 105, 406, 434, 419 ], "score": 1.0, "content": "Rank-One Model Editing (ROME). Being able to insert a new knowledge tuple", "type": "text" }, { "bbox": [ 434, 406, 493, 418 ], "score": 0.93, "content": "t ^ { * } = ( s , r , o ^ { * } )", "type": "inline_equation" }, { "bbox": [ 493, 406, 505, 419 ], "score": 1.0, "content": "in", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 417, 506, 430 ], "spans": [ { "bbox": [ 105, 417, 209, 430 ], "score": 1.0, "content": "place of the current tuple", "type": "text" }, { "bbox": [ 209, 417, 265, 429 ], "score": 0.93, "content": "t ^ { c } = \\left( s , r , o ^ { c } \\right)", "type": "inline_equation" }, { "bbox": [ 265, 417, 506, 430 ], "score": 1.0, "content": "with both generalization and specificity would demonstrate", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 429, 374, 441 ], "spans": [ { "bbox": [ 106, 429, 374, 441 ], "score": 1.0, "content": "fine-grained understanding of the association-storage mechanisms.", "type": "text" } ], "index": 20 } ], "index": 18.5, "bbox_fs": [ 105, 395, 506, 441 ] }, { "type": "title", "bbox": [ 106, 454, 489, 466 ], "lines": [ { "bbox": [ 104, 451, 491, 470 ], "spans": [ { "bbox": [ 104, 451, 491, 470 ], "score": 1.0, "content": "3.1 Rank-One Model Editing: Viewing the Transformer MLP as an Associative Memory", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 106, 473, 506, 552 ], "lines": [ { "bbox": [ 103, 471, 506, 489 ], "spans": [ { "bbox": [ 103, 471, 168, 489 ], "score": 1.0, "content": "We view W (l)proj", "type": "text" }, { "bbox": [ 170, 475, 506, 489 ], "score": 1.0, "content": "as a linear associative memory (Kohonen, 1972; Anderson, 1972). This perspective", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 485, 506, 498 ], "spans": [ { "bbox": [ 105, 485, 251, 498 ], "score": 1.0, "content": "observes that any linear operation", "type": "text" }, { "bbox": [ 251, 486, 263, 496 ], "score": 0.65, "content": "W", "type": "inline_equation" }, { "bbox": [ 263, 485, 506, 498 ], "score": 1.0, "content": "can operate as a key–value store for a set of vector keys", "type": "text" } ], "index": 23 }, { "bbox": [ 107, 496, 506, 510 ], "spans": [ { "bbox": [ 107, 497, 190, 509 ], "score": 0.92, "content": "K = [ k _ { 1 } \\ | \\ k _ { 2 } \\ | \\ . \\ . \\ . ]", "type": "inline_equation" }, { "bbox": [ 190, 496, 326, 510 ], "score": 1.0, "content": "and corresponding vector values", "type": "text" }, { "bbox": [ 327, 497, 408, 509 ], "score": 0.91, "content": "V = \\left[ v _ { 1 } \\mid v _ { 2 } \\mid \\ldots \\right]", "type": "inline_equation" }, { "bbox": [ 408, 496, 459, 510 ], "score": 1.0, "content": ", by solving", "type": "text" }, { "bbox": [ 459, 497, 502, 507 ], "score": 0.89, "content": "W K \\approx V", "type": "inline_equation" }, { "bbox": [ 503, 496, 506, 510 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 508, 506, 520 ], "spans": [ { "bbox": [ 105, 508, 412, 520 ], "score": 1.0, "content": "whose squared error is minimized using the Moore-Penrose pseudoinverse:", "type": "text" }, { "bbox": [ 413, 508, 461, 518 ], "score": 0.91, "content": "\\dot { W } = V { \\bar { K ^ { + } } }", "type": "inline_equation" }, { "bbox": [ 461, 508, 506, 520 ], "score": 1.0, "content": ". Bau et al.", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 518, 506, 532 ], "spans": [ { "bbox": [ 105, 518, 283, 532 ], "score": 1.0, "content": "(2020) observed that a new key–value pair", "type": "text" }, { "bbox": [ 283, 519, 315, 531 ], "score": 0.93, "content": "( k _ { * } , v _ { * } )", "type": "inline_equation" }, { "bbox": [ 315, 518, 506, 532 ], "score": 1.0, "content": "can be inserted optimally into the memory by", "type": "text" } ], "index": 26 }, { "bbox": [ 104, 528, 506, 544 ], "spans": [ { "bbox": [ 104, 528, 506, 544 ], "score": 1.0, "content": "solving a constrained least-squares problem. In a convolutional network, Bau et al. solve this using", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 541, 447, 552 ], "spans": [ { "bbox": [ 105, 541, 447, 552 ], "score": 1.0, "content": "an optimization, but in a fully-connected layer, we can derive a closed form solution:", "type": "text" } ], "index": 28 } ], "index": 25, "bbox_fs": [ 103, 471, 506, 552 ] }, { "type": "interline_equation", "bbox": [ 175, 559, 469, 575 ], "lines": [ { "bbox": [ 175, 559, 469, 575 ], "spans": [ { "bbox": [ 175, 559, 469, 575 ], "score": 0.66, "content": "\\mathrm { ~ e ~ } \\Vert \\hat { W } K - V \\Vert \\mathrm { ~ s u c h ~ t h a t ~ } \\hat { W } k _ { * } = v _ { * } \\quad \\mathrm { b y ~ s e t t i n g ~ } \\hat { W } = W + \\Lambda ( C ^ { - 1 } k _ { * } ) ^ { T } .", "type": "interline_equation", "image_path": "812a0ecccfad6534321ab6f067c0a9e82bc6dd7692683887a95ab0e4a7320554.jpg" } ] } ], "index": 29, "virtual_lines": [ { "bbox": [ 175, 559, 469, 575 ], "spans": [], "index": 29 } ] }, { "type": "text", "bbox": [ 106, 583, 504, 639 ], "lines": [ { "bbox": [ 105, 583, 506, 596 ], "spans": [ { "bbox": [ 105, 583, 127, 596 ], "score": 1.0, "content": "Here", "type": "text" }, { "bbox": [ 128, 584, 140, 594 ], "score": 0.57, "content": "W", "type": "inline_equation" }, { "bbox": [ 140, 583, 225, 596 ], "score": 1.0, "content": "is the original matrix,", "type": "text" }, { "bbox": [ 225, 583, 271, 594 ], "score": 0.92, "content": "C = K K ^ { T }", "type": "inline_equation" }, { "bbox": [ 272, 583, 506, 596 ], "score": 1.0, "content": "is a constant that we pre-cache by estimating the uncentered", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 593, 503, 608 ], "spans": [ { "bbox": [ 105, 593, 161, 608 ], "score": 1.0, "content": "covariance of", "type": "text" }, { "bbox": [ 161, 595, 168, 605 ], "score": 0.83, "content": "k", "type": "inline_equation" }, { "bbox": [ 168, 593, 378, 608 ], "score": 1.0, "content": "from a sample of Wikipedia text (Appendix E.5), and", "type": "text" }, { "bbox": [ 378, 594, 503, 607 ], "score": 0.92, "content": "\\Lambda = \\mathbf { \\bar { \\Phi } } ( v _ { * } - W k _ { * } ) / ( C ^ { - 1 } k _ { * } ) ^ { T } k _ { * }", "type": "inline_equation" } ], "index": 31 }, { "bbox": [ 106, 606, 505, 618 ], "spans": [ { "bbox": [ 106, 606, 505, 618 ], "score": 1.0, "content": "is a vector proportional to the residual error of the new key–value pair on the original memory matrix", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 616, 506, 630 ], "spans": [ { "bbox": [ 105, 616, 506, 630 ], "score": 1.0, "content": "(full derivation in Appendix A). Because of this simple algebraic structure, we can insert any fact", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 626, 471, 641 ], "spans": [ { "bbox": [ 105, 626, 161, 641 ], "score": 1.0, "content": "directly once", "type": "text" }, { "bbox": [ 162, 628, 193, 640 ], "score": 0.92, "content": "( k _ { * } , v _ { * } )", "type": "inline_equation" }, { "bbox": [ 193, 626, 427, 641 ], "score": 1.0, "content": "is computed. All that remains is to choose the appropriate", "type": "text" }, { "bbox": [ 427, 628, 438, 639 ], "score": 0.89, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 438, 626, 456, 641 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 457, 629, 467, 639 ], "score": 0.84, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 467, 626, 471, 641 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 34 } ], "index": 32, "bbox_fs": [ 105, 583, 506, 641 ] }, { "type": "text", "bbox": [ 107, 644, 505, 699 ], "lines": [ { "bbox": [ 106, 644, 505, 656 ], "spans": [ { "bbox": [ 106, 644, 183, 656 ], "score": 1.0, "content": "Step 1: Choosing", "type": "text" }, { "bbox": [ 183, 644, 194, 655 ], "score": 0.86, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 194, 644, 505, 656 ], "score": 1.0, "content": "to Select the Subject. Based on the decisive role of MLP inputs at the final", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 655, 505, 667 ], "spans": [ { "bbox": [ 106, 655, 505, 667 ], "score": 1.0, "content": "subject token (Section 2), we shall choose inputs that represent the subject at its last token as the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 664, 506, 680 ], "spans": [ { "bbox": [ 105, 664, 155, 680 ], "score": 1.0, "content": "lookup key", "type": "text" }, { "bbox": [ 155, 666, 165, 677 ], "score": 0.88, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 166, 664, 277, 680 ], "score": 1.0, "content": ". Specifically, we compute", "type": "text" }, { "bbox": [ 277, 666, 288, 677 ], "score": 0.88, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 288, 664, 451, 680 ], "score": 1.0, "content": "by collecting activations: We pass text", "type": "text" }, { "bbox": [ 451, 668, 458, 676 ], "score": 0.75, "content": "x", "type": "inline_equation" }, { "bbox": [ 459, 664, 506, 680 ], "score": 1.0, "content": "containing", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 676, 505, 690 ], "spans": [ { "bbox": [ 105, 676, 153, 690 ], "score": 1.0, "content": "the subject", "type": "text" }, { "bbox": [ 153, 679, 159, 687 ], "score": 0.73, "content": "s", "type": "inline_equation" }, { "bbox": [ 159, 676, 194, 690 ], "score": 1.0, "content": "through", "type": "text" }, { "bbox": [ 195, 677, 203, 687 ], "score": 0.79, "content": "G", "type": "inline_equation" }, { "bbox": [ 204, 676, 260, 690 ], "score": 1.0, "content": "; then at layer", "type": "text" }, { "bbox": [ 261, 677, 269, 687 ], "score": 0.85, "content": "l ^ { * }", "type": "inline_equation" }, { "bbox": [ 270, 676, 386, 690 ], "score": 1.0, "content": "and last subject token index", "type": "text" }, { "bbox": [ 387, 678, 391, 687 ], "score": 0.76, "content": "i", "type": "inline_equation" }, { "bbox": [ 391, 676, 505, 690 ], "score": 1.0, "content": ", we read the value after the", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "non-linearity inside the MLP (Figure 4d). Because the state will vary depending on tokens that", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 72, 498, 86 ], "spans": [ { "bbox": [ 105, 72, 140, 86 ], "score": 1.0, "content": "precede", "type": "text", "cross_page": true }, { "bbox": [ 140, 75, 146, 83 ], "score": 0.77, "content": "s", "type": "inline_equation", "cross_page": true }, { "bbox": [ 146, 72, 204, 86 ], "score": 1.0, "content": "in text, we set", "type": "text", "cross_page": true }, { "bbox": [ 205, 73, 216, 83 ], "score": 0.88, "content": "k _ { * }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 216, 72, 488, 86 ], "score": 1.0, "content": "to an average value over a small set of texts ending with the subject", "type": "text", "cross_page": true }, { "bbox": [ 488, 75, 493, 82 ], "score": 0.72, "content": "s", "type": "inline_equation", "cross_page": true }, { "bbox": [ 494, 72, 498, 86 ], "score": 1.0, "content": ":", "type": "text", "cross_page": true } ], "index": 0 } ], "index": 37, "bbox_fs": [ 105, 644, 506, 700 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 103, 72, 497, 85 ], "lines": [ { "bbox": [ 105, 72, 498, 86 ], "spans": [ { "bbox": [ 105, 72, 140, 86 ], "score": 1.0, "content": "precede", "type": "text" }, { "bbox": [ 140, 75, 146, 83 ], "score": 0.77, "content": "s", "type": "inline_equation" }, { "bbox": [ 146, 72, 204, 86 ], "score": 1.0, "content": "in text, we set", "type": "text" }, { "bbox": [ 205, 73, 216, 83 ], "score": 0.88, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 216, 72, 488, 86 ], "score": 1.0, "content": "to an average value over a small set of texts ending with the subject", "type": "text" }, { "bbox": [ 488, 75, 493, 82 ], "score": 0.72, "content": "s", "type": "inline_equation" }, { "bbox": [ 494, 72, 498, 86 ], "score": 1.0, "content": ":", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "interline_equation", "bbox": [ 161, 90, 450, 126 ], "lines": [ { "bbox": [ 161, 90, 450, 126 ], "spans": [ { "bbox": [ 161, 90, 450, 126 ], "score": 0.92, "content": "k _ { * } = \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } k ( x _ { j } + s ) , \\mathrm { ~ w h e r e ~ } k ( x ) = \\sigma \\left( W _ { f c } ^ { ( l ^ { * } ) } \\gamma ( a _ { [ x ] , i } ^ { ( l ^ { * } ) } + h _ { [ x ] , i } ^ { ( l ^ { * } - 1 ) } ) \\right) .", "type": "interline_equation", "image_path": "d4f6daa2319c1c15912713da18554698b275ac89ab2709b22950b690f1995056.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 161, 90, 450, 102.0 ], "spans": [], "index": 1 }, { "bbox": [ 161, 102.0, 450, 114.0 ], "spans": [], "index": 2 }, { "bbox": [ 161, 114.0, 450, 126.0 ], "spans": [], "index": 3 } ] }, { "type": "text", "bbox": [ 105, 131, 485, 143 ], "lines": [ { "bbox": [ 105, 129, 482, 145 ], "spans": [ { "bbox": [ 105, 129, 198, 145 ], "score": 1.0, "content": "In practice, we sample", "type": "text" }, { "bbox": [ 199, 133, 209, 144 ], "score": 0.87, "content": "x _ { j }", "type": "inline_equation" }, { "bbox": [ 210, 129, 473, 145 ], "score": 1.0, "content": "by generating 50 random token sequences of length 2 to 10 using", "type": "text" }, { "bbox": [ 474, 132, 482, 141 ], "score": 0.8, "content": "G", "type": "inline_equation" } ], "index": 4 } ], "index": 4 }, { "type": "text", "bbox": [ 104, 147, 505, 171 ], "lines": [ { "bbox": [ 106, 147, 505, 160 ], "spans": [ { "bbox": [ 106, 147, 180, 160 ], "score": 1.0, "content": "Step 2: Choosing", "type": "text" }, { "bbox": [ 180, 150, 190, 158 ], "score": 0.84, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 191, 147, 442, 160 ], "score": 1.0, "content": "to Recall the Fact. Next, we wish to choose some vector value", "type": "text" }, { "bbox": [ 442, 150, 452, 158 ], "score": 0.86, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 453, 147, 505, 160 ], "score": 1.0, "content": "that encodes", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 158, 504, 171 ], "spans": [ { "bbox": [ 106, 158, 173, 171 ], "score": 1.0, "content": "the new relation", "type": "text" }, { "bbox": [ 174, 159, 200, 171 ], "score": 0.92, "content": "( r , o ^ { * } )", "type": "inline_equation" }, { "bbox": [ 200, 158, 266, 171 ], "score": 1.0, "content": "as a property of", "type": "text" }, { "bbox": [ 266, 160, 272, 168 ], "score": 0.73, "content": "s", "type": "inline_equation" }, { "bbox": [ 272, 158, 305, 171 ], "score": 1.0, "content": ". We set", "type": "text" }, { "bbox": [ 306, 159, 385, 171 ], "score": 0.92, "content": "v _ { * } = \\mathrm { a r g m i n } _ { z } \\mathcal { L } ( z )", "type": "inline_equation" }, { "bbox": [ 386, 158, 469, 171 ], "score": 1.0, "content": ", where the objective", "type": "text" }, { "bbox": [ 470, 158, 490, 171 ], "score": 0.92, "content": "\\mathcal { L } ( z )", "type": "inline_equation" }, { "bbox": [ 491, 158, 504, 171 ], "score": 1.0, "content": "is:", "type": "text" } ], "index": 6 } ], "index": 5.5 }, { "type": "interline_equation", "bbox": [ 133, 176, 478, 219 ], "lines": [ { "bbox": [ 133, 176, 478, 219 ], "spans": [ { "bbox": [ 133, 176, 478, 219 ], "score": 0.91, "content": "\\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\underbrace { - \\log { \\mathbb { P } } _ { G ( m _ { i } ^ { ( t ^ { * } ) } : = z ) } [ o ^ { * } \\mid x _ { j } + p ] } _ { \\mathrm { ( a ) M a x i m i z i n g ~ \\textstyle o ^ { * } ~ p r o b a b i l i t y } } + \\underbrace { D _ { \\mathrm { K L } } ( \\mathbb { P } _ { G ( m _ { i ^ { \\prime } } ^ { ( t ^ { * } ) } : = z ) } [ x \\mid p ^ { \\prime } ] \\| \\mathbb { P } _ { G } [ x \\mid p ^ { \\prime } ] ) } _ { \\mathrm { ( b ) C o n r o l i n g ~ e s s e n c e d i r t } } .", "type": "interline_equation", "image_path": "0c6fb3c47350f248ba3314f3eab9be5916393c05b4a9d4dadddd23edf8560b8c.jpg" } ] } ], "index": 8, "virtual_lines": [ { "bbox": [ 133, 176, 478, 190.33333333333334 ], "spans": [], "index": 7 }, { "bbox": [ 133, 190.33333333333334, 478, 204.66666666666669 ], "spans": [], "index": 8 }, { "bbox": [ 133, 204.66666666666669, 478, 219.00000000000003 ], "spans": [], "index": 9 } ] }, { "type": "text", "bbox": [ 106, 225, 505, 314 ], "lines": [ { "bbox": [ 105, 225, 505, 237 ], "spans": [ { "bbox": [ 105, 225, 259, 237 ], "score": 1.0, "content": "The first term (Eqn. 4a) seeks a vector", "type": "text" }, { "bbox": [ 260, 227, 266, 235 ], "score": 0.76, "content": "z", "type": "inline_equation" }, { "bbox": [ 267, 225, 505, 237 ], "score": 1.0, "content": "that, when substituted as the output of the MLP at the token", "type": "text" } ], "index": 10 }, { "bbox": [ 107, 235, 506, 250 ], "spans": [ { "bbox": [ 107, 238, 111, 246 ], "score": 0.77, "content": "i", "type": "inline_equation" }, { "bbox": [ 111, 235, 243, 250 ], "score": 1.0, "content": "at the end of the subject (notated", "type": "text" }, { "bbox": [ 243, 235, 303, 249 ], "score": 0.89, "content": "G ( m _ { i } ^ { ( l ^ { * } ) } : = z ) ^ { \\backslash }", "type": "inline_equation" }, { "bbox": [ 303, 235, 506, 250 ], "score": 1.0, "content": "), will cause the network to predict the target object", "type": "text" } ], "index": 11 }, { "bbox": [ 107, 246, 506, 260 ], "spans": [ { "bbox": [ 107, 248, 117, 258 ], "score": 0.84, "content": "o ^ { * }", "type": "inline_equation" }, { "bbox": [ 117, 246, 252, 260 ], "score": 1.0, "content": "in response to the factual prompt", "type": "text" }, { "bbox": [ 253, 250, 259, 259 ], "score": 0.79, "content": "p", "type": "inline_equation" }, { "bbox": [ 259, 246, 506, 260 ], "score": 1.0, "content": ". The second term (Eqn. 4b) minimizes the KL divergence of", "type": "text" } ], "index": 12 }, { "bbox": [ 104, 258, 506, 271 ], "spans": [ { "bbox": [ 104, 258, 216, 271 ], "score": 1.0, "content": "predictions for the prompt", "type": "text" }, { "bbox": [ 216, 258, 226, 270 ], "score": 0.86, "content": "p ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 226, 258, 280, 271 ], "score": 1.0, "content": "(of the form", "type": "text" }, { "bbox": [ 280, 258, 324, 270 ], "score": 0.52, "content": "\\mathbf { \\cdots } \\{ \\mathrm { s u b j e c t } \\}", "type": "inline_equation" }, { "bbox": [ 324, 258, 506, 271 ], "score": 1.0, "content": "is a”) to the unchanged model, which helps", "type": "text" } ], "index": 13 }, { "bbox": [ 104, 269, 506, 282 ], "spans": [ { "bbox": [ 104, 269, 506, 282 ], "score": 1.0, "content": "preserve the model’s understanding of the subject’s essence. To be clear, the optimization does not", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 280, 506, 293 ], "spans": [ { "bbox": [ 105, 280, 362, 293 ], "score": 1.0, "content": "directly alter model weights; it identifies a vector representation", "type": "text" }, { "bbox": [ 363, 282, 373, 291 ], "score": 0.86, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 374, 280, 506, 293 ], "score": 1.0, "content": "that, when output at the targeted", "type": "text" } ], "index": 15 }, { "bbox": [ 104, 290, 507, 304 ], "spans": [ { "bbox": [ 104, 290, 276, 304 ], "score": 1.0, "content": "MLP module, represents the new property", "type": "text" }, { "bbox": [ 276, 291, 302, 303 ], "score": 0.92, "content": "( r , o ^ { * } )", "type": "inline_equation" }, { "bbox": [ 302, 290, 362, 304 ], "score": 1.0, "content": "for the subject", "type": "text" }, { "bbox": [ 362, 293, 368, 301 ], "score": 0.35, "content": "s", "type": "inline_equation" }, { "bbox": [ 369, 290, 453, 304 ], "score": 1.0, "content": ". Note that, similar to", "type": "text" }, { "bbox": [ 454, 291, 465, 302 ], "score": 0.89, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 465, 290, 507, 304 ], "score": 1.0, "content": "selection,", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 302, 507, 315 ], "spans": [ { "bbox": [ 106, 304, 117, 313 ], "score": 0.83, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 118, 302, 300, 315 ], "score": 1.0, "content": "optimization also uses the random prefix texts", "type": "text" }, { "bbox": [ 300, 304, 311, 314 ], "score": 0.87, "content": "x _ { j }", "type": "inline_equation" }, { "bbox": [ 311, 302, 507, 315 ], "score": 1.0, "content": "to encourage robustness under differing contexts.", "type": "text" } ], "index": 17 } ], "index": 13.5 }, { "type": "text", "bbox": [ 106, 318, 504, 352 ], "lines": [ { "bbox": [ 105, 316, 506, 333 ], "spans": [ { "bbox": [ 105, 316, 368, 333 ], "score": 1.0, "content": "Step 3: Inserting the Fact. Once we have computed the pair", "type": "text" }, { "bbox": [ 369, 319, 401, 329 ], "score": 0.86, "content": "( k _ { * } , v _ { * } )", "type": "inline_equation" }, { "bbox": [ 401, 316, 506, 333 ], "score": 1.0, "content": "to represent the full fact", "type": "text" } ], "index": 18 }, { "bbox": [ 103, 326, 506, 344 ], "spans": [ { "bbox": [ 103, 326, 350, 344 ], "score": 1.0, "content": "(s, r, o⇤), we apply Eqn. 2, updating the MLP weights W (l)proj", "type": "text" }, { "bbox": [ 350, 329, 506, 343 ], "score": 1.0, "content": "with a rank-one update that inserts the", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 340, 454, 353 ], "spans": [ { "bbox": [ 106, 340, 454, 353 ], "score": 1.0, "content": "new key–value association directly. For full implementation details, see Appendix E.5.", "type": "text" } ], "index": 20 } ], "index": 19 }, { "type": "title", "bbox": [ 106, 364, 376, 376 ], "lines": [ { "bbox": [ 105, 364, 377, 379 ], "spans": [ { "bbox": [ 105, 364, 377, 379 ], "score": 1.0, "content": "3.2 Evaluating ROME: Zero-Shot Relation Extraction (zsRE)", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 107, 384, 504, 407 ], "lines": [ { "bbox": [ 105, 384, 505, 397 ], "spans": [ { "bbox": [ 105, 384, 505, 397 ], "score": 1.0, "content": "We wish to test our localized factual association hypothesis: can storing a single new vector association", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 396, 430, 407 ], "spans": [ { "bbox": [ 106, 396, 430, 407 ], "score": 1.0, "content": "using ROME insert a substantial, generalized factual association into the model?", "type": "text" } ], "index": 23 } ], "index": 22.5 }, { "type": "text", "bbox": [ 106, 412, 505, 489 ], "lines": [ { "bbox": [ 105, 412, 506, 425 ], "spans": [ { "bbox": [ 105, 412, 506, 425 ], "score": 1.0, "content": "A natural question is how ROME compares to other model-editing methods, which use direct", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 424, 506, 436 ], "spans": [ { "bbox": [ 106, 424, 506, 436 ], "score": 1.0, "content": "optimization or hypernetworks to incorporate a single new training example into a network. For", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 434, 505, 447 ], "spans": [ { "bbox": [ 106, 434, 505, 447 ], "score": 1.0, "content": "baselines, we examine Fine-Tuning (FT), which applies Adam with early stopping at one layer to", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 444, 506, 458 ], "spans": [ { "bbox": [ 105, 444, 146, 458 ], "score": 1.0, "content": "minimize", "type": "text" }, { "bbox": [ 146, 445, 207, 457 ], "score": 0.9, "content": "- \\log \\mathbb { P } \\left[ o ^ { * } \\mid x \\right]", "type": "inline_equation" }, { "bbox": [ 208, 444, 312, 458 ], "score": 1.0, "content": ". Constrained Fine-Tuning", "type": "text" }, { "bbox": [ 312, 445, 343, 456 ], "score": 0.71, "content": "\\mathbf { \\left( F T + L \\right) }", "type": "inline_equation" }, { "bbox": [ 343, 444, 506, 458 ], "score": 1.0, "content": "(Zhu et al., 2020) additionally imposes a", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 456, 506, 469 ], "spans": [ { "bbox": [ 105, 456, 172, 469 ], "score": 1.0, "content": "parameter-space", "type": "text" }, { "bbox": [ 172, 457, 189, 467 ], "score": 0.88, "content": "L _ { \\infty }", "type": "inline_equation" }, { "bbox": [ 189, 456, 506, 469 ], "score": 1.0, "content": "norm constraint on weight changes. We also test two hypernetworks: Knowledge", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 466, 505, 480 ], "spans": [ { "bbox": [ 106, 466, 135, 480 ], "score": 1.0, "content": "Editor", "type": "text" }, { "bbox": [ 135, 467, 156, 478 ], "score": 0.27, "content": "\\mathbf { ( K E ) }", "type": "inline_equation" }, { "bbox": [ 157, 466, 505, 480 ], "score": 1.0, "content": "(De Cao et al., 2021) and MEND (Mitchell et al., 2021), both of which learn auxiliary", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 478, 445, 490 ], "spans": [ { "bbox": [ 106, 478, 252, 490 ], "score": 1.0, "content": "models to predict weight changes to", "type": "text" }, { "bbox": [ 253, 478, 262, 488 ], "score": 0.7, "content": "G", "type": "inline_equation" }, { "bbox": [ 262, 478, 445, 490 ], "score": 1.0, "content": ". Further details are described in Appendix E.", "type": "text" } ], "index": 30 } ], "index": 27 }, { "type": "text", "bbox": [ 107, 494, 297, 635 ], "lines": [ { "bbox": [ 107, 495, 298, 505 ], "spans": [ { "bbox": [ 107, 495, 298, 505 ], "score": 1.0, "content": "We first evaluate ROME on the Zero-Shot Re-", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 505, 297, 516 ], "spans": [ { "bbox": [ 106, 505, 297, 516 ], "score": 1.0, "content": "lation Extraction (zsRE) task used in Mitchell", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 516, 298, 527 ], "spans": [ { "bbox": [ 105, 516, 298, 527 ], "score": 1.0, "content": "et al. (2021) and De Cao et al. (2021). Our", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 527, 297, 538 ], "spans": [ { "bbox": [ 106, 527, 297, 538 ], "score": 1.0, "content": "evaluation slice contains 10,000 records, each", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 538, 298, 550 ], "spans": [ { "bbox": [ 106, 538, 298, 550 ], "score": 1.0, "content": "containing one factual statement, its paraphrase,", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 549, 298, 561 ], "spans": [ { "bbox": [ 105, 549, 298, 561 ], "score": 1.0, "content": "and one unrelated factual statement. “Efficacy”", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 559, 297, 572 ], "spans": [ { "bbox": [ 105, 559, 297, 572 ], "score": 1.0, "content": "and “Paraphrase” measure post-edit accuracy", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 569, 298, 583 ], "spans": [ { "bbox": [ 106, 570, 208, 583 ], "score": 0.91, "content": "\\mathbb { I } \\big [ o ^ { * } = \\mathrm { a r g m a x } _ { o } \\mathbb { P } _ { G ^ { \\prime } } \\left[ o \\right] \\big ]", "type": "inline_equation" }, { "bbox": [ 208, 569, 298, 583 ], "score": 1.0, "content": "of the statement and", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 581, 298, 594 ], "spans": [ { "bbox": [ 105, 581, 298, 594 ], "score": 1.0, "content": "its paraphrase, respectively, while “Specificity”", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 592, 298, 604 ], "spans": [ { "bbox": [ 105, 592, 298, 604 ], "score": 1.0, "content": "measures the edited model’s accuracy on an un-", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 603, 297, 615 ], "spans": [ { "bbox": [ 106, 603, 297, 615 ], "score": 1.0, "content": "related fact. Table 1 shows the results: ROME is", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 614, 297, 627 ], "spans": [ { "bbox": [ 105, 614, 297, 627 ], "score": 1.0, "content": "competitive with hypernetworks and fine-tuning", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 625, 297, 637 ], "spans": [ { "bbox": [ 106, 625, 297, 637 ], "score": 1.0, "content": "methods despite its simplicity. We find that it", "type": "text" } ], "index": 43 } ], "index": 37 }, { "type": "table", "bbox": [ 305, 507, 504, 623 ], "blocks": [ { "type": "table_caption", "bbox": [ 321, 495, 486, 506 ], "group_id": 0, "lines": [ { "bbox": [ 320, 494, 488, 508 ], "spans": [ { "bbox": [ 320, 494, 488, 508 ], "score": 1.0, "content": "Table 1: zsRE Editing Results on GPT-2 XL.", "type": "text" } ], "index": 44 } ], "index": 44 }, { "type": "table_body", "bbox": [ 305, 507, 504, 623 ], "group_id": 0, "lines": [ { "bbox": [ 305, 507, 504, 623 ], "spans": [ { "bbox": [ 305, 507, 504, 623 ], "score": 0.98, "html": "
EditorEfficacy 个 Paraphrase 个 Specificity 个
GPT-2 XL22.2 (±0.5) 21.3 (±0.5) 24.2 (±0.5)
FT99.6 (±0.1) 82.1 (±0.6) 23.2(±0.5)
FT+L92.3 (±0.4) 47.2 (±0.7) 23.4(±0.5)
KE65.5 (±0.6) 61.4(±0.6) 24.9 (±0.5)
KE-zsRE92.4 (±0.3) 90.0 (±0.3) 23.8 (±0.5)
MEND75.9 (±0.5) 65.3 (±0.6) 24.1(±0.5)
MEND-zsRE 99.4 (±0.1)99.3 (±0.1) 24.1(±0.5)
ROME99.8 (±0.0) 88.1(±0.5) 24.2 (±0.5)
", "type": "table", "image_path": "0e85568b3db85a4ec77f72778120f03fa7931c55642ffa5391e3f2bdfa77943d.jpg" } ] } ], "index": 48.5, "virtual_lines": [ { "bbox": [ 305, 507, 504, 521.5 ], "spans": [], "index": 45 }, { "bbox": [ 305, 521.5, 504, 536.0 ], "spans": [], "index": 46 }, { "bbox": [ 305, 536.0, 504, 550.5 ], "spans": [], "index": 47 }, { "bbox": [ 305, 550.5, 504, 565.0 ], "spans": [], "index": 48 }, { "bbox": [ 305, 565.0, 504, 579.5 ], "spans": [], "index": 49 }, { "bbox": [ 305, 579.5, 504, 594.0 ], "spans": [], "index": 50 }, { "bbox": [ 305, 594.0, 504, 608.5 ], "spans": [], "index": 51 }, { "bbox": [ 305, 608.5, 504, 623.0 ], "spans": [], "index": 52 } ] } ], "index": 46.25 }, { "type": "text", "bbox": [ 107, 636, 505, 702 ], "lines": [ { "bbox": [ 105, 635, 505, 649 ], "spans": [ { "bbox": [ 105, 635, 505, 649 ], "score": 1.0, "content": "is not hard for ROME to insert an association that can be regurgitated by the model. Robustness", "type": "text" } ], "index": 53 }, { "bbox": [ 106, 648, 505, 659 ], "spans": [ { "bbox": [ 106, 648, 505, 659 ], "score": 1.0, "content": "under paraphrase is also strong, although it comes short of custom-tuned hyperparameter networks", "type": "text" } ], "index": 54 }, { "bbox": [ 105, 657, 506, 669 ], "spans": [ { "bbox": [ 105, 657, 506, 669 ], "score": 1.0, "content": "KE-zsRE and MEND-zsRE, which we explicitly trained on the zsRE data distribution.7 We find that", "type": "text" } ], "index": 55 }, { "bbox": [ 105, 668, 505, 681 ], "spans": [ { "bbox": [ 105, 668, 505, 681 ], "score": 1.0, "content": "zsRE’s specificity score is not a sensitive measure of model damage, since these prompts are sampled", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 678, 506, 694 ], "spans": [ { "bbox": [ 105, 678, 506, 694 ], "score": 1.0, "content": "from a large space of possible facts, whereas bleedover is most likely to occur on related neighboring", "type": "text" } ], "index": 57 }, { "bbox": [ 106, 691, 338, 702 ], "spans": [ { "bbox": [ 106, 691, 338, 702 ], "score": 1.0, "content": "subjects. Appendix C has additional experimental details.", "type": "text" } ], "index": 58 } ], "index": 55.5 } ], "page_idx": 5, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 105, 711, 493, 722 ], "lines": [ { "bbox": [ 107, 709, 494, 723 ], "spans": [ { "bbox": [ 107, 709, 494, 723 ], "score": 1.0, "content": "7 Out-of-the-box, they are trained on a WikiText generation task (Mitchell et al., 2021; De Cao et al., 2021).", "type": "text" } ] } ] }, { "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": "text", "bbox": [ 103, 72, 497, 85 ], "lines": [], "index": 0, "bbox_fs": [ 105, 72, 498, 86 ], "lines_deleted": true }, { "type": "interline_equation", "bbox": [ 161, 90, 450, 126 ], "lines": [ { "bbox": [ 161, 90, 450, 126 ], "spans": [ { "bbox": [ 161, 90, 450, 126 ], "score": 0.92, "content": "k _ { * } = \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } k ( x _ { j } + s ) , \\mathrm { ~ w h e r e ~ } k ( x ) = \\sigma \\left( W _ { f c } ^ { ( l ^ { * } ) } \\gamma ( a _ { [ x ] , i } ^ { ( l ^ { * } ) } + h _ { [ x ] , i } ^ { ( l ^ { * } - 1 ) } ) \\right) .", "type": "interline_equation", "image_path": "d4f6daa2319c1c15912713da18554698b275ac89ab2709b22950b690f1995056.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 161, 90, 450, 102.0 ], "spans": [], "index": 1 }, { "bbox": [ 161, 102.0, 450, 114.0 ], "spans": [], "index": 2 }, { "bbox": [ 161, 114.0, 450, 126.0 ], "spans": [], "index": 3 } ] }, { "type": "text", "bbox": [ 105, 131, 485, 143 ], "lines": [ { "bbox": [ 105, 129, 482, 145 ], "spans": [ { "bbox": [ 105, 129, 198, 145 ], "score": 1.0, "content": "In practice, we sample", "type": "text" }, { "bbox": [ 199, 133, 209, 144 ], "score": 0.87, "content": "x _ { j }", "type": "inline_equation" }, { "bbox": [ 210, 129, 473, 145 ], "score": 1.0, "content": "by generating 50 random token sequences of length 2 to 10 using", "type": "text" }, { "bbox": [ 474, 132, 482, 141 ], "score": 0.8, "content": "G", "type": "inline_equation" } ], "index": 4 } ], "index": 4, "bbox_fs": [ 105, 129, 482, 145 ] }, { "type": "text", "bbox": [ 104, 147, 505, 171 ], "lines": [ { "bbox": [ 106, 147, 505, 160 ], "spans": [ { "bbox": [ 106, 147, 180, 160 ], "score": 1.0, "content": "Step 2: Choosing", "type": "text" }, { "bbox": [ 180, 150, 190, 158 ], "score": 0.84, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 191, 147, 442, 160 ], "score": 1.0, "content": "to Recall the Fact. Next, we wish to choose some vector value", "type": "text" }, { "bbox": [ 442, 150, 452, 158 ], "score": 0.86, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 453, 147, 505, 160 ], "score": 1.0, "content": "that encodes", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 158, 504, 171 ], "spans": [ { "bbox": [ 106, 158, 173, 171 ], "score": 1.0, "content": "the new relation", "type": "text" }, { "bbox": [ 174, 159, 200, 171 ], "score": 0.92, "content": "( r , o ^ { * } )", "type": "inline_equation" }, { "bbox": [ 200, 158, 266, 171 ], "score": 1.0, "content": "as a property of", "type": "text" }, { "bbox": [ 266, 160, 272, 168 ], "score": 0.73, "content": "s", "type": "inline_equation" }, { "bbox": [ 272, 158, 305, 171 ], "score": 1.0, "content": ". We set", "type": "text" }, { "bbox": [ 306, 159, 385, 171 ], "score": 0.92, "content": "v _ { * } = \\mathrm { a r g m i n } _ { z } \\mathcal { L } ( z )", "type": "inline_equation" }, { "bbox": [ 386, 158, 469, 171 ], "score": 1.0, "content": ", where the objective", "type": "text" }, { "bbox": [ 470, 158, 490, 171 ], "score": 0.92, "content": "\\mathcal { L } ( z )", "type": "inline_equation" }, { "bbox": [ 491, 158, 504, 171 ], "score": 1.0, "content": "is:", "type": "text" } ], "index": 6 } ], "index": 5.5, "bbox_fs": [ 106, 147, 505, 171 ] }, { "type": "interline_equation", "bbox": [ 133, 176, 478, 219 ], "lines": [ { "bbox": [ 133, 176, 478, 219 ], "spans": [ { "bbox": [ 133, 176, 478, 219 ], "score": 0.91, "content": "\\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\underbrace { - \\log { \\mathbb { P } } _ { G ( m _ { i } ^ { ( t ^ { * } ) } : = z ) } [ o ^ { * } \\mid x _ { j } + p ] } _ { \\mathrm { ( a ) M a x i m i z i n g ~ \\textstyle o ^ { * } ~ p r o b a b i l i t y } } + \\underbrace { D _ { \\mathrm { K L } } ( \\mathbb { P } _ { G ( m _ { i ^ { \\prime } } ^ { ( t ^ { * } ) } : = z ) } [ x \\mid p ^ { \\prime } ] \\| \\mathbb { P } _ { G } [ x \\mid p ^ { \\prime } ] ) } _ { \\mathrm { ( b ) C o n r o l i n g ~ e s s e n c e d i r t } } .", "type": "interline_equation", "image_path": "0c6fb3c47350f248ba3314f3eab9be5916393c05b4a9d4dadddd23edf8560b8c.jpg" } ] } ], "index": 8, "virtual_lines": [ { "bbox": [ 133, 176, 478, 190.33333333333334 ], "spans": [], "index": 7 }, { "bbox": [ 133, 190.33333333333334, 478, 204.66666666666669 ], "spans": [], "index": 8 }, { "bbox": [ 133, 204.66666666666669, 478, 219.00000000000003 ], "spans": [], "index": 9 } ] }, { "type": "text", "bbox": [ 106, 225, 505, 314 ], "lines": [ { "bbox": [ 105, 225, 505, 237 ], "spans": [ { "bbox": [ 105, 225, 259, 237 ], "score": 1.0, "content": "The first term (Eqn. 4a) seeks a vector", "type": "text" }, { "bbox": [ 260, 227, 266, 235 ], "score": 0.76, "content": "z", "type": "inline_equation" }, { "bbox": [ 267, 225, 505, 237 ], "score": 1.0, "content": "that, when substituted as the output of the MLP at the token", "type": "text" } ], "index": 10 }, { "bbox": [ 107, 235, 506, 250 ], "spans": [ { "bbox": [ 107, 238, 111, 246 ], "score": 0.77, "content": "i", "type": "inline_equation" }, { "bbox": [ 111, 235, 243, 250 ], "score": 1.0, "content": "at the end of the subject (notated", "type": "text" }, { "bbox": [ 243, 235, 303, 249 ], "score": 0.89, "content": "G ( m _ { i } ^ { ( l ^ { * } ) } : = z ) ^ { \\backslash }", "type": "inline_equation" }, { "bbox": [ 303, 235, 506, 250 ], "score": 1.0, "content": "), will cause the network to predict the target object", "type": "text" } ], "index": 11 }, { "bbox": [ 107, 246, 506, 260 ], "spans": [ { "bbox": [ 107, 248, 117, 258 ], "score": 0.84, "content": "o ^ { * }", "type": "inline_equation" }, { "bbox": [ 117, 246, 252, 260 ], "score": 1.0, "content": "in response to the factual prompt", "type": "text" }, { "bbox": [ 253, 250, 259, 259 ], "score": 0.79, "content": "p", "type": "inline_equation" }, { "bbox": [ 259, 246, 506, 260 ], "score": 1.0, "content": ". The second term (Eqn. 4b) minimizes the KL divergence of", "type": "text" } ], "index": 12 }, { "bbox": [ 104, 258, 506, 271 ], "spans": [ { "bbox": [ 104, 258, 216, 271 ], "score": 1.0, "content": "predictions for the prompt", "type": "text" }, { "bbox": [ 216, 258, 226, 270 ], "score": 0.86, "content": "p ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 226, 258, 280, 271 ], "score": 1.0, "content": "(of the form", "type": "text" }, { "bbox": [ 280, 258, 324, 270 ], "score": 0.52, "content": "\\mathbf { \\cdots } \\{ \\mathrm { s u b j e c t } \\}", "type": "inline_equation" }, { "bbox": [ 324, 258, 506, 271 ], "score": 1.0, "content": "is a”) to the unchanged model, which helps", "type": "text" } ], "index": 13 }, { "bbox": [ 104, 269, 506, 282 ], "spans": [ { "bbox": [ 104, 269, 506, 282 ], "score": 1.0, "content": "preserve the model’s understanding of the subject’s essence. To be clear, the optimization does not", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 280, 506, 293 ], "spans": [ { "bbox": [ 105, 280, 362, 293 ], "score": 1.0, "content": "directly alter model weights; it identifies a vector representation", "type": "text" }, { "bbox": [ 363, 282, 373, 291 ], "score": 0.86, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 374, 280, 506, 293 ], "score": 1.0, "content": "that, when output at the targeted", "type": "text" } ], "index": 15 }, { "bbox": [ 104, 290, 507, 304 ], "spans": [ { "bbox": [ 104, 290, 276, 304 ], "score": 1.0, "content": "MLP module, represents the new property", "type": "text" }, { "bbox": [ 276, 291, 302, 303 ], "score": 0.92, "content": "( r , o ^ { * } )", "type": "inline_equation" }, { "bbox": [ 302, 290, 362, 304 ], "score": 1.0, "content": "for the subject", "type": "text" }, { "bbox": [ 362, 293, 368, 301 ], "score": 0.35, "content": "s", "type": "inline_equation" }, { "bbox": [ 369, 290, 453, 304 ], "score": 1.0, "content": ". Note that, similar to", "type": "text" }, { "bbox": [ 454, 291, 465, 302 ], "score": 0.89, "content": "k _ { * }", "type": "inline_equation" }, { "bbox": [ 465, 290, 507, 304 ], "score": 1.0, "content": "selection,", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 302, 507, 315 ], "spans": [ { "bbox": [ 106, 304, 117, 313 ], "score": 0.83, "content": "v _ { * }", "type": "inline_equation" }, { "bbox": [ 118, 302, 300, 315 ], "score": 1.0, "content": "optimization also uses the random prefix texts", "type": "text" }, { "bbox": [ 300, 304, 311, 314 ], "score": 0.87, "content": "x _ { j }", "type": "inline_equation" }, { "bbox": [ 311, 302, 507, 315 ], "score": 1.0, "content": "to encourage robustness under differing contexts.", "type": "text" } ], "index": 17 } ], "index": 13.5, "bbox_fs": [ 104, 225, 507, 315 ] }, { "type": "text", "bbox": [ 106, 318, 504, 352 ], "lines": [ { "bbox": [ 105, 316, 506, 333 ], "spans": [ { "bbox": [ 105, 316, 368, 333 ], "score": 1.0, "content": "Step 3: Inserting the Fact. Once we have computed the pair", "type": "text" }, { "bbox": [ 369, 319, 401, 329 ], "score": 0.86, "content": "( k _ { * } , v _ { * } )", "type": "inline_equation" }, { "bbox": [ 401, 316, 506, 333 ], "score": 1.0, "content": "to represent the full fact", "type": "text" } ], "index": 18 }, { "bbox": [ 103, 326, 506, 344 ], "spans": [ { "bbox": [ 103, 326, 350, 344 ], "score": 1.0, "content": "(s, r, o⇤), we apply Eqn. 2, updating the MLP weights W (l)proj", "type": "text" }, { "bbox": [ 350, 329, 506, 343 ], "score": 1.0, "content": "with a rank-one update that inserts the", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 340, 454, 353 ], "spans": [ { "bbox": [ 106, 340, 454, 353 ], "score": 1.0, "content": "new key–value association directly. For full implementation details, see Appendix E.5.", "type": "text" } ], "index": 20 } ], "index": 19, "bbox_fs": [ 103, 316, 506, 353 ] }, { "type": "title", "bbox": [ 106, 364, 376, 376 ], "lines": [ { "bbox": [ 105, 364, 377, 379 ], "spans": [ { "bbox": [ 105, 364, 377, 379 ], "score": 1.0, "content": "3.2 Evaluating ROME: Zero-Shot Relation Extraction (zsRE)", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 107, 384, 504, 407 ], "lines": [ { "bbox": [ 105, 384, 505, 397 ], "spans": [ { "bbox": [ 105, 384, 505, 397 ], "score": 1.0, "content": "We wish to test our localized factual association hypothesis: can storing a single new vector association", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 396, 430, 407 ], "spans": [ { "bbox": [ 106, 396, 430, 407 ], "score": 1.0, "content": "using ROME insert a substantial, generalized factual association into the model?", "type": "text" } ], "index": 23 } ], "index": 22.5, "bbox_fs": [ 105, 384, 505, 407 ] }, { "type": "text", "bbox": [ 106, 412, 505, 489 ], "lines": [ { "bbox": [ 105, 412, 506, 425 ], "spans": [ { "bbox": [ 105, 412, 506, 425 ], "score": 1.0, "content": "A natural question is how ROME compares to other model-editing methods, which use direct", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 424, 506, 436 ], "spans": [ { "bbox": [ 106, 424, 506, 436 ], "score": 1.0, "content": "optimization or hypernetworks to incorporate a single new training example into a network. For", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 434, 505, 447 ], "spans": [ { "bbox": [ 106, 434, 505, 447 ], "score": 1.0, "content": "baselines, we examine Fine-Tuning (FT), which applies Adam with early stopping at one layer to", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 444, 506, 458 ], "spans": [ { "bbox": [ 105, 444, 146, 458 ], "score": 1.0, "content": "minimize", "type": "text" }, { "bbox": [ 146, 445, 207, 457 ], "score": 0.9, "content": "- \\log \\mathbb { P } \\left[ o ^ { * } \\mid x \\right]", "type": "inline_equation" }, { "bbox": [ 208, 444, 312, 458 ], "score": 1.0, "content": ". Constrained Fine-Tuning", "type": "text" }, { "bbox": [ 312, 445, 343, 456 ], "score": 0.71, "content": "\\mathbf { \\left( F T + L \\right) }", "type": "inline_equation" }, { "bbox": [ 343, 444, 506, 458 ], "score": 1.0, "content": "(Zhu et al., 2020) additionally imposes a", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 456, 506, 469 ], "spans": [ { "bbox": [ 105, 456, 172, 469 ], "score": 1.0, "content": "parameter-space", "type": "text" }, { "bbox": [ 172, 457, 189, 467 ], "score": 0.88, "content": "L _ { \\infty }", "type": "inline_equation" }, { "bbox": [ 189, 456, 506, 469 ], "score": 1.0, "content": "norm constraint on weight changes. We also test two hypernetworks: Knowledge", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 466, 505, 480 ], "spans": [ { "bbox": [ 106, 466, 135, 480 ], "score": 1.0, "content": "Editor", "type": "text" }, { "bbox": [ 135, 467, 156, 478 ], "score": 0.27, "content": "\\mathbf { ( K E ) }", "type": "inline_equation" }, { "bbox": [ 157, 466, 505, 480 ], "score": 1.0, "content": "(De Cao et al., 2021) and MEND (Mitchell et al., 2021), both of which learn auxiliary", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 478, 445, 490 ], "spans": [ { "bbox": [ 106, 478, 252, 490 ], "score": 1.0, "content": "models to predict weight changes to", "type": "text" }, { "bbox": [ 253, 478, 262, 488 ], "score": 0.7, "content": "G", "type": "inline_equation" }, { "bbox": [ 262, 478, 445, 490 ], "score": 1.0, "content": ". Further details are described in Appendix E.", "type": "text" } ], "index": 30 } ], "index": 27, "bbox_fs": [ 105, 412, 506, 490 ] }, { "type": "text", "bbox": [ 107, 494, 297, 635 ], "lines": [ { "bbox": [ 107, 495, 298, 505 ], "spans": [ { "bbox": [ 107, 495, 298, 505 ], "score": 1.0, "content": "We first evaluate ROME on the Zero-Shot Re-", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 505, 297, 516 ], "spans": [ { "bbox": [ 106, 505, 297, 516 ], "score": 1.0, "content": "lation Extraction (zsRE) task used in Mitchell", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 516, 298, 527 ], "spans": [ { "bbox": [ 105, 516, 298, 527 ], "score": 1.0, "content": "et al. (2021) and De Cao et al. (2021). Our", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 527, 297, 538 ], "spans": [ { "bbox": [ 106, 527, 297, 538 ], "score": 1.0, "content": "evaluation slice contains 10,000 records, each", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 538, 298, 550 ], "spans": [ { "bbox": [ 106, 538, 298, 550 ], "score": 1.0, "content": "containing one factual statement, its paraphrase,", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 549, 298, 561 ], "spans": [ { "bbox": [ 105, 549, 298, 561 ], "score": 1.0, "content": "and one unrelated factual statement. “Efficacy”", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 559, 297, 572 ], "spans": [ { "bbox": [ 105, 559, 297, 572 ], "score": 1.0, "content": "and “Paraphrase” measure post-edit accuracy", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 569, 298, 583 ], "spans": [ { "bbox": [ 106, 570, 208, 583 ], "score": 0.91, "content": "\\mathbb { I } \\big [ o ^ { * } = \\mathrm { a r g m a x } _ { o } \\mathbb { P } _ { G ^ { \\prime } } \\left[ o \\right] \\big ]", "type": "inline_equation" }, { "bbox": [ 208, 569, 298, 583 ], "score": 1.0, "content": "of the statement and", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 581, 298, 594 ], "spans": [ { "bbox": [ 105, 581, 298, 594 ], "score": 1.0, "content": "its paraphrase, respectively, while “Specificity”", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 592, 298, 604 ], "spans": [ { "bbox": [ 105, 592, 298, 604 ], "score": 1.0, "content": "measures the edited model’s accuracy on an un-", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 603, 297, 615 ], "spans": [ { "bbox": [ 106, 603, 297, 615 ], "score": 1.0, "content": "related fact. Table 1 shows the results: ROME is", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 614, 297, 627 ], "spans": [ { "bbox": [ 105, 614, 297, 627 ], "score": 1.0, "content": "competitive with hypernetworks and fine-tuning", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 625, 297, 637 ], "spans": [ { "bbox": [ 106, 625, 297, 637 ], "score": 1.0, "content": "methods despite its simplicity. We find that it", "type": "text" } ], "index": 43 } ], "index": 37, "bbox_fs": [ 105, 495, 298, 637 ] }, { "type": "table", "bbox": [ 305, 507, 504, 623 ], "blocks": [ { "type": "table_caption", "bbox": [ 321, 495, 486, 506 ], "group_id": 0, "lines": [ { "bbox": [ 320, 494, 488, 508 ], "spans": [ { "bbox": [ 320, 494, 488, 508 ], "score": 1.0, "content": "Table 1: zsRE Editing Results on GPT-2 XL.", "type": "text" } ], "index": 44 } ], "index": 44 }, { "type": "table_body", "bbox": [ 305, 507, 504, 623 ], "group_id": 0, "lines": [ { "bbox": [ 305, 507, 504, 623 ], "spans": [ { "bbox": [ 305, 507, 504, 623 ], "score": 0.98, "html": "
EditorEfficacy 个 Paraphrase 个 Specificity 个
GPT-2 XL22.2 (±0.5) 21.3 (±0.5) 24.2 (±0.5)
FT99.6 (±0.1) 82.1 (±0.6) 23.2(±0.5)
FT+L92.3 (±0.4) 47.2 (±0.7) 23.4(±0.5)
KE65.5 (±0.6) 61.4(±0.6) 24.9 (±0.5)
KE-zsRE92.4 (±0.3) 90.0 (±0.3) 23.8 (±0.5)
MEND75.9 (±0.5) 65.3 (±0.6) 24.1(±0.5)
MEND-zsRE 99.4 (±0.1)99.3 (±0.1) 24.1(±0.5)
ROME99.8 (±0.0) 88.1(±0.5) 24.2 (±0.5)
", "type": "table", "image_path": "0e85568b3db85a4ec77f72778120f03fa7931c55642ffa5391e3f2bdfa77943d.jpg" } ] } ], "index": 48.5, "virtual_lines": [ { "bbox": [ 305, 507, 504, 521.5 ], "spans": [], "index": 45 }, { "bbox": [ 305, 521.5, 504, 536.0 ], "spans": [], "index": 46 }, { "bbox": [ 305, 536.0, 504, 550.5 ], "spans": [], "index": 47 }, { "bbox": [ 305, 550.5, 504, 565.0 ], "spans": [], "index": 48 }, { "bbox": [ 305, 565.0, 504, 579.5 ], "spans": [], "index": 49 }, { "bbox": [ 305, 579.5, 504, 594.0 ], "spans": [], "index": 50 }, { "bbox": [ 305, 594.0, 504, 608.5 ], "spans": [], "index": 51 }, { "bbox": [ 305, 608.5, 504, 623.0 ], "spans": [], "index": 52 } ] } ], "index": 46.25 }, { "type": "text", "bbox": [ 107, 636, 505, 702 ], "lines": [ { "bbox": [ 105, 635, 505, 649 ], "spans": [ { "bbox": [ 105, 635, 505, 649 ], "score": 1.0, "content": "is not hard for ROME to insert an association that can be regurgitated by the model. Robustness", "type": "text" } ], "index": 53 }, { "bbox": [ 106, 648, 505, 659 ], "spans": [ { "bbox": [ 106, 648, 505, 659 ], "score": 1.0, "content": "under paraphrase is also strong, although it comes short of custom-tuned hyperparameter networks", "type": "text" } ], "index": 54 }, { "bbox": [ 105, 657, 506, 669 ], "spans": [ { "bbox": [ 105, 657, 506, 669 ], "score": 1.0, "content": "KE-zsRE and MEND-zsRE, which we explicitly trained on the zsRE data distribution.7 We find that", "type": "text" } ], "index": 55 }, { "bbox": [ 105, 668, 505, 681 ], "spans": [ { "bbox": [ 105, 668, 505, 681 ], "score": 1.0, "content": "zsRE’s specificity score is not a sensitive measure of model damage, since these prompts are sampled", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 678, 506, 694 ], "spans": [ { "bbox": [ 105, 678, 506, 694 ], "score": 1.0, "content": "from a large space of possible facts, whereas bleedover is most likely to occur on related neighboring", "type": "text" } ], "index": 57 }, { "bbox": [ 106, 691, 338, 702 ], "spans": [ { "bbox": [ 106, 691, 338, 702 ], "score": 1.0, "content": "subjects. Appendix C has additional experimental details.", "type": "text" } ], "index": 58 } ], "index": 55.5, "bbox_fs": [ 105, 635, 506, 702 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 106, 68, 504, 165 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 68, 504, 165 ], "group_id": 0, "lines": [ { "bbox": [ 106, 68, 504, 165 ], "spans": [ { "bbox": [ 106, 68, 504, 165 ], "score": 0.957, "type": "image", "image_path": "5a2a7a1c4f15eec825b5bc2c6f590368fcee6879815bfc273dd860b9cd34e2db.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 68, 504, 100.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 106, 100.33333333333334, 504, 132.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 106, 132.66666666666669, 504, 165.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 168, 506, 199 ], "group_id": 0, "lines": [ { "bbox": [ 106, 167, 506, 180 ], "spans": [ { "bbox": [ 106, 167, 506, 180 ], "score": 1.0, "content": "Figure 5: ROME edits are benchmarked at each layer-and-token combination in GPT-2-XL. The target token is", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 177, 506, 191 ], "spans": [ { "bbox": [ 105, 177, 255, 191 ], "score": 1.0, "content": "determined by selecting the token index", "type": "text" }, { "bbox": [ 255, 179, 260, 187 ], "score": 0.46, "content": "_ { i }", "type": "inline_equation" }, { "bbox": [ 260, 177, 506, 191 ], "score": 1.0, "content": "where the key representation is collected (Eqn. 3). ROME editing", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 189, 497, 199 ], "spans": [ { "bbox": [ 106, 189, 497, 199 ], "score": 1.0, "content": "results confirm the importance of mid-layer MLP layers at the final subject token, where performance peaks.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "title", "bbox": [ 106, 212, 340, 224 ], "lines": [ { "bbox": [ 105, 211, 340, 225 ], "spans": [ { "bbox": [ 105, 211, 340, 225 ], "score": 1.0, "content": "3.3 Evaluating ROME: Our COUNTERFACT Dataset", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 107, 233, 505, 266 ], "lines": [ { "bbox": [ 105, 232, 506, 246 ], "spans": [ { "bbox": [ 105, 232, 506, 246 ], "score": 1.0, "content": "While standard model-editing metrics on zsRE are a reasonable starting point for evaluating ROME,", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 243, 505, 257 ], "spans": [ { "bbox": [ 105, 243, 505, 257 ], "score": 1.0, "content": "they do not provide detailed insights that would allow us to distinguish superficial wording changes", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 254, 425, 268 ], "spans": [ { "bbox": [ 105, 254, 425, 268 ], "score": 1.0, "content": "from deeper modifications that correspond to a meaningful change about a fact.", "type": "text" } ], "index": 9 } ], "index": 8 }, { "type": "text", "bbox": [ 107, 271, 505, 392 ], "lines": [ { "bbox": [ 105, 271, 505, 284 ], "spans": [ { "bbox": [ 105, 271, 505, 284 ], "score": 1.0, "content": "In particular, we wish to measure the efficacy of significant changes. Hase et al. (2021) observed", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 281, 506, 294 ], "spans": [ { "bbox": [ 105, 281, 506, 294 ], "score": 1.0, "content": "that standard model-editing benchmarks underestimate difficulty by often testing only proposals that", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 293, 505, 305 ], "spans": [ { "bbox": [ 106, 293, 442, 305 ], "score": 1.0, "content": "the model previously scored as likely. We compile a set of more difficult false facts", "type": "text" }, { "bbox": [ 442, 293, 477, 304 ], "score": 0.94, "content": "( s , r , o ^ { * } )", "type": "inline_equation" }, { "bbox": [ 477, 293, 505, 305 ], "score": 1.0, "content": ": these", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 304, 506, 316 ], "spans": [ { "bbox": [ 106, 304, 368, 316 ], "score": 1.0, "content": "counterfactuals start with low scores compared to the correct facts", "type": "text" }, { "bbox": [ 368, 304, 403, 316 ], "score": 0.92, "content": "( s , r , o ^ { c } )", "type": "inline_equation" }, { "bbox": [ 403, 304, 506, 316 ], "score": 1.0, "content": ". Our Efficacy Score (ES)", "type": "text" } ], "index": 13 }, { "bbox": [ 104, 313, 506, 328 ], "spans": [ { "bbox": [ 104, 313, 273, 328 ], "score": 1.0, "content": "is the portion of cases for which we have", "type": "text" }, { "bbox": [ 274, 315, 330, 327 ], "score": 0.94, "content": "\\mathbb { P } [ o ^ { * } ] > \\mathbb { P } [ o ^ { c } ]", "type": "inline_equation" }, { "bbox": [ 330, 313, 506, 328 ], "score": 1.0, "content": "post-edit, and Efficacy Magnitude (EM) is", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 324, 506, 339 ], "spans": [ { "bbox": [ 105, 324, 189, 339 ], "score": 1.0, "content": "the mean difference", "type": "text" }, { "bbox": [ 190, 325, 244, 338 ], "score": 0.94, "content": "\\mathbb { P } [ o ^ { * } ] - \\mathbb { P } [ o ^ { c } ]", "type": "inline_equation" }, { "bbox": [ 245, 324, 506, 339 ], "score": 1.0, "content": ". Then, to measure generalization, with each counterfactual we", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 336, 506, 349 ], "spans": [ { "bbox": [ 106, 337, 298, 349 ], "score": 1.0, "content": "gather a set of rephrased prompts equivalent to", "type": "text" }, { "bbox": [ 298, 336, 320, 349 ], "score": 0.92, "content": "( s , r )", "type": "inline_equation" }, { "bbox": [ 320, 337, 506, 349 ], "score": 1.0, "content": "and report Paraphrase Scores (PS) and (PM),", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 347, 506, 361 ], "spans": [ { "bbox": [ 105, 347, 478, 361 ], "score": 1.0, "content": "computed similarly to ES and EM. To measure specificity, we collect a set of nearby subjects", "type": "text" }, { "bbox": [ 479, 349, 489, 358 ], "score": 0.85, "content": "s _ { n }", "type": "inline_equation" }, { "bbox": [ 490, 347, 506, 361 ], "score": 1.0, "content": "for", "type": "text" } ], "index": 17 }, { "bbox": [ 104, 356, 502, 372 ], "spans": [ { "bbox": [ 104, 356, 134, 372 ], "score": 1.0, "content": "which", "type": "text" }, { "bbox": [ 134, 358, 174, 370 ], "score": 0.93, "content": "( s _ { n } , r , o ^ { c } )", "type": "inline_equation" }, { "bbox": [ 175, 356, 446, 372 ], "score": 1.0, "content": "holds true. Because we do not wish to alter these subjects, we test", "type": "text" }, { "bbox": [ 447, 358, 502, 370 ], "score": 0.93, "content": "\\mathbb { P } [ o ^ { c } ] > \\mathbb { P } [ o ^ { * } ]", "type": "inline_equation" } ], "index": 18 }, { "bbox": [ 105, 370, 506, 381 ], "spans": [ { "bbox": [ 105, 370, 506, 381 ], "score": 1.0, "content": "reporting the success fraction as Neighborhood Score (NS) and difference as (NM). To test the", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 381, 479, 393 ], "spans": [ { "bbox": [ 105, 381, 479, 393 ], "score": 1.0, "content": "generalization–specificity tradeoff, we report the harmonic mean of ES, PS, NS as Score (S).", "type": "text" } ], "index": 20 } ], "index": 15 }, { "type": "text", "bbox": [ 107, 397, 317, 517 ], "lines": [ { "bbox": [ 106, 396, 317, 408 ], "spans": [ { "bbox": [ 106, 396, 317, 408 ], "score": 1.0, "content": "We also wish to measure semantic consistency of", "type": "text" } ], "index": 21 }, { "bbox": [ 107, 407, 318, 420 ], "spans": [ { "bbox": [ 107, 408, 118, 418 ], "score": 0.83, "content": "G ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 118, 407, 318, 420 ], "score": 1.0, "content": "’s generations. To do so, we generate text start-", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 418, 318, 430 ], "spans": [ { "bbox": [ 105, 418, 143, 430 ], "score": 1.0, "content": "ing with", "type": "text" }, { "bbox": [ 143, 420, 149, 429 ], "score": 0.67, "content": "s", "type": "inline_equation" }, { "bbox": [ 150, 418, 318, 430 ], "score": 1.0, "content": "and report (RS) as the cos similarity be-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 429, 318, 441 ], "spans": [ { "bbox": [ 105, 429, 318, 441 ], "score": 1.0, "content": "tween the unigram TF-IDF vectors of generated texts,", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 440, 317, 453 ], "spans": [ { "bbox": [ 105, 440, 317, 453 ], "score": 1.0, "content": "compared to reference texts about subjects sharing", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 451, 317, 464 ], "spans": [ { "bbox": [ 105, 451, 184, 464 ], "score": 1.0, "content": "the target property", "type": "text" }, { "bbox": [ 185, 452, 195, 461 ], "score": 0.84, "content": "o ^ { * }", "type": "inline_equation" }, { "bbox": [ 195, 451, 317, 464 ], "score": 1.0, "content": ". Finally, we monitor fluency", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 462, 317, 474 ], "spans": [ { "bbox": [ 105, 462, 317, 474 ], "score": 1.0, "content": "degradations by measuring the weighted average of", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 471, 317, 485 ], "spans": [ { "bbox": [ 105, 471, 317, 485 ], "score": 1.0, "content": "bi- and tri-gram entropies (Zhang et al., 2018) given", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 483, 317, 497 ], "spans": [ { "bbox": [ 105, 483, 119, 497 ], "score": 1.0, "content": "by", "type": "text" }, { "bbox": [ 119, 484, 207, 496 ], "score": 0.81, "content": "\\begin{array} { r } { - \\sum _ { k } f ( k ) \\log _ { 2 } f ( k ) } \\end{array}", "type": "inline_equation" }, { "bbox": [ 207, 483, 239, 497 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 239, 484, 257, 496 ], "score": 0.91, "content": "f ( \\cdot )", "type": "inline_equation" }, { "bbox": [ 257, 483, 284, 497 ], "score": 1.0, "content": "is the", "type": "text" }, { "bbox": [ 284, 486, 291, 494 ], "score": 0.78, "content": "n", "type": "inline_equation" }, { "bbox": [ 291, 483, 317, 497 ], "score": 1.0, "content": "-gram", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 495, 317, 507 ], "spans": [ { "bbox": [ 106, 495, 317, 507 ], "score": 1.0, "content": "frequency distribution, which we report as (GE); this", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 505, 298, 518 ], "spans": [ { "bbox": [ 105, 505, 298, 518 ], "score": 1.0, "content": "quantity drops if text generations are repetitive.", "type": "text" } ], "index": 31 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 522, 317, 587 ], "lines": [ { "bbox": [ 106, 522, 317, 534 ], "spans": [ { "bbox": [ 106, 522, 317, 534 ], "score": 1.0, "content": "In order to facilitate the above measurements, we", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 532, 317, 545 ], "spans": [ { "bbox": [ 106, 532, 317, 545 ], "score": 1.0, "content": "introduce COUNTERFACT, a challenging evaluation", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 542, 317, 556 ], "spans": [ { "bbox": [ 105, 542, 317, 556 ], "score": 1.0, "content": "dataset for evaluating counterfactual edits in language", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 555, 317, 567 ], "spans": [ { "bbox": [ 106, 555, 317, 567 ], "score": 1.0, "content": "models. Containing 21,919 records with a diverse", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 565, 318, 578 ], "spans": [ { "bbox": [ 106, 565, 318, 578 ], "score": 1.0, "content": "set of subjects, relations, and linguistic variations,", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 576, 318, 588 ], "spans": [ { "bbox": [ 106, 576, 318, 588 ], "score": 1.0, "content": "COUNTERFACT’s goal is to differentiate robust stor-", "type": "text" } ], "index": 37 } ], "index": 34.5 }, { "type": "table", "bbox": [ 328, 410, 504, 508 ], "blocks": [ { "type": "table_caption", "bbox": [ 342, 399, 487, 409 ], "group_id": 0, "lines": [ { "bbox": [ 341, 397, 487, 411 ], "spans": [ { "bbox": [ 341, 397, 487, 411 ], "score": 1.0, "content": "Table 2: COUNTERFACT Composition", "type": "text" } ], "index": 38 } ], "index": 38 }, { "type": "table_body", "bbox": [ 328, 410, 504, 508 ], "group_id": 0, "lines": [ { "bbox": [ 328, 410, 504, 508 ], "spans": [ { "bbox": [ 328, 410, 504, 508 ], "score": 0.975, "html": "
Per Per Item Total Relation Record
Records 21919 645 1
Subjects 20391 624 1
Objects 749 60 1
Counterfactual Statements 21595 635 1
Paraphrase Prompts 42876 1262 2
Neighborhood Prompts 82650 2441 10
Generation Prompts 62346 1841 3
", "type": "table", "image_path": "56dd6344eec28b2447ac46d99df4b8d20fb1b574bc947630b3237d3833c61664.jpg" } ] } ], "index": 42, "virtual_lines": [ { "bbox": [ 328, 410, 504, 424.0 ], "spans": [], "index": 39 }, { "bbox": [ 328, 424.0, 504, 438.0 ], "spans": [], "index": 40 }, { "bbox": [ 328, 438.0, 504, 452.0 ], "spans": [], "index": 41 }, { "bbox": [ 328, 452.0, 504, 466.0 ], "spans": [], "index": 42 }, { "bbox": [ 328, 466.0, 504, 480.0 ], "spans": [], "index": 43 }, { "bbox": [ 328, 480.0, 504, 494.0 ], "spans": [], "index": 44 }, { "bbox": [ 328, 494.0, 504, 508.0 ], "spans": [], "index": 45 } ] } ], "index": 40.0 }, { "type": "table", "bbox": [ 325, 527, 502, 582 ], "blocks": [ { "type": "table_caption", "bbox": [ 330, 514, 498, 524 ], "group_id": 1, "lines": [ { "bbox": [ 329, 513, 500, 526 ], "spans": [ { "bbox": [ 329, 513, 500, 526 ], "score": 1.0, "content": "Table 3: Comparison to Existing Benchmarks", "type": "text" } ], "index": 46 } ], "index": 46 }, { "type": "table_body", "bbox": [ 325, 527, 502, 582 ], "group_id": 1, "lines": [ { "bbox": [ 325, 527, 502, 582 ], "spans": [ { "bbox": [ 325, 527, 502, 582 ], "score": 0.968, "html": "
CriterionSQuAD zSRE FEVER WikiTextPARAREL CF
Efficacy<<xx<<xxxvvv<>
Generalization
Bleedover
Consistency<<xxx/xxxx
FluencyX<<xxx
", "type": "table", "image_path": "0cc2be581f070ea69322025c4a04be9eb9c48b44947f2353c02b9f51712ad1cc.jpg" } ] } ], "index": 48.5, "virtual_lines": [ { "bbox": [ 325, 527, 502, 540.75 ], "spans": [], "index": 47 }, { "bbox": [ 325, 540.75, 502, 554.5 ], "spans": [], "index": 48 }, { "bbox": [ 325, 554.5, 502, 568.25 ], "spans": [], "index": 49 }, { "bbox": [ 325, 568.25, 502, 582.0 ], "spans": [], "index": 50 } ] } ], "index": 47.25 }, { "type": "text", "bbox": [ 107, 588, 505, 610 ], "lines": [ { "bbox": [ 105, 587, 505, 600 ], "spans": [ { "bbox": [ 105, 587, 505, 600 ], "score": 1.0, "content": "age of new facts from the superficial regurgitation of target words. See Appendix D for additional", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 597, 450, 611 ], "spans": [ { "bbox": [ 105, 597, 450, 611 ], "score": 1.0, "content": "technical details about its construction, and Table 2 for a summary of its composition.", "type": "text" } ], "index": 52 } ], "index": 51.5 }, { "type": "title", "bbox": [ 107, 624, 448, 636 ], "lines": [ { "bbox": [ 104, 622, 450, 640 ], "spans": [ { "bbox": [ 104, 622, 450, 640 ], "score": 1.0, "content": "3.4 Confirming the Importance of Decisive States Identified by Causal Tracing", "type": "text" } ], "index": 53 } ], "index": 53 }, { "type": "text", "bbox": [ 107, 645, 506, 722 ], "lines": [ { "bbox": [ 105, 644, 507, 658 ], "spans": [ { "bbox": [ 105, 644, 507, 658 ], "score": 1.0, "content": "In Section 2, we used Causal Tracing to identify decisive hidden states. To confirm that factual asso-", "type": "text" } ], "index": 54 }, { "bbox": [ 106, 656, 506, 668 ], "spans": [ { "bbox": [ 106, 656, 506, 668 ], "score": 1.0, "content": "ciations are indeed stored in the MLP modules that output those states, we test ROME’s effectiveness", "type": "text" } ], "index": 55 }, { "bbox": [ 105, 667, 505, 680 ], "spans": [ { "bbox": [ 105, 667, 505, 680 ], "score": 1.0, "content": "when targeted at various layers and tokens. Figure 5 plots four metrics evaluating both generalization", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 678, 506, 690 ], "spans": [ { "bbox": [ 105, 678, 506, 690 ], "score": 1.0, "content": "(a,b,d) and specificity (c). We observe strong correlations with the causal analysis; rewrites are most", "type": "text" } ], "index": 57 }, { "bbox": [ 106, 689, 506, 701 ], "spans": [ { "bbox": [ 106, 689, 506, 701 ], "score": 1.0, "content": "successful at the last subject token, where both specificity and generalization peak at middle layers.", "type": "text" } ], "index": 58 }, { "bbox": [ 106, 700, 505, 712 ], "spans": [ { "bbox": [ 106, 700, 505, 712 ], "score": 1.0, "content": "Targeting earlier or later tokens results in poor generalization and/or specificity. Furthermore, the", "type": "text" } ], "index": 59 }, { "bbox": [ 105, 710, 505, 725 ], "spans": [ { "bbox": [ 105, 710, 505, 725 ], "score": 1.0, "content": "layers at which edits generalize best correspond to the middle layers of the early site identified by", "type": "text" } ], "index": 60 } ], "index": 57 } ], "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": "image", "bbox": [ 106, 68, 504, 165 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 68, 504, 165 ], "group_id": 0, "lines": [ { "bbox": [ 106, 68, 504, 165 ], "spans": [ { "bbox": [ 106, 68, 504, 165 ], "score": 0.957, "type": "image", "image_path": "5a2a7a1c4f15eec825b5bc2c6f590368fcee6879815bfc273dd860b9cd34e2db.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 68, 504, 100.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 106, 100.33333333333334, 504, 132.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 106, 132.66666666666669, 504, 165.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 168, 506, 199 ], "group_id": 0, "lines": [ { "bbox": [ 106, 167, 506, 180 ], "spans": [ { "bbox": [ 106, 167, 506, 180 ], "score": 1.0, "content": "Figure 5: ROME edits are benchmarked at each layer-and-token combination in GPT-2-XL. The target token is", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 177, 506, 191 ], "spans": [ { "bbox": [ 105, 177, 255, 191 ], "score": 1.0, "content": "determined by selecting the token index", "type": "text" }, { "bbox": [ 255, 179, 260, 187 ], "score": 0.46, "content": "_ { i }", "type": "inline_equation" }, { "bbox": [ 260, 177, 506, 191 ], "score": 1.0, "content": "where the key representation is collected (Eqn. 3). ROME editing", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 189, 497, 199 ], "spans": [ { "bbox": [ 106, 189, 497, 199 ], "score": 1.0, "content": "results confirm the importance of mid-layer MLP layers at the final subject token, where performance peaks.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "title", "bbox": [ 106, 212, 340, 224 ], "lines": [ { "bbox": [ 105, 211, 340, 225 ], "spans": [ { "bbox": [ 105, 211, 340, 225 ], "score": 1.0, "content": "3.3 Evaluating ROME: Our COUNTERFACT Dataset", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 107, 233, 505, 266 ], "lines": [ { "bbox": [ 105, 232, 506, 246 ], "spans": [ { "bbox": [ 105, 232, 506, 246 ], "score": 1.0, "content": "While standard model-editing metrics on zsRE are a reasonable starting point for evaluating ROME,", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 243, 505, 257 ], "spans": [ { "bbox": [ 105, 243, 505, 257 ], "score": 1.0, "content": "they do not provide detailed insights that would allow us to distinguish superficial wording changes", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 254, 425, 268 ], "spans": [ { "bbox": [ 105, 254, 425, 268 ], "score": 1.0, "content": "from deeper modifications that correspond to a meaningful change about a fact.", "type": "text" } ], "index": 9 } ], "index": 8, "bbox_fs": [ 105, 232, 506, 268 ] }, { "type": "text", "bbox": [ 107, 271, 505, 392 ], "lines": [ { "bbox": [ 105, 271, 505, 284 ], "spans": [ { "bbox": [ 105, 271, 505, 284 ], "score": 1.0, "content": "In particular, we wish to measure the efficacy of significant changes. Hase et al. (2021) observed", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 281, 506, 294 ], "spans": [ { "bbox": [ 105, 281, 506, 294 ], "score": 1.0, "content": "that standard model-editing benchmarks underestimate difficulty by often testing only proposals that", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 293, 505, 305 ], "spans": [ { "bbox": [ 106, 293, 442, 305 ], "score": 1.0, "content": "the model previously scored as likely. We compile a set of more difficult false facts", "type": "text" }, { "bbox": [ 442, 293, 477, 304 ], "score": 0.94, "content": "( s , r , o ^ { * } )", "type": "inline_equation" }, { "bbox": [ 477, 293, 505, 305 ], "score": 1.0, "content": ": these", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 304, 506, 316 ], "spans": [ { "bbox": [ 106, 304, 368, 316 ], "score": 1.0, "content": "counterfactuals start with low scores compared to the correct facts", "type": "text" }, { "bbox": [ 368, 304, 403, 316 ], "score": 0.92, "content": "( s , r , o ^ { c } )", "type": "inline_equation" }, { "bbox": [ 403, 304, 506, 316 ], "score": 1.0, "content": ". Our Efficacy Score (ES)", "type": "text" } ], "index": 13 }, { "bbox": [ 104, 313, 506, 328 ], "spans": [ { "bbox": [ 104, 313, 273, 328 ], "score": 1.0, "content": "is the portion of cases for which we have", "type": "text" }, { "bbox": [ 274, 315, 330, 327 ], "score": 0.94, "content": "\\mathbb { P } [ o ^ { * } ] > \\mathbb { P } [ o ^ { c } ]", "type": "inline_equation" }, { "bbox": [ 330, 313, 506, 328 ], "score": 1.0, "content": "post-edit, and Efficacy Magnitude (EM) is", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 324, 506, 339 ], "spans": [ { "bbox": [ 105, 324, 189, 339 ], "score": 1.0, "content": "the mean difference", "type": "text" }, { "bbox": [ 190, 325, 244, 338 ], "score": 0.94, "content": "\\mathbb { P } [ o ^ { * } ] - \\mathbb { P } [ o ^ { c } ]", "type": "inline_equation" }, { "bbox": [ 245, 324, 506, 339 ], "score": 1.0, "content": ". Then, to measure generalization, with each counterfactual we", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 336, 506, 349 ], "spans": [ { "bbox": [ 106, 337, 298, 349 ], "score": 1.0, "content": "gather a set of rephrased prompts equivalent to", "type": "text" }, { "bbox": [ 298, 336, 320, 349 ], "score": 0.92, "content": "( s , r )", "type": "inline_equation" }, { "bbox": [ 320, 337, 506, 349 ], "score": 1.0, "content": "and report Paraphrase Scores (PS) and (PM),", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 347, 506, 361 ], "spans": [ { "bbox": [ 105, 347, 478, 361 ], "score": 1.0, "content": "computed similarly to ES and EM. To measure specificity, we collect a set of nearby subjects", "type": "text" }, { "bbox": [ 479, 349, 489, 358 ], "score": 0.85, "content": "s _ { n }", "type": "inline_equation" }, { "bbox": [ 490, 347, 506, 361 ], "score": 1.0, "content": "for", "type": "text" } ], "index": 17 }, { "bbox": [ 104, 356, 502, 372 ], "spans": [ { "bbox": [ 104, 356, 134, 372 ], "score": 1.0, "content": "which", "type": "text" }, { "bbox": [ 134, 358, 174, 370 ], "score": 0.93, "content": "( s _ { n } , r , o ^ { c } )", "type": "inline_equation" }, { "bbox": [ 175, 356, 446, 372 ], "score": 1.0, "content": "holds true. Because we do not wish to alter these subjects, we test", "type": "text" }, { "bbox": [ 447, 358, 502, 370 ], "score": 0.93, "content": "\\mathbb { P } [ o ^ { c } ] > \\mathbb { P } [ o ^ { * } ]", "type": "inline_equation" } ], "index": 18 }, { "bbox": [ 105, 370, 506, 381 ], "spans": [ { "bbox": [ 105, 370, 506, 381 ], "score": 1.0, "content": "reporting the success fraction as Neighborhood Score (NS) and difference as (NM). To test the", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 381, 479, 393 ], "spans": [ { "bbox": [ 105, 381, 479, 393 ], "score": 1.0, "content": "generalization–specificity tradeoff, we report the harmonic mean of ES, PS, NS as Score (S).", "type": "text" } ], "index": 20 } ], "index": 15, "bbox_fs": [ 104, 271, 506, 393 ] }, { "type": "text", "bbox": [ 107, 397, 317, 517 ], "lines": [ { "bbox": [ 106, 396, 317, 408 ], "spans": [ { "bbox": [ 106, 396, 317, 408 ], "score": 1.0, "content": "We also wish to measure semantic consistency of", "type": "text" } ], "index": 21 }, { "bbox": [ 107, 407, 318, 420 ], "spans": [ { "bbox": [ 107, 408, 118, 418 ], "score": 0.83, "content": "G ^ { \\prime }", "type": "inline_equation" }, { "bbox": [ 118, 407, 318, 420 ], "score": 1.0, "content": "’s generations. To do so, we generate text start-", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 418, 318, 430 ], "spans": [ { "bbox": [ 105, 418, 143, 430 ], "score": 1.0, "content": "ing with", "type": "text" }, { "bbox": [ 143, 420, 149, 429 ], "score": 0.67, "content": "s", "type": "inline_equation" }, { "bbox": [ 150, 418, 318, 430 ], "score": 1.0, "content": "and report (RS) as the cos similarity be-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 429, 318, 441 ], "spans": [ { "bbox": [ 105, 429, 318, 441 ], "score": 1.0, "content": "tween the unigram TF-IDF vectors of generated texts,", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 440, 317, 453 ], "spans": [ { "bbox": [ 105, 440, 317, 453 ], "score": 1.0, "content": "compared to reference texts about subjects sharing", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 451, 317, 464 ], "spans": [ { "bbox": [ 105, 451, 184, 464 ], "score": 1.0, "content": "the target property", "type": "text" }, { "bbox": [ 185, 452, 195, 461 ], "score": 0.84, "content": "o ^ { * }", "type": "inline_equation" }, { "bbox": [ 195, 451, 317, 464 ], "score": 1.0, "content": ". Finally, we monitor fluency", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 462, 317, 474 ], "spans": [ { "bbox": [ 105, 462, 317, 474 ], "score": 1.0, "content": "degradations by measuring the weighted average of", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 471, 317, 485 ], "spans": [ { "bbox": [ 105, 471, 317, 485 ], "score": 1.0, "content": "bi- and tri-gram entropies (Zhang et al., 2018) given", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 483, 317, 497 ], "spans": [ { "bbox": [ 105, 483, 119, 497 ], "score": 1.0, "content": "by", "type": "text" }, { "bbox": [ 119, 484, 207, 496 ], "score": 0.81, "content": "\\begin{array} { r } { - \\sum _ { k } f ( k ) \\log _ { 2 } f ( k ) } \\end{array}", "type": "inline_equation" }, { "bbox": [ 207, 483, 239, 497 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 239, 484, 257, 496 ], "score": 0.91, "content": "f ( \\cdot )", "type": "inline_equation" }, { "bbox": [ 257, 483, 284, 497 ], "score": 1.0, "content": "is the", "type": "text" }, { "bbox": [ 284, 486, 291, 494 ], "score": 0.78, "content": "n", "type": "inline_equation" }, { "bbox": [ 291, 483, 317, 497 ], "score": 1.0, "content": "-gram", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 495, 317, 507 ], "spans": [ { "bbox": [ 106, 495, 317, 507 ], "score": 1.0, "content": "frequency distribution, which we report as (GE); this", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 505, 298, 518 ], "spans": [ { "bbox": [ 105, 505, 298, 518 ], "score": 1.0, "content": "quantity drops if text generations are repetitive.", "type": "text" } ], "index": 31 } ], "index": 26, "bbox_fs": [ 105, 396, 318, 518 ] }, { "type": "text", "bbox": [ 107, 522, 317, 587 ], "lines": [ { "bbox": [ 106, 522, 317, 534 ], "spans": [ { "bbox": [ 106, 522, 317, 534 ], "score": 1.0, "content": "In order to facilitate the above measurements, we", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 532, 317, 545 ], "spans": [ { "bbox": [ 106, 532, 317, 545 ], "score": 1.0, "content": "introduce COUNTERFACT, a challenging evaluation", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 542, 317, 556 ], "spans": [ { "bbox": [ 105, 542, 317, 556 ], "score": 1.0, "content": "dataset for evaluating counterfactual edits in language", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 555, 317, 567 ], "spans": [ { "bbox": [ 106, 555, 317, 567 ], "score": 1.0, "content": "models. Containing 21,919 records with a diverse", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 565, 318, 578 ], "spans": [ { "bbox": [ 106, 565, 318, 578 ], "score": 1.0, "content": "set of subjects, relations, and linguistic variations,", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 576, 318, 588 ], "spans": [ { "bbox": [ 106, 576, 318, 588 ], "score": 1.0, "content": "COUNTERFACT’s goal is to differentiate robust stor-", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 587, 505, 600 ], "spans": [ { "bbox": [ 105, 587, 505, 600 ], "score": 1.0, "content": "age of new facts from the superficial regurgitation of target words. See Appendix D for additional", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 597, 450, 611 ], "spans": [ { "bbox": [ 105, 597, 450, 611 ], "score": 1.0, "content": "technical details about its construction, and Table 2 for a summary of its composition.", "type": "text" } ], "index": 52 } ], "index": 34.5, "bbox_fs": [ 105, 522, 318, 588 ] }, { "type": "table", "bbox": [ 328, 410, 504, 508 ], "blocks": [ { "type": "table_caption", "bbox": [ 342, 399, 487, 409 ], "group_id": 0, "lines": [ { "bbox": [ 341, 397, 487, 411 ], "spans": [ { "bbox": [ 341, 397, 487, 411 ], "score": 1.0, "content": "Table 2: COUNTERFACT Composition", "type": "text" } ], "index": 38 } ], "index": 38 }, { "type": "table_body", "bbox": [ 328, 410, 504, 508 ], "group_id": 0, "lines": [ { "bbox": [ 328, 410, 504, 508 ], "spans": [ { "bbox": [ 328, 410, 504, 508 ], "score": 0.975, "html": "
Per Per Item Total Relation Record
Records 21919 645 1
Subjects 20391 624 1
Objects 749 60 1
Counterfactual Statements 21595 635 1
Paraphrase Prompts 42876 1262 2
Neighborhood Prompts 82650 2441 10
Generation Prompts 62346 1841 3
", "type": "table", "image_path": "56dd6344eec28b2447ac46d99df4b8d20fb1b574bc947630b3237d3833c61664.jpg" } ] } ], "index": 42, "virtual_lines": [ { "bbox": [ 328, 410, 504, 424.0 ], "spans": [], "index": 39 }, { "bbox": [ 328, 424.0, 504, 438.0 ], "spans": [], "index": 40 }, { "bbox": [ 328, 438.0, 504, 452.0 ], "spans": [], "index": 41 }, { "bbox": [ 328, 452.0, 504, 466.0 ], "spans": [], "index": 42 }, { "bbox": [ 328, 466.0, 504, 480.0 ], "spans": [], "index": 43 }, { "bbox": [ 328, 480.0, 504, 494.0 ], "spans": [], "index": 44 }, { "bbox": [ 328, 494.0, 504, 508.0 ], "spans": [], "index": 45 } ] } ], "index": 40.0 }, { "type": "table", "bbox": [ 325, 527, 502, 582 ], "blocks": [ { "type": "table_caption", "bbox": [ 330, 514, 498, 524 ], "group_id": 1, "lines": [ { "bbox": [ 329, 513, 500, 526 ], "spans": [ { "bbox": [ 329, 513, 500, 526 ], "score": 1.0, "content": "Table 3: Comparison to Existing Benchmarks", "type": "text" } ], "index": 46 } ], "index": 46 }, { "type": "table_body", "bbox": [ 325, 527, 502, 582 ], "group_id": 1, "lines": [ { "bbox": [ 325, 527, 502, 582 ], "spans": [ { "bbox": [ 325, 527, 502, 582 ], "score": 0.968, "html": "
CriterionSQuAD zSRE FEVER WikiTextPARAREL CF
Efficacy<<xx<<xxxvvv<>
Generalization
Bleedover
Consistency<<xxx/xxxx
FluencyX<<xxx
", "type": "table", "image_path": "0cc2be581f070ea69322025c4a04be9eb9c48b44947f2353c02b9f51712ad1cc.jpg" } ] } ], "index": 48.5, "virtual_lines": [ { "bbox": [ 325, 527, 502, 540.75 ], "spans": [], "index": 47 }, { "bbox": [ 325, 540.75, 502, 554.5 ], "spans": [], "index": 48 }, { "bbox": [ 325, 554.5, 502, 568.25 ], "spans": [], "index": 49 }, { "bbox": [ 325, 568.25, 502, 582.0 ], "spans": [], "index": 50 } ] } ], "index": 47.25 }, { "type": "text", "bbox": [ 107, 588, 505, 610 ], "lines": [], "index": 51.5, "bbox_fs": [ 105, 587, 505, 611 ], "lines_deleted": true }, { "type": "title", "bbox": [ 107, 624, 448, 636 ], "lines": [ { "bbox": [ 104, 622, 450, 640 ], "spans": [ { "bbox": [ 104, 622, 450, 640 ], "score": 1.0, "content": "3.4 Confirming the Importance of Decisive States Identified by Causal Tracing", "type": "text" } ], "index": 53 } ], "index": 53 }, { "type": "text", "bbox": [ 107, 645, 506, 722 ], "lines": [ { "bbox": [ 105, 644, 507, 658 ], "spans": [ { "bbox": [ 105, 644, 507, 658 ], "score": 1.0, "content": "In Section 2, we used Causal Tracing to identify decisive hidden states. To confirm that factual asso-", "type": "text" } ], "index": 54 }, { "bbox": [ 106, 656, 506, 668 ], "spans": [ { "bbox": [ 106, 656, 506, 668 ], "score": 1.0, "content": "ciations are indeed stored in the MLP modules that output those states, we test ROME’s effectiveness", "type": "text" } ], "index": 55 }, { "bbox": [ 105, 667, 505, 680 ], "spans": [ { "bbox": [ 105, 667, 505, 680 ], "score": 1.0, "content": "when targeted at various layers and tokens. Figure 5 plots four metrics evaluating both generalization", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 678, 506, 690 ], "spans": [ { "bbox": [ 105, 678, 506, 690 ], "score": 1.0, "content": "(a,b,d) and specificity (c). We observe strong correlations with the causal analysis; rewrites are most", "type": "text" } ], "index": 57 }, { "bbox": [ 106, 689, 506, 701 ], "spans": [ { "bbox": [ 106, 689, 506, 701 ], "score": 1.0, "content": "successful at the last subject token, where both specificity and generalization peak at middle layers.", "type": "text" } ], "index": 58 }, { "bbox": [ 106, 700, 505, 712 ], "spans": [ { "bbox": [ 106, 700, 505, 712 ], "score": 1.0, "content": "Targeting earlier or later tokens results in poor generalization and/or specificity. Furthermore, the", "type": "text" } ], "index": 59 }, { "bbox": [ 105, 710, 505, 725 ], "spans": [ { "bbox": [ 105, 710, 505, 725 ], "score": 1.0, "content": "layers at which edits generalize best correspond to the middle layers of the early site identified by", "type": "text" } ], "index": 60 } ], "index": 57, "bbox_fs": [ 105, 644, 507, 725 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 106, 118, 506, 309 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 71, 505, 112 ], "group_id": 0, "lines": [ { "bbox": [ 105, 70, 505, 83 ], "spans": [ { "bbox": [ 105, 70, 254, 83 ], "score": 1.0, "content": "Table 4: Quantitative Editing Results.", "type": "text" }, { "bbox": [ 254, 71, 272, 81 ], "score": 0.86, "content": "9 5 \\%", "type": "inline_equation" }, { "bbox": [ 273, 70, 505, 83 ], "score": 1.0, "content": "confidence intervals are in parentheses. Green numbers indicate", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 82, 505, 92 ], "spans": [ { "bbox": [ 106, 82, 505, 92 ], "score": 1.0, "content": "columnwise maxima, whereas red numbers indicate a clear failure on either generalization or specificity. The", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 91, 505, 103 ], "spans": [ { "bbox": [ 106, 91, 505, 103 ], "score": 1.0, "content": "presence of red in a column might explain excellent results in another. For example, on GPT-J, FT achieves", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 101, 361, 112 ], "spans": [ { "bbox": [ 106, 101, 128, 110 ], "score": 0.84, "content": "\\bar { 1 } 0 0 \\%", "type": "inline_equation" }, { "bbox": [ 128, 101, 199, 112 ], "score": 1.0, "content": "efficacy, but nearly", "type": "text" }, { "bbox": [ 199, 101, 217, 110 ], "score": 0.86, "content": "90 \\%", "type": "inline_equation" }, { "bbox": [ 217, 101, 361, 112 ], "score": 1.0, "content": "of neighborhood prompts are incorrect.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "table_body", "bbox": [ 106, 118, 506, 309 ], "group_id": 0, "lines": [ { "bbox": [ 106, 118, 506, 309 ], "spans": [ { "bbox": [ 106, 118, 506, 309 ], "score": 0.985, "html": "
EditorScore S↑EfficacyGeneralizationSpecificityFluencyConsistency
ES↑EM个PS个PM个NS↑NM↑GE个RS个
GPT-2 XL30.522.2 (0.9)-4.8 (0.3)24.7 (0.8)-5.0 (0.3)78.1 (0.6)5.0 (0.2)626.6 (0.3)31.9 (0.2)
FT65.1100.0 (0.0)98.8 (0.1)87.9 (0.6)46.6 (0.8)40.4 (0.7)-6.2 (0.4)607.1 (1.1)40.5 (0.3)
FT+L66.999.1 (0.2)91.5 (0.5)48.7 (1.0)28.9 (0.8)70.3 (0.7)3.5 (0.3)621.4 (1.0)37.4 (0.3)
KN35.628.7 (1.0)-3.4 (0.3)28.0 (0.9)-3.3 (0.2)72.9 (0.7)3.7 (0.2)570.4 (2.3)30.3 (0.3)
KE52.284.3 (0.8)33.9 (0.9)75.4 (0.8)14.6 (0.6)30.9 (0.7)-11.0 (0.5)586.6 (2.1)31.2 (0.3)
KE-CF18.199.9 (0.1)97.0 (0.2)95.8 (0.4)59.2 (0.8)6.9 (0.3)-63.2 (0.7)383.0 (4.1)24.5 (0.4)
MEND57.999.1 (0.2)70.9 (0.8)65.4 (0.9)12.2 (0.6)37.9 (0.7)-11.6 (0.5)624.2 (0.4)34.8 (0.3)
MEND-CF14.9100.0 (0.0)99.2 (0.1)97.0 (0.3)65.6 (0.7)5.5 (0.3)-69.9 (0.6)570.0 (2.1)33.2 (0.3)
ROME89.2100.0 (0.1)97.9 (0.2)96.4 (0.3)62.7 (0.8)75.4 (0.7)4.2 (0.2)621.9 (0.5)41.9 (0.3)
GPT-J23.616.3 (1.6)-7.2 (0.7)18.6 (1.5)-7.4 (0.6)83.0 (1.1)7.3 (0.5)621.8 (0.6)29.8 (0.5)
FT25.5100.0 (0.0)99.9 (0.0)96.6 (0.6)71.0 (1.5)10.3 (0.8)-50.7 (1.3)387.8 (7.3)24.6 (0.8)
FT+L68.799.6 (0.3)95.0 (0.6)47.9 (1.9)30.4 (1.5)78.6 (1.2)6.8 (0.5)622.8 (0.6)35.5 (0.5)
MEND63.297.4 (0.7)71.5 (1.6)53.6 (1.9)11.0 (1.3)53.9 (1.4)-6.0 (0.9)620.5 (0.7)32.6 (0.5)
ROME91.599.9 (0.1)99.4 (0.3)99.1 (0.3)74.1 (1.3)78.9 (1.2)5.2 (0.5)620.1 (0.9)43.0 (0.6)
", "type": "table", "image_path": "81f7c5ee028d178abcc9e36d3316ee2dd561914da80719413243efebfe5b138d.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 106, 118, 506, 181.66666666666666 ], "spans": [], "index": 4 }, { "bbox": [ 106, 181.66666666666666, 506, 245.33333333333331 ], "spans": [], "index": 5 }, { "bbox": [ 106, 245.33333333333331, 506, 309.0 ], "spans": [], "index": 6 } ] } ], "index": 3.25 }, { "type": "text", "bbox": [ 108, 330, 504, 363 ], "lines": [ { "bbox": [ 106, 330, 505, 343 ], "spans": [ { "bbox": [ 106, 330, 505, 343 ], "score": 1.0, "content": "Causal Tracing, with generalization peaking at the 18th layer. This evidence suggests that we have an", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 342, 505, 353 ], "spans": [ { "bbox": [ 105, 342, 505, 353 ], "score": 1.0, "content": "accurate understanding not only of where factual associations are stored, but also how. Appendix I", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 352, 477, 365 ], "spans": [ { "bbox": [ 106, 352, 477, 365 ], "score": 1.0, "content": "furthermore demonstrates that editing the late-layer attention modules leads to regurgitation.", "type": "text" } ], "index": 9 } ], "index": 8 }, { "type": "text", "bbox": [ 107, 368, 505, 500 ], "lines": [ { "bbox": [ 105, 369, 506, 381 ], "spans": [ { "bbox": [ 105, 369, 506, 381 ], "score": 1.0, "content": "Table 4 showcases quantitative results on GPT-2 XL (1.5B) and GPT-J (6B) over 7,500 and 2,000-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 379, 506, 392 ], "spans": [ { "bbox": [ 105, 379, 506, 392 ], "score": 1.0, "content": "record test sets in COUNTERFACT, respectively. In this experiment, in addition to the baselines tested", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 391, 505, 403 ], "spans": [ { "bbox": [ 106, 391, 505, 403 ], "score": 1.0, "content": "above, we compare with a method based on neuron interpretability, Knowledge Neurons (KN) (Dai", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 400, 507, 415 ], "spans": [ { "bbox": [ 105, 400, 507, 415 ], "score": 1.0, "content": "et al., 2022), which first selects neurons associated with knowledge via gradient-based attribution,", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 412, 506, 425 ], "spans": [ { "bbox": [ 105, 412, 506, 425 ], "score": 1.0, "content": "then modifies MLP weights at corresponding rows by adding scaled embedding vectors. We observe", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 423, 506, 435 ], "spans": [ { "bbox": [ 105, 423, 506, 435 ], "score": 1.0, "content": "that all tested methods other than ROME exhibit one or both of the following problems: (F1)", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 433, 506, 448 ], "spans": [ { "bbox": [ 105, 433, 506, 448 ], "score": 1.0, "content": "overfitting to the counterfactual statement and failing to generalize, or (F2) underfitting and predicting", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 443, 506, 459 ], "spans": [ { "bbox": [ 105, 443, 506, 459 ], "score": 1.0, "content": "the same new output for unrelated subjects. FT achieves high generalization at the cost of making", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 456, 505, 468 ], "spans": [ { "bbox": [ 105, 456, 362, 468 ], "score": 1.0, "content": "mistakes on most neighboring entities (F2); the reverse is true of", "type": "text" }, { "bbox": [ 362, 456, 387, 466 ], "score": 0.77, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 387, 456, 505, 468 ], "score": 1.0, "content": "(F1). KE- and MEND-edited", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 466, 505, 480 ], "spans": [ { "bbox": [ 105, 466, 235, 480 ], "score": 1.0, "content": "models exhibit issues with both", "type": "text" }, { "bbox": [ 235, 467, 263, 478 ], "score": 0.78, "content": "\\mathrm { F } 1 { + } \\mathrm { F } 2", "type": "inline_equation" }, { "bbox": [ 263, 466, 505, 480 ], "score": 1.0, "content": "; generalization, consistency, and bleedover are poor despite", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 477, 507, 491 ], "spans": [ { "bbox": [ 105, 477, 404, 491 ], "score": 1.0, "content": "high efficacy, indicating regurgitation. KN is unable to make effective edits", "type": "text" }, { "bbox": [ 405, 478, 436, 489 ], "score": 0.59, "content": "( \\mathrm { F } 1 { + } \\mathrm { F } 2 )", "type": "inline_equation" }, { "bbox": [ 436, 477, 507, 491 ], "score": 1.0, "content": "). By comparison,", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 487, 334, 502 ], "spans": [ { "bbox": [ 105, 487, 334, 502 ], "score": 1.0, "content": "ROME demonstrates both generalization and specificity.", "type": "text" } ], "index": 21 } ], "index": 15.5 }, { "type": "title", "bbox": [ 108, 514, 263, 525 ], "lines": [ { "bbox": [ 104, 512, 264, 529 ], "spans": [ { "bbox": [ 104, 512, 264, 529 ], "score": 1.0, "content": "3.5 Comparing Generation Results", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 106, 534, 505, 655 ], "lines": [ { "bbox": [ 106, 534, 506, 547 ], "spans": [ { "bbox": [ 106, 534, 506, 547 ], "score": 1.0, "content": "Figure 6 compares generated text after applying the counterfactual “Pierre Curie’s area of work is", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 545, 506, 558 ], "spans": [ { "bbox": [ 105, 545, 506, 558 ], "score": 1.0, "content": "medicine” to GPT-2 XL (he is actually a physicist). Generalization: In this case, FT and ROME", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 556, 506, 569 ], "spans": [ { "bbox": [ 105, 556, 506, 569 ], "score": 1.0, "content": "generalize well to paraphrases, describing the subject as a physician rather than a physicist for various", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 567, 504, 579 ], "spans": [ { "bbox": [ 106, 567, 228, 579 ], "score": 1.0, "content": "wordings. On the other hand,", "type": "text" }, { "bbox": [ 229, 567, 254, 578 ], "score": 0.78, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 255, 567, 504, 579 ], "score": 1.0, "content": ", KE and MEND fail to generalize to paraphrases, alternately", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 577, 505, 591 ], "spans": [ { "bbox": [ 105, 577, 505, 591 ], "score": 1.0, "content": "describing the subject as either (c,d,e1) in medicine or (c1,e,d1) in physics depending on the prompt’s", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 588, 506, 602 ], "spans": [ { "bbox": [ 105, 588, 506, 602 ], "score": 1.0, "content": "wording. KE (d) demonstrates a problem with fluency, favoring nonsense repetition of the word", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 600, 505, 613 ], "spans": [ { "bbox": [ 106, 600, 505, 613 ], "score": 1.0, "content": "medicine. Specificity: FT, KE, and MEND have problems with specificity, changing the profession", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 611, 506, 623 ], "spans": [ { "bbox": [ 106, 611, 506, 623 ], "score": 1.0, "content": "of a totally unrelated subject. Before editing, GPT-2 XL describes Robert Millikan as an astronomer", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 622, 506, 634 ], "spans": [ { "bbox": [ 106, 622, 506, 634 ], "score": 1.0, "content": "(in reality he is a different type of physicist), but after editing Pierre Curie’s profession, Millikan is", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 632, 507, 646 ], "spans": [ { "bbox": [ 105, 632, 231, 646 ], "score": 1.0, "content": "described as (b1) a biologist by", "type": "text" }, { "bbox": [ 231, 633, 256, 643 ], "score": 0.81, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 257, 632, 507, 646 ], "score": 1.0, "content": "and (d2, e2) a medical scientist by KE and MEND. In contrast,", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 644, 492, 656 ], "spans": [ { "bbox": [ 106, 644, 492, 656 ], "score": 1.0, "content": "ROME is specific, leaving Millikan’s field unchanged. See Appendix G for additional examples.", "type": "text" } ], "index": 33 } ], "index": 28 }, { "type": "title", "bbox": [ 107, 668, 209, 680 ], "lines": [ { "bbox": [ 106, 668, 210, 681 ], "spans": [ { "bbox": [ 106, 668, 210, 681 ], "score": 1.0, "content": "3.6 Human evaluation", "type": "text" } ], "index": 34 } ], "index": 34 }, { "type": "text", "bbox": [ 108, 689, 505, 722 ], "lines": [ { "bbox": [ 105, 688, 506, 702 ], "spans": [ { "bbox": [ 105, 688, 506, 702 ], "score": 1.0, "content": "To evaluate the quality of generated text after applying ROME, we ask 15 volunteers to evaluate", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 700, 505, 712 ], "spans": [ { "bbox": [ 105, 700, 505, 712 ], "score": 1.0, "content": "models by comparing generated text samples on the basis of both fluency and consistency with the", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 711, 499, 723 ], "spans": [ { "bbox": [ 106, 711, 283, 723 ], "score": 1.0, "content": "inserted fact. Evaluators compare ROME to", "type": "text" }, { "bbox": [ 284, 711, 308, 721 ], "score": 0.78, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 309, 711, 499, 723 ], "score": 1.0, "content": "on models modified to insert 50 different facts.", "type": "text" } ], "index": 37 } ], "index": 36 } ], "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": "table", "bbox": [ 106, 118, 506, 309 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 71, 505, 112 ], "group_id": 0, "lines": [ { "bbox": [ 105, 70, 505, 83 ], "spans": [ { "bbox": [ 105, 70, 254, 83 ], "score": 1.0, "content": "Table 4: Quantitative Editing Results.", "type": "text" }, { "bbox": [ 254, 71, 272, 81 ], "score": 0.86, "content": "9 5 \\%", "type": "inline_equation" }, { "bbox": [ 273, 70, 505, 83 ], "score": 1.0, "content": "confidence intervals are in parentheses. Green numbers indicate", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 82, 505, 92 ], "spans": [ { "bbox": [ 106, 82, 505, 92 ], "score": 1.0, "content": "columnwise maxima, whereas red numbers indicate a clear failure on either generalization or specificity. The", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 91, 505, 103 ], "spans": [ { "bbox": [ 106, 91, 505, 103 ], "score": 1.0, "content": "presence of red in a column might explain excellent results in another. For example, on GPT-J, FT achieves", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 101, 361, 112 ], "spans": [ { "bbox": [ 106, 101, 128, 110 ], "score": 0.84, "content": "\\bar { 1 } 0 0 \\%", "type": "inline_equation" }, { "bbox": [ 128, 101, 199, 112 ], "score": 1.0, "content": "efficacy, but nearly", "type": "text" }, { "bbox": [ 199, 101, 217, 110 ], "score": 0.86, "content": "90 \\%", "type": "inline_equation" }, { "bbox": [ 217, 101, 361, 112 ], "score": 1.0, "content": "of neighborhood prompts are incorrect.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "table_body", "bbox": [ 106, 118, 506, 309 ], "group_id": 0, "lines": [ { "bbox": [ 106, 118, 506, 309 ], "spans": [ { "bbox": [ 106, 118, 506, 309 ], "score": 0.985, "html": "
EditorScore S↑EfficacyGeneralizationSpecificityFluencyConsistency
ES↑EM个PS个PM个NS↑NM↑GE个RS个
GPT-2 XL30.522.2 (0.9)-4.8 (0.3)24.7 (0.8)-5.0 (0.3)78.1 (0.6)5.0 (0.2)626.6 (0.3)31.9 (0.2)
FT65.1100.0 (0.0)98.8 (0.1)87.9 (0.6)46.6 (0.8)40.4 (0.7)-6.2 (0.4)607.1 (1.1)40.5 (0.3)
FT+L66.999.1 (0.2)91.5 (0.5)48.7 (1.0)28.9 (0.8)70.3 (0.7)3.5 (0.3)621.4 (1.0)37.4 (0.3)
KN35.628.7 (1.0)-3.4 (0.3)28.0 (0.9)-3.3 (0.2)72.9 (0.7)3.7 (0.2)570.4 (2.3)30.3 (0.3)
KE52.284.3 (0.8)33.9 (0.9)75.4 (0.8)14.6 (0.6)30.9 (0.7)-11.0 (0.5)586.6 (2.1)31.2 (0.3)
KE-CF18.199.9 (0.1)97.0 (0.2)95.8 (0.4)59.2 (0.8)6.9 (0.3)-63.2 (0.7)383.0 (4.1)24.5 (0.4)
MEND57.999.1 (0.2)70.9 (0.8)65.4 (0.9)12.2 (0.6)37.9 (0.7)-11.6 (0.5)624.2 (0.4)34.8 (0.3)
MEND-CF14.9100.0 (0.0)99.2 (0.1)97.0 (0.3)65.6 (0.7)5.5 (0.3)-69.9 (0.6)570.0 (2.1)33.2 (0.3)
ROME89.2100.0 (0.1)97.9 (0.2)96.4 (0.3)62.7 (0.8)75.4 (0.7)4.2 (0.2)621.9 (0.5)41.9 (0.3)
GPT-J23.616.3 (1.6)-7.2 (0.7)18.6 (1.5)-7.4 (0.6)83.0 (1.1)7.3 (0.5)621.8 (0.6)29.8 (0.5)
FT25.5100.0 (0.0)99.9 (0.0)96.6 (0.6)71.0 (1.5)10.3 (0.8)-50.7 (1.3)387.8 (7.3)24.6 (0.8)
FT+L68.799.6 (0.3)95.0 (0.6)47.9 (1.9)30.4 (1.5)78.6 (1.2)6.8 (0.5)622.8 (0.6)35.5 (0.5)
MEND63.297.4 (0.7)71.5 (1.6)53.6 (1.9)11.0 (1.3)53.9 (1.4)-6.0 (0.9)620.5 (0.7)32.6 (0.5)
ROME91.599.9 (0.1)99.4 (0.3)99.1 (0.3)74.1 (1.3)78.9 (1.2)5.2 (0.5)620.1 (0.9)43.0 (0.6)
", "type": "table", "image_path": "81f7c5ee028d178abcc9e36d3316ee2dd561914da80719413243efebfe5b138d.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 106, 118, 506, 181.66666666666666 ], "spans": [], "index": 4 }, { "bbox": [ 106, 181.66666666666666, 506, 245.33333333333331 ], "spans": [], "index": 5 }, { "bbox": [ 106, 245.33333333333331, 506, 309.0 ], "spans": [], "index": 6 } ] } ], "index": 3.25 }, { "type": "text", "bbox": [ 108, 330, 504, 363 ], "lines": [ { "bbox": [ 106, 330, 505, 343 ], "spans": [ { "bbox": [ 106, 330, 505, 343 ], "score": 1.0, "content": "Causal Tracing, with generalization peaking at the 18th layer. This evidence suggests that we have an", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 342, 505, 353 ], "spans": [ { "bbox": [ 105, 342, 505, 353 ], "score": 1.0, "content": "accurate understanding not only of where factual associations are stored, but also how. Appendix I", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 352, 477, 365 ], "spans": [ { "bbox": [ 106, 352, 477, 365 ], "score": 1.0, "content": "furthermore demonstrates that editing the late-layer attention modules leads to regurgitation.", "type": "text" } ], "index": 9 } ], "index": 8, "bbox_fs": [ 105, 330, 505, 365 ] }, { "type": "text", "bbox": [ 107, 368, 505, 500 ], "lines": [ { "bbox": [ 105, 369, 506, 381 ], "spans": [ { "bbox": [ 105, 369, 506, 381 ], "score": 1.0, "content": "Table 4 showcases quantitative results on GPT-2 XL (1.5B) and GPT-J (6B) over 7,500 and 2,000-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 379, 506, 392 ], "spans": [ { "bbox": [ 105, 379, 506, 392 ], "score": 1.0, "content": "record test sets in COUNTERFACT, respectively. In this experiment, in addition to the baselines tested", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 391, 505, 403 ], "spans": [ { "bbox": [ 106, 391, 505, 403 ], "score": 1.0, "content": "above, we compare with a method based on neuron interpretability, Knowledge Neurons (KN) (Dai", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 400, 507, 415 ], "spans": [ { "bbox": [ 105, 400, 507, 415 ], "score": 1.0, "content": "et al., 2022), which first selects neurons associated with knowledge via gradient-based attribution,", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 412, 506, 425 ], "spans": [ { "bbox": [ 105, 412, 506, 425 ], "score": 1.0, "content": "then modifies MLP weights at corresponding rows by adding scaled embedding vectors. We observe", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 423, 506, 435 ], "spans": [ { "bbox": [ 105, 423, 506, 435 ], "score": 1.0, "content": "that all tested methods other than ROME exhibit one or both of the following problems: (F1)", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 433, 506, 448 ], "spans": [ { "bbox": [ 105, 433, 506, 448 ], "score": 1.0, "content": "overfitting to the counterfactual statement and failing to generalize, or (F2) underfitting and predicting", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 443, 506, 459 ], "spans": [ { "bbox": [ 105, 443, 506, 459 ], "score": 1.0, "content": "the same new output for unrelated subjects. FT achieves high generalization at the cost of making", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 456, 505, 468 ], "spans": [ { "bbox": [ 105, 456, 362, 468 ], "score": 1.0, "content": "mistakes on most neighboring entities (F2); the reverse is true of", "type": "text" }, { "bbox": [ 362, 456, 387, 466 ], "score": 0.77, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 387, 456, 505, 468 ], "score": 1.0, "content": "(F1). KE- and MEND-edited", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 466, 505, 480 ], "spans": [ { "bbox": [ 105, 466, 235, 480 ], "score": 1.0, "content": "models exhibit issues with both", "type": "text" }, { "bbox": [ 235, 467, 263, 478 ], "score": 0.78, "content": "\\mathrm { F } 1 { + } \\mathrm { F } 2", "type": "inline_equation" }, { "bbox": [ 263, 466, 505, 480 ], "score": 1.0, "content": "; generalization, consistency, and bleedover are poor despite", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 477, 507, 491 ], "spans": [ { "bbox": [ 105, 477, 404, 491 ], "score": 1.0, "content": "high efficacy, indicating regurgitation. KN is unable to make effective edits", "type": "text" }, { "bbox": [ 405, 478, 436, 489 ], "score": 0.59, "content": "( \\mathrm { F } 1 { + } \\mathrm { F } 2 )", "type": "inline_equation" }, { "bbox": [ 436, 477, 507, 491 ], "score": 1.0, "content": "). By comparison,", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 487, 334, 502 ], "spans": [ { "bbox": [ 105, 487, 334, 502 ], "score": 1.0, "content": "ROME demonstrates both generalization and specificity.", "type": "text" } ], "index": 21 } ], "index": 15.5, "bbox_fs": [ 105, 369, 507, 502 ] }, { "type": "title", "bbox": [ 108, 514, 263, 525 ], "lines": [ { "bbox": [ 104, 512, 264, 529 ], "spans": [ { "bbox": [ 104, 512, 264, 529 ], "score": 1.0, "content": "3.5 Comparing Generation Results", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 106, 534, 505, 655 ], "lines": [ { "bbox": [ 106, 534, 506, 547 ], "spans": [ { "bbox": [ 106, 534, 506, 547 ], "score": 1.0, "content": "Figure 6 compares generated text after applying the counterfactual “Pierre Curie’s area of work is", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 545, 506, 558 ], "spans": [ { "bbox": [ 105, 545, 506, 558 ], "score": 1.0, "content": "medicine” to GPT-2 XL (he is actually a physicist). Generalization: In this case, FT and ROME", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 556, 506, 569 ], "spans": [ { "bbox": [ 105, 556, 506, 569 ], "score": 1.0, "content": "generalize well to paraphrases, describing the subject as a physician rather than a physicist for various", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 567, 504, 579 ], "spans": [ { "bbox": [ 106, 567, 228, 579 ], "score": 1.0, "content": "wordings. On the other hand,", "type": "text" }, { "bbox": [ 229, 567, 254, 578 ], "score": 0.78, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 255, 567, 504, 579 ], "score": 1.0, "content": ", KE and MEND fail to generalize to paraphrases, alternately", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 577, 505, 591 ], "spans": [ { "bbox": [ 105, 577, 505, 591 ], "score": 1.0, "content": "describing the subject as either (c,d,e1) in medicine or (c1,e,d1) in physics depending on the prompt’s", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 588, 506, 602 ], "spans": [ { "bbox": [ 105, 588, 506, 602 ], "score": 1.0, "content": "wording. KE (d) demonstrates a problem with fluency, favoring nonsense repetition of the word", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 600, 505, 613 ], "spans": [ { "bbox": [ 106, 600, 505, 613 ], "score": 1.0, "content": "medicine. Specificity: FT, KE, and MEND have problems with specificity, changing the profession", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 611, 506, 623 ], "spans": [ { "bbox": [ 106, 611, 506, 623 ], "score": 1.0, "content": "of a totally unrelated subject. Before editing, GPT-2 XL describes Robert Millikan as an astronomer", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 622, 506, 634 ], "spans": [ { "bbox": [ 106, 622, 506, 634 ], "score": 1.0, "content": "(in reality he is a different type of physicist), but after editing Pierre Curie’s profession, Millikan is", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 632, 507, 646 ], "spans": [ { "bbox": [ 105, 632, 231, 646 ], "score": 1.0, "content": "described as (b1) a biologist by", "type": "text" }, { "bbox": [ 231, 633, 256, 643 ], "score": 0.81, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 257, 632, 507, 646 ], "score": 1.0, "content": "and (d2, e2) a medical scientist by KE and MEND. In contrast,", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 644, 492, 656 ], "spans": [ { "bbox": [ 106, 644, 492, 656 ], "score": 1.0, "content": "ROME is specific, leaving Millikan’s field unchanged. See Appendix G for additional examples.", "type": "text" } ], "index": 33 } ], "index": 28, "bbox_fs": [ 105, 534, 507, 656 ] }, { "type": "title", "bbox": [ 107, 668, 209, 680 ], "lines": [ { "bbox": [ 106, 668, 210, 681 ], "spans": [ { "bbox": [ 106, 668, 210, 681 ], "score": 1.0, "content": "3.6 Human evaluation", "type": "text" } ], "index": 34 } ], "index": 34 }, { "type": "text", "bbox": [ 108, 689, 505, 722 ], "lines": [ { "bbox": [ 105, 688, 506, 702 ], "spans": [ { "bbox": [ 105, 688, 506, 702 ], "score": 1.0, "content": "To evaluate the quality of generated text after applying ROME, we ask 15 volunteers to evaluate", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 700, 505, 712 ], "spans": [ { "bbox": [ 105, 700, 505, 712 ], "score": 1.0, "content": "models by comparing generated text samples on the basis of both fluency and consistency with the", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 711, 499, 723 ], "spans": [ { "bbox": [ 106, 711, 283, 723 ], "score": 1.0, "content": "inserted fact. Evaluators compare ROME to", "type": "text" }, { "bbox": [ 284, 711, 308, 721 ], "score": 0.78, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 309, 711, 499, 723 ], "score": 1.0, "content": "on models modified to insert 50 different facts.", "type": "text" } ], "index": 37 } ], "index": 36, "bbox_fs": [ 105, 688, 506, 723 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 108, 70, 503, 240 ], "blocks": [ { "type": "image_body", "bbox": [ 108, 70, 503, 240 ], "group_id": 0, "lines": [ { "bbox": [ 108, 70, 503, 240 ], "spans": [ { "bbox": [ 108, 70, 503, 240 ], "score": 0.638, "type": "image", "image_path": "c650a665431eec32346f08d0d8cf41a19b248635548033b50246702212f597f6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 108, 70, 503, 126.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 108, 126.66666666666666, 503, 183.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 108, 183.33333333333331, 503, 239.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 247, 505, 278 ], "group_id": 0, "lines": [ { "bbox": [ 105, 245, 505, 259 ], "spans": [ { "bbox": [ 105, 245, 505, 259 ], "score": 1.0, "content": "Figure 6: Comparison of generated text. Prompts are italicized, green and red indicate keywords reflecting", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 255, 505, 270 ], "spans": [ { "bbox": [ 105, 255, 505, 270 ], "score": 1.0, "content": "correct and incorrect behavior, respectively, and blue indicates a factually-incorrect keyword that was already", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 267, 351, 279 ], "spans": [ { "bbox": [ 105, 267, 144, 279 ], "score": 1.0, "content": "present in", "type": "text" }, { "bbox": [ 144, 268, 153, 276 ], "score": 0.75, "content": "G", "type": "inline_equation" }, { "bbox": [ 153, 267, 351, 279 ], "score": 1.0, "content": "before rewriting. See Section 3.5 for detailed analysis.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 290, 505, 356 ], "lines": [ { "bbox": [ 106, 291, 505, 302 ], "spans": [ { "bbox": [ 106, 291, 505, 302 ], "score": 1.0, "content": "We find that evaluators are 1.8 times more likely to rate ROME as more consistent with the inserted", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 301, 505, 314 ], "spans": [ { "bbox": [ 105, 301, 160, 314 ], "score": 1.0, "content": "fact than the", "type": "text" }, { "bbox": [ 160, 302, 186, 312 ], "score": 0.81, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 186, 301, 505, 314 ], "score": 1.0, "content": "model, confirming the efficacy and generalization of the model that has been", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 312, 506, 325 ], "spans": [ { "bbox": [ 105, 312, 506, 325 ], "score": 1.0, "content": "observed in our other metrics. However, evaluators find text generated by ROME to be somewhat", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 323, 506, 336 ], "spans": [ { "bbox": [ 105, 323, 255, 336 ], "score": 1.0, "content": "less fluent than models editing using", "type": "text" }, { "bbox": [ 255, 324, 280, 334 ], "score": 0.68, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 280, 323, 506, 336 ], "score": 1.0, "content": ", rating ROME as 1.3 times less likely to be more fluent", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 334, 505, 347 ], "spans": [ { "bbox": [ 105, 334, 140, 347 ], "score": 1.0, "content": "than the", "type": "text" }, { "bbox": [ 140, 334, 165, 345 ], "score": 0.76, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 165, 334, 505, 347 ], "score": 1.0, "content": "model, suggesting that ROME introduces some loss in fluency that is not captured by", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 346, 454, 357 ], "spans": [ { "bbox": [ 106, 346, 454, 357 ], "score": 1.0, "content": "our other metrics. Further details of the human evaluation can be found in Appendix J.", "type": "text" } ], "index": 11 } ], "index": 8.5 }, { "type": "title", "bbox": [ 107, 370, 180, 382 ], "lines": [ { "bbox": [ 105, 370, 181, 384 ], "spans": [ { "bbox": [ 105, 370, 181, 384 ], "score": 1.0, "content": "3.7 Limitations", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 107, 391, 505, 457 ], "lines": [ { "bbox": [ 106, 390, 506, 403 ], "spans": [ { "bbox": [ 106, 390, 506, 403 ], "score": 1.0, "content": "The purpose of ROME is to serve as a tool for understanding mechanisms of knowledge storage: it", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 402, 506, 414 ], "spans": [ { "bbox": [ 106, 402, 506, 414 ], "score": 1.0, "content": "only edits a single fact at a time, and it is not intended as a practical method for large-scale model", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 412, 506, 425 ], "spans": [ { "bbox": [ 106, 412, 506, 425 ], "score": 1.0, "content": "training. Associations edited by ROME are directional, for example, “The iconic landmark in Seattle", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 423, 506, 436 ], "spans": [ { "bbox": [ 106, 423, 506, 436 ], "score": 1.0, "content": "is the Space Needle” is stored separately from “The Space Needle is the iconic landmark in Seattle,”", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 435, 504, 446 ], "spans": [ { "bbox": [ 106, 435, 504, 446 ], "score": 1.0, "content": "so altering both requires two edits. A scalable approach for multiple simultaneous edits built upon", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 446, 479, 457 ], "spans": [ { "bbox": [ 106, 446, 479, 457 ], "score": 1.0, "content": "the ideas in ROME is developed in Meng, Sen Sharma, Andonian, Belinkov, and Bau (2022).", "type": "text" } ], "index": 18 } ], "index": 15.5 }, { "type": "text", "bbox": [ 107, 462, 505, 528 ], "lines": [ { "bbox": [ 105, 461, 507, 474 ], "spans": [ { "bbox": [ 105, 461, 507, 474 ], "score": 1.0, "content": "ROME and Causal Tracing have shed light on factual association within GPT, but we have not inves-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 471, 507, 487 ], "spans": [ { "bbox": [ 105, 471, 507, 487 ], "score": 1.0, "content": "tigated other kinds of learned beliefs such as logical, spatial, or numerical knowledge. Furthermore,", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 483, 506, 497 ], "spans": [ { "bbox": [ 105, 483, 506, 497 ], "score": 1.0, "content": "our understanding of the structure of the vector spaces that represent learned attributes remains", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 494, 505, 506 ], "spans": [ { "bbox": [ 105, 494, 505, 506 ], "score": 1.0, "content": "incomplete. Even when a model’s stored factual association is changed successfully, the model will", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 505, 506, 518 ], "spans": [ { "bbox": [ 105, 505, 506, 518 ], "score": 1.0, "content": "guess plausible new facts that have no basis in evidence and that are likely to be false. This may limit", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 517, 329, 528 ], "spans": [ { "bbox": [ 106, 517, 329, 528 ], "score": 1.0, "content": "the usefulness of a language model as a source of facts.", "type": "text" } ], "index": 24 } ], "index": 21.5 }, { "type": "title", "bbox": [ 107, 549, 197, 563 ], "lines": [ { "bbox": [ 105, 549, 198, 564 ], "spans": [ { "bbox": [ 105, 549, 198, 564 ], "score": 1.0, "content": "4 Related Work", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 579, 505, 722 ], "lines": [ { "bbox": [ 106, 579, 505, 591 ], "spans": [ { "bbox": [ 106, 579, 505, 591 ], "score": 1.0, "content": "The question of what a model learns is a fundamental problem that has been approached from several", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 590, 506, 603 ], "spans": [ { "bbox": [ 105, 590, 506, 603 ], "score": 1.0, "content": "directions. One line of work studies which properties are encoded in internal model representations,", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 602, 505, 614 ], "spans": [ { "bbox": [ 106, 602, 505, 614 ], "score": 1.0, "content": "most commonly by training a probing classifier to predict said properties from the representations", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 611, 506, 625 ], "spans": [ { "bbox": [ 105, 611, 506, 625 ], "score": 1.0, "content": "(Ettinger et al., 2016; Adi et al., 2017; Hupkes et al., 2018; Conneau et al., 2018; Belinkov et al.,", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 622, 506, 636 ], "spans": [ { "bbox": [ 105, 622, 506, 636 ], "score": 1.0, "content": "2017; Belinkov & Glass, 2019, inter alia). However, such approaches suffer from various limitations,", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 635, 505, 646 ], "spans": [ { "bbox": [ 106, 635, 505, 646 ], "score": 1.0, "content": "notably being dissociated from the network’s behavior (Belinkov, 2021). In contrast, causal effects", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 644, 506, 658 ], "spans": [ { "bbox": [ 105, 644, 506, 658 ], "score": 1.0, "content": "have been used to probe important information within a network in a way that avoids misleading", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 655, 506, 669 ], "spans": [ { "bbox": [ 105, 655, 506, 669 ], "score": 1.0, "content": "spurious correlations. Vig et al. (2020b,a) introduced the use of causal mediation analysis to identify", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 666, 506, 680 ], "spans": [ { "bbox": [ 106, 666, 506, 680 ], "score": 1.0, "content": "individual neurons that contribute to biased gender assumptions, and Finlayson et al. (2021) have", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 678, 506, 690 ], "spans": [ { "bbox": [ 106, 678, 506, 690 ], "score": 1.0, "content": "used a similar methodology to investigate mechanisms of syntactic agreement in language models.", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 687, 506, 702 ], "spans": [ { "bbox": [ 105, 687, 506, 702 ], "score": 1.0, "content": "Feder et al. (2021) described a framework that applies interventions on representations and weights to", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 700, 505, 713 ], "spans": [ { "bbox": [ 106, 700, 505, 713 ], "score": 1.0, "content": "understand the causal structure of models. Elazar et al. (2021b) proposed erasing specific information", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 710, 505, 724 ], "spans": [ { "bbox": [ 105, 710, 505, 724 ], "score": 1.0, "content": "from a representation in order to measure its causal effect. Extending these ideas, our Causal Tracing", "type": "text" } ], "index": 38 } ], "index": 32 } ], "page_idx": 8, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 741, 308, 750 ], "lines": [ { "bbox": [ 302, 741, 309, 752 ], "spans": [ { "bbox": [ 302, 741, 309, 752 ], "score": 1.0, "content": "9", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 108, 70, 503, 240 ], "blocks": [ { "type": "image_body", "bbox": [ 108, 70, 503, 240 ], "group_id": 0, "lines": [ { "bbox": [ 108, 70, 503, 240 ], "spans": [ { "bbox": [ 108, 70, 503, 240 ], "score": 0.638, "type": "image", "image_path": "c650a665431eec32346f08d0d8cf41a19b248635548033b50246702212f597f6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 108, 70, 503, 126.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 108, 126.66666666666666, 503, 183.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 108, 183.33333333333331, 503, 239.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 247, 505, 278 ], "group_id": 0, "lines": [ { "bbox": [ 105, 245, 505, 259 ], "spans": [ { "bbox": [ 105, 245, 505, 259 ], "score": 1.0, "content": "Figure 6: Comparison of generated text. Prompts are italicized, green and red indicate keywords reflecting", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 255, 505, 270 ], "spans": [ { "bbox": [ 105, 255, 505, 270 ], "score": 1.0, "content": "correct and incorrect behavior, respectively, and blue indicates a factually-incorrect keyword that was already", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 267, 351, 279 ], "spans": [ { "bbox": [ 105, 267, 144, 279 ], "score": 1.0, "content": "present in", "type": "text" }, { "bbox": [ 144, 268, 153, 276 ], "score": 0.75, "content": "G", "type": "inline_equation" }, { "bbox": [ 153, 267, 351, 279 ], "score": 1.0, "content": "before rewriting. See Section 3.5 for detailed analysis.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 290, 505, 356 ], "lines": [ { "bbox": [ 106, 291, 505, 302 ], "spans": [ { "bbox": [ 106, 291, 505, 302 ], "score": 1.0, "content": "We find that evaluators are 1.8 times more likely to rate ROME as more consistent with the inserted", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 301, 505, 314 ], "spans": [ { "bbox": [ 105, 301, 160, 314 ], "score": 1.0, "content": "fact than the", "type": "text" }, { "bbox": [ 160, 302, 186, 312 ], "score": 0.81, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 186, 301, 505, 314 ], "score": 1.0, "content": "model, confirming the efficacy and generalization of the model that has been", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 312, 506, 325 ], "spans": [ { "bbox": [ 105, 312, 506, 325 ], "score": 1.0, "content": "observed in our other metrics. However, evaluators find text generated by ROME to be somewhat", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 323, 506, 336 ], "spans": [ { "bbox": [ 105, 323, 255, 336 ], "score": 1.0, "content": "less fluent than models editing using", "type": "text" }, { "bbox": [ 255, 324, 280, 334 ], "score": 0.68, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 280, 323, 506, 336 ], "score": 1.0, "content": ", rating ROME as 1.3 times less likely to be more fluent", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 334, 505, 347 ], "spans": [ { "bbox": [ 105, 334, 140, 347 ], "score": 1.0, "content": "than the", "type": "text" }, { "bbox": [ 140, 334, 165, 345 ], "score": 0.76, "content": "\\mathrm { F T + L }", "type": "inline_equation" }, { "bbox": [ 165, 334, 505, 347 ], "score": 1.0, "content": "model, suggesting that ROME introduces some loss in fluency that is not captured by", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 346, 454, 357 ], "spans": [ { "bbox": [ 106, 346, 454, 357 ], "score": 1.0, "content": "our other metrics. Further details of the human evaluation can be found in Appendix J.", "type": "text" } ], "index": 11 } ], "index": 8.5, "bbox_fs": [ 105, 291, 506, 357 ] }, { "type": "title", "bbox": [ 107, 370, 180, 382 ], "lines": [ { "bbox": [ 105, 370, 181, 384 ], "spans": [ { "bbox": [ 105, 370, 181, 384 ], "score": 1.0, "content": "3.7 Limitations", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 107, 391, 505, 457 ], "lines": [ { "bbox": [ 106, 390, 506, 403 ], "spans": [ { "bbox": [ 106, 390, 506, 403 ], "score": 1.0, "content": "The purpose of ROME is to serve as a tool for understanding mechanisms of knowledge storage: it", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 402, 506, 414 ], "spans": [ { "bbox": [ 106, 402, 506, 414 ], "score": 1.0, "content": "only edits a single fact at a time, and it is not intended as a practical method for large-scale model", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 412, 506, 425 ], "spans": [ { "bbox": [ 106, 412, 506, 425 ], "score": 1.0, "content": "training. Associations edited by ROME are directional, for example, “The iconic landmark in Seattle", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 423, 506, 436 ], "spans": [ { "bbox": [ 106, 423, 506, 436 ], "score": 1.0, "content": "is the Space Needle” is stored separately from “The Space Needle is the iconic landmark in Seattle,”", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 435, 504, 446 ], "spans": [ { "bbox": [ 106, 435, 504, 446 ], "score": 1.0, "content": "so altering both requires two edits. A scalable approach for multiple simultaneous edits built upon", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 446, 479, 457 ], "spans": [ { "bbox": [ 106, 446, 479, 457 ], "score": 1.0, "content": "the ideas in ROME is developed in Meng, Sen Sharma, Andonian, Belinkov, and Bau (2022).", "type": "text" } ], "index": 18 } ], "index": 15.5, "bbox_fs": [ 106, 390, 506, 457 ] }, { "type": "text", "bbox": [ 107, 462, 505, 528 ], "lines": [ { "bbox": [ 105, 461, 507, 474 ], "spans": [ { "bbox": [ 105, 461, 507, 474 ], "score": 1.0, "content": "ROME and Causal Tracing have shed light on factual association within GPT, but we have not inves-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 471, 507, 487 ], "spans": [ { "bbox": [ 105, 471, 507, 487 ], "score": 1.0, "content": "tigated other kinds of learned beliefs such as logical, spatial, or numerical knowledge. Furthermore,", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 483, 506, 497 ], "spans": [ { "bbox": [ 105, 483, 506, 497 ], "score": 1.0, "content": "our understanding of the structure of the vector spaces that represent learned attributes remains", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 494, 505, 506 ], "spans": [ { "bbox": [ 105, 494, 505, 506 ], "score": 1.0, "content": "incomplete. Even when a model’s stored factual association is changed successfully, the model will", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 505, 506, 518 ], "spans": [ { "bbox": [ 105, 505, 506, 518 ], "score": 1.0, "content": "guess plausible new facts that have no basis in evidence and that are likely to be false. This may limit", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 517, 329, 528 ], "spans": [ { "bbox": [ 106, 517, 329, 528 ], "score": 1.0, "content": "the usefulness of a language model as a source of facts.", "type": "text" } ], "index": 24 } ], "index": 21.5, "bbox_fs": [ 105, 461, 507, 528 ] }, { "type": "title", "bbox": [ 107, 549, 197, 563 ], "lines": [ { "bbox": [ 105, 549, 198, 564 ], "spans": [ { "bbox": [ 105, 549, 198, 564 ], "score": 1.0, "content": "4 Related Work", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 579, 505, 722 ], "lines": [ { "bbox": [ 106, 579, 505, 591 ], "spans": [ { "bbox": [ 106, 579, 505, 591 ], "score": 1.0, "content": "The question of what a model learns is a fundamental problem that has been approached from several", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 590, 506, 603 ], "spans": [ { "bbox": [ 105, 590, 506, 603 ], "score": 1.0, "content": "directions. One line of work studies which properties are encoded in internal model representations,", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 602, 505, 614 ], "spans": [ { "bbox": [ 106, 602, 505, 614 ], "score": 1.0, "content": "most commonly by training a probing classifier to predict said properties from the representations", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 611, 506, 625 ], "spans": [ { "bbox": [ 105, 611, 506, 625 ], "score": 1.0, "content": "(Ettinger et al., 2016; Adi et al., 2017; Hupkes et al., 2018; Conneau et al., 2018; Belinkov et al.,", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 622, 506, 636 ], "spans": [ { "bbox": [ 105, 622, 506, 636 ], "score": 1.0, "content": "2017; Belinkov & Glass, 2019, inter alia). However, such approaches suffer from various limitations,", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 635, 505, 646 ], "spans": [ { "bbox": [ 106, 635, 505, 646 ], "score": 1.0, "content": "notably being dissociated from the network’s behavior (Belinkov, 2021). In contrast, causal effects", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 644, 506, 658 ], "spans": [ { "bbox": [ 105, 644, 506, 658 ], "score": 1.0, "content": "have been used to probe important information within a network in a way that avoids misleading", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 655, 506, 669 ], "spans": [ { "bbox": [ 105, 655, 506, 669 ], "score": 1.0, "content": "spurious correlations. Vig et al. (2020b,a) introduced the use of causal mediation analysis to identify", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 666, 506, 680 ], "spans": [ { "bbox": [ 106, 666, 506, 680 ], "score": 1.0, "content": "individual neurons that contribute to biased gender assumptions, and Finlayson et al. (2021) have", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 678, 506, 690 ], "spans": [ { "bbox": [ 106, 678, 506, 690 ], "score": 1.0, "content": "used a similar methodology to investigate mechanisms of syntactic agreement in language models.", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 687, 506, 702 ], "spans": [ { "bbox": [ 105, 687, 506, 702 ], "score": 1.0, "content": "Feder et al. (2021) described a framework that applies interventions on representations and weights to", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 700, 505, 713 ], "spans": [ { "bbox": [ 106, 700, 505, 713 ], "score": 1.0, "content": "understand the causal structure of models. Elazar et al. (2021b) proposed erasing specific information", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 710, 505, 724 ], "spans": [ { "bbox": [ 105, 710, 505, 724 ], "score": 1.0, "content": "from a representation in order to measure its causal effect. Extending these ideas, our Causal Tracing", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 73, 505, 85 ], "spans": [ { "bbox": [ 105, 73, 505, 85 ], "score": 1.0, "content": "method introduces paired interventions that allow explicit measurement of causal indirect effects", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 105, 83, 298, 96 ], "spans": [ { "bbox": [ 105, 83, 298, 96 ], "score": 1.0, "content": "(Pearl, 2001) of individual hidden state vectors.", "type": "text", "cross_page": true } ], "index": 1 } ], "index": 32, "bbox_fs": [ 105, 579, 506, 724 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 73, 504, 94 ], "lines": [ { "bbox": [ 105, 73, 505, 85 ], "spans": [ { "bbox": [ 105, 73, 505, 85 ], "score": 1.0, "content": "method introduces paired interventions that allow explicit measurement of causal indirect effects", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 83, 298, 96 ], "spans": [ { "bbox": [ 105, 83, 298, 96 ], "score": 1.0, "content": "(Pearl, 2001) of individual hidden state vectors.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 107, 100, 505, 220 ], "lines": [ { "bbox": [ 105, 99, 505, 112 ], "spans": [ { "bbox": [ 105, 99, 505, 112 ], "score": 1.0, "content": "Another line of work aims to assess the knowledge within LMs by evaluating whether the model", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 110, 506, 124 ], "spans": [ { "bbox": [ 105, 110, 506, 124 ], "score": 1.0, "content": "predict pieces of knowledge. A common strategy is to define a fill-in-the-blank prompt, and let a", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 121, 506, 135 ], "spans": [ { "bbox": [ 105, 121, 506, 135 ], "score": 1.0, "content": "masked LM complete it (Petroni et al., 2019, 2020). Later work showed that knowledge extraction", "type": "text" } ], "index": 4 }, { "bbox": [ 104, 131, 506, 147 ], "spans": [ { "bbox": [ 104, 131, 506, 147 ], "score": 1.0, "content": "can be improved by diversifying the prompts (Jiang et al., 2020; Zhong et al., 2021), or by fine-tuning", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 143, 506, 156 ], "spans": [ { "bbox": [ 105, 143, 506, 156 ], "score": 1.0, "content": "a model on open-domain textual facts (Roberts et al., 2020). However, constructing prompts from", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 154, 506, 168 ], "spans": [ { "bbox": [ 105, 154, 506, 168 ], "score": 1.0, "content": "supervised knowledge extraction data risks learning new knowledge instead of recalling existing", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 164, 506, 178 ], "spans": [ { "bbox": [ 105, 164, 506, 178 ], "score": 1.0, "content": "knowledge in an LM (Zhong et al., 2021). More recently, Elazar et al. (2021a) introduced ParaRel, a", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 177, 506, 188 ], "spans": [ { "bbox": [ 106, 177, 506, 188 ], "score": 1.0, "content": "curated dataset of paraphrased prompts and facts. We use it as a basis for constructing COUNTER-", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 187, 505, 200 ], "spans": [ { "bbox": [ 106, 187, 505, 200 ], "score": 1.0, "content": "FACT, which enables fine-grained measurements of knowledge extraction and editing along multiple", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 199, 506, 210 ], "spans": [ { "bbox": [ 106, 199, 506, 210 ], "score": 1.0, "content": "dimensions. Different from prior work, we do not strive to extract the most knowledge from a model,", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 210, 405, 221 ], "spans": [ { "bbox": [ 106, 210, 405, 221 ], "score": 1.0, "content": "but rather wish to understand mechanisms of knowledge recall in a model.", "type": "text" } ], "index": 12 } ], "index": 7 }, { "type": "text", "bbox": [ 107, 225, 505, 368 ], "lines": [ { "bbox": [ 105, 225, 506, 238 ], "spans": [ { "bbox": [ 105, 225, 506, 238 ], "score": 1.0, "content": "Finally, a few studies aim to localize and modify the computation of knowledge within transformers.", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 235, 506, 249 ], "spans": [ { "bbox": [ 105, 235, 506, 249 ], "score": 1.0, "content": "Geva et al. (2021) identify the MLP layers in a (masked LM) transformer as key–value memories", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 246, 505, 260 ], "spans": [ { "bbox": [ 105, 246, 505, 260 ], "score": 1.0, "content": "of entities and information associated with that entity. Building on this finding, Dai et al. (2022)", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 257, 506, 271 ], "spans": [ { "bbox": [ 105, 257, 506, 271 ], "score": 1.0, "content": "demonstrate a method to edit facts in BERT by writing the embedding of the object into certain rows", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 268, 507, 282 ], "spans": [ { "bbox": [ 105, 268, 507, 282 ], "score": 1.0, "content": "of the MLP matrix. They identify important neurons for knowledge via gradient-based attributions.", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 280, 506, 293 ], "spans": [ { "bbox": [ 106, 280, 506, 293 ], "score": 1.0, "content": "De Cao et al. (2021) train a hyper-network to predict a weight update at test time, which will alter a", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 290, 506, 303 ], "spans": [ { "bbox": [ 106, 290, 506, 303 ], "score": 1.0, "content": "fact. They experiment with BERT and BART (Lewis et al., 2020), a sequence-to-sequence model, and", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 301, 506, 314 ], "spans": [ { "bbox": [ 105, 301, 506, 314 ], "score": 1.0, "content": "focus on models fine-tuned for question answering. Mitchell et al. (2021) presents a hyper-network", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 313, 505, 325 ], "spans": [ { "bbox": [ 106, 313, 505, 325 ], "score": 1.0, "content": "method that learns to transform the decomposed terms of the gradient in order to efficiently predict", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 323, 505, 335 ], "spans": [ { "bbox": [ 105, 323, 505, 335 ], "score": 1.0, "content": "a knowledge update, and demonstrates the ability to scale up to large models including T5 (Raffel", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 334, 506, 347 ], "spans": [ { "bbox": [ 105, 334, 506, 347 ], "score": 1.0, "content": "et al., 2020) and GPT-J (Wang & Komatsuzaki, 2021). We compare with all these methods in our", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 345, 506, 358 ], "spans": [ { "bbox": [ 105, 345, 506, 358 ], "score": 1.0, "content": "experiments, and find that our single-layer ROME parameter intervention has comparable capabilities,", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 357, 397, 369 ], "spans": [ { "bbox": [ 106, 357, 397, 369 ], "score": 1.0, "content": "avoiding failures in specificity and generalization seen in other methods.", "type": "text" } ], "index": 25 } ], "index": 19 }, { "type": "title", "bbox": [ 107, 383, 182, 396 ], "lines": [ { "bbox": [ 104, 381, 185, 399 ], "spans": [ { "bbox": [ 104, 381, 185, 399 ], "score": 1.0, "content": "5 Conclusion", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 409, 505, 486 ], "lines": [ { "bbox": [ 106, 409, 505, 422 ], "spans": [ { "bbox": [ 106, 409, 505, 422 ], "score": 1.0, "content": "We have clarified information flow during knowledge recall in autoregressive transformers, and", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 420, 506, 432 ], "spans": [ { "bbox": [ 105, 420, 506, 432 ], "score": 1.0, "content": "we have exploited this understanding to develop a simple, principled model editor called ROME.", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 432, 506, 444 ], "spans": [ { "bbox": [ 106, 432, 506, 444 ], "score": 1.0, "content": "Our experiments provide insight into how facts are stored and demonstrate the feasibility of direct", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 442, 505, 454 ], "spans": [ { "bbox": [ 105, 442, 505, 454 ], "score": 1.0, "content": "manipulation of computational mechanisms in large pretrained models. While the methods in this", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 453, 506, 465 ], "spans": [ { "bbox": [ 105, 453, 506, 465 ], "score": 1.0, "content": "paper serve to test the locality of knowledge within a model, they apply only to editing a single fact", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 464, 505, 476 ], "spans": [ { "bbox": [ 105, 464, 505, 476 ], "score": 1.0, "content": "at once. Adapting the approach to scale up to many more facts is the subject of other work such", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 475, 350, 487 ], "spans": [ { "bbox": [ 105, 475, 350, 487 ], "score": 1.0, "content": "as Meng, Sen Sharma, Andonian, Belinkov, and Bau (2022).", "type": "text" } ], "index": 33 } ], "index": 30 }, { "type": "text", "bbox": [ 107, 491, 503, 513 ], "lines": [ { "bbox": [ 105, 489, 505, 504 ], "spans": [ { "bbox": [ 105, 489, 505, 504 ], "score": 1.0, "content": "Code, interactive notebooks, dataset, benchmarks, and further visualizations are open-sourced at", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 501, 237, 514 ], "spans": [ { "bbox": [ 105, 501, 237, 514 ], "score": 1.0, "content": "https://rome.baulab.info.", "type": "text" } ], "index": 35 } ], "index": 34.5 }, { "type": "title", "bbox": [ 107, 529, 243, 542 ], "lines": [ { "bbox": [ 105, 527, 244, 544 ], "spans": [ { "bbox": [ 105, 527, 244, 544 ], "score": 1.0, "content": "6 Ethical Considerations", "type": "text" } ], "index": 36 } ], "index": 36 }, { "type": "text", "bbox": [ 107, 555, 505, 626 ], "lines": [ { "bbox": [ 105, 554, 506, 569 ], "spans": [ { "bbox": [ 105, 554, 506, 569 ], "score": 1.0, "content": "By explaining large autoregressive transformer language models’ internal organization and developing", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 565, 505, 578 ], "spans": [ { "bbox": [ 105, 565, 505, 578 ], "score": 1.0, "content": "a fast method for modifying stored knowledge, our work potentially improves the transparency of", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 575, 506, 588 ], "spans": [ { "bbox": [ 106, 575, 506, 588 ], "score": 1.0, "content": "these systems and reduces the energy consumed to correct their errors. However, the capability to", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 583, 507, 599 ], "spans": [ { "bbox": [ 105, 583, 507, 599 ], "score": 1.0, "content": "directly edit large models also has the potential for abuse, such as adding malicious misinformation,", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 595, 506, 608 ], "spans": [ { "bbox": [ 106, 595, 506, 608 ], "score": 1.0, "content": "bias, or other adversarial data to a model. Because of these concerns as well as our observations of", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 605, 506, 618 ], "spans": [ { "bbox": [ 105, 605, 506, 618 ], "score": 1.0, "content": "guessing behavior, we stress that large language models should not be used as an authoritative source", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 614, 268, 628 ], "spans": [ { "bbox": [ 105, 614, 268, 628 ], "score": 1.0, "content": "of factual knowledge in critical settings.", "type": "text" } ], "index": 43 } ], "index": 40 }, { "type": "title", "bbox": [ 108, 641, 207, 655 ], "lines": [ { "bbox": [ 105, 640, 208, 658 ], "spans": [ { "bbox": [ 105, 640, 208, 658 ], "score": 1.0, "content": "Acknowledgements", "type": "text" } ], "index": 44 } ], "index": 44 }, { "type": "text", "bbox": [ 107, 667, 505, 722 ], "lines": [ { "bbox": [ 105, 667, 506, 681 ], "spans": [ { "bbox": [ 105, 667, 506, 681 ], "score": 1.0, "content": "We are grateful to Antonio Torralba, Martin Wattenberg, and Bill Ferguson, whose insightful discussions,", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 678, 505, 690 ], "spans": [ { "bbox": [ 106, 678, 505, 690 ], "score": 1.0, "content": "financial support, and encouragement enabled this project. KM, DB and YB were supported by an AI Alignment", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 689, 506, 701 ], "spans": [ { "bbox": [ 105, 689, 506, 701 ], "score": 1.0, "content": "grant from Open Philanthropy. KM and DB were supported by DARPA SAIL-ON HR0011-20-C-0022 and XAI", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 700, 506, 713 ], "spans": [ { "bbox": [ 105, 700, 506, 713 ], "score": 1.0, "content": "FA8750-18-C-0004. YB was supported by the ISRAEL SCIENCE FOUNDATION (grant No. 448/20) and an", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 710, 297, 725 ], "spans": [ { "bbox": [ 105, 710, 297, 725 ], "score": 1.0, "content": "Azrieli Foundation Early Career Faculty Fellowship.", "type": "text" } ], "index": 49 } ], "index": 47 } ], "page_idx": 9, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 301, 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": "text", "bbox": [ 107, 73, 504, 94 ], "lines": [], "index": 0.5, "bbox_fs": [ 105, 73, 505, 96 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 100, 505, 220 ], "lines": [ { "bbox": [ 105, 99, 505, 112 ], "spans": [ { "bbox": [ 105, 99, 505, 112 ], "score": 1.0, "content": "Another line of work aims to assess the knowledge within LMs by evaluating whether the model", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 110, 506, 124 ], "spans": [ { "bbox": [ 105, 110, 506, 124 ], "score": 1.0, "content": "predict pieces of knowledge. A common strategy is to define a fill-in-the-blank prompt, and let a", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 121, 506, 135 ], "spans": [ { "bbox": [ 105, 121, 506, 135 ], "score": 1.0, "content": "masked LM complete it (Petroni et al., 2019, 2020). Later work showed that knowledge extraction", "type": "text" } ], "index": 4 }, { "bbox": [ 104, 131, 506, 147 ], "spans": [ { "bbox": [ 104, 131, 506, 147 ], "score": 1.0, "content": "can be improved by diversifying the prompts (Jiang et al., 2020; Zhong et al., 2021), or by fine-tuning", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 143, 506, 156 ], "spans": [ { "bbox": [ 105, 143, 506, 156 ], "score": 1.0, "content": "a model on open-domain textual facts (Roberts et al., 2020). However, constructing prompts from", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 154, 506, 168 ], "spans": [ { "bbox": [ 105, 154, 506, 168 ], "score": 1.0, "content": "supervised knowledge extraction data risks learning new knowledge instead of recalling existing", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 164, 506, 178 ], "spans": [ { "bbox": [ 105, 164, 506, 178 ], "score": 1.0, "content": "knowledge in an LM (Zhong et al., 2021). More recently, Elazar et al. (2021a) introduced ParaRel, a", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 177, 506, 188 ], "spans": [ { "bbox": [ 106, 177, 506, 188 ], "score": 1.0, "content": "curated dataset of paraphrased prompts and facts. We use it as a basis for constructing COUNTER-", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 187, 505, 200 ], "spans": [ { "bbox": [ 106, 187, 505, 200 ], "score": 1.0, "content": "FACT, which enables fine-grained measurements of knowledge extraction and editing along multiple", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 199, 506, 210 ], "spans": [ { "bbox": [ 106, 199, 506, 210 ], "score": 1.0, "content": "dimensions. Different from prior work, we do not strive to extract the most knowledge from a model,", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 210, 405, 221 ], "spans": [ { "bbox": [ 106, 210, 405, 221 ], "score": 1.0, "content": "but rather wish to understand mechanisms of knowledge recall in a model.", "type": "text" } ], "index": 12 } ], "index": 7, "bbox_fs": [ 104, 99, 506, 221 ] }, { "type": "text", "bbox": [ 107, 225, 505, 368 ], "lines": [ { "bbox": [ 105, 225, 506, 238 ], "spans": [ { "bbox": [ 105, 225, 506, 238 ], "score": 1.0, "content": "Finally, a few studies aim to localize and modify the computation of knowledge within transformers.", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 235, 506, 249 ], "spans": [ { "bbox": [ 105, 235, 506, 249 ], "score": 1.0, "content": "Geva et al. (2021) identify the MLP layers in a (masked LM) transformer as key–value memories", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 246, 505, 260 ], "spans": [ { "bbox": [ 105, 246, 505, 260 ], "score": 1.0, "content": "of entities and information associated with that entity. Building on this finding, Dai et al. (2022)", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 257, 506, 271 ], "spans": [ { "bbox": [ 105, 257, 506, 271 ], "score": 1.0, "content": "demonstrate a method to edit facts in BERT by writing the embedding of the object into certain rows", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 268, 507, 282 ], "spans": [ { "bbox": [ 105, 268, 507, 282 ], "score": 1.0, "content": "of the MLP matrix. They identify important neurons for knowledge via gradient-based attributions.", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 280, 506, 293 ], "spans": [ { "bbox": [ 106, 280, 506, 293 ], "score": 1.0, "content": "De Cao et al. (2021) train a hyper-network to predict a weight update at test time, which will alter a", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 290, 506, 303 ], "spans": [ { "bbox": [ 106, 290, 506, 303 ], "score": 1.0, "content": "fact. They experiment with BERT and BART (Lewis et al., 2020), a sequence-to-sequence model, and", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 301, 506, 314 ], "spans": [ { "bbox": [ 105, 301, 506, 314 ], "score": 1.0, "content": "focus on models fine-tuned for question answering. Mitchell et al. (2021) presents a hyper-network", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 313, 505, 325 ], "spans": [ { "bbox": [ 106, 313, 505, 325 ], "score": 1.0, "content": "method that learns to transform the decomposed terms of the gradient in order to efficiently predict", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 323, 505, 335 ], "spans": [ { "bbox": [ 105, 323, 505, 335 ], "score": 1.0, "content": "a knowledge update, and demonstrates the ability to scale up to large models including T5 (Raffel", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 334, 506, 347 ], "spans": [ { "bbox": [ 105, 334, 506, 347 ], "score": 1.0, "content": "et al., 2020) and GPT-J (Wang & Komatsuzaki, 2021). We compare with all these methods in our", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 345, 506, 358 ], "spans": [ { "bbox": [ 105, 345, 506, 358 ], "score": 1.0, "content": "experiments, and find that our single-layer ROME parameter intervention has comparable capabilities,", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 357, 397, 369 ], "spans": [ { "bbox": [ 106, 357, 397, 369 ], "score": 1.0, "content": "avoiding failures in specificity and generalization seen in other methods.", "type": "text" } ], "index": 25 } ], "index": 19, "bbox_fs": [ 105, 225, 507, 369 ] }, { "type": "title", "bbox": [ 107, 383, 182, 396 ], "lines": [ { "bbox": [ 104, 381, 185, 399 ], "spans": [ { "bbox": [ 104, 381, 185, 399 ], "score": 1.0, "content": "5 Conclusion", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 409, 505, 486 ], "lines": [ { "bbox": [ 106, 409, 505, 422 ], "spans": [ { "bbox": [ 106, 409, 505, 422 ], "score": 1.0, "content": "We have clarified information flow during knowledge recall in autoregressive transformers, and", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 420, 506, 432 ], "spans": [ { "bbox": [ 105, 420, 506, 432 ], "score": 1.0, "content": "we have exploited this understanding to develop a simple, principled model editor called ROME.", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 432, 506, 444 ], "spans": [ { "bbox": [ 106, 432, 506, 444 ], "score": 1.0, "content": "Our experiments provide insight into how facts are stored and demonstrate the feasibility of direct", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 442, 505, 454 ], "spans": [ { "bbox": [ 105, 442, 505, 454 ], "score": 1.0, "content": "manipulation of computational mechanisms in large pretrained models. While the methods in this", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 453, 506, 465 ], "spans": [ { "bbox": [ 105, 453, 506, 465 ], "score": 1.0, "content": "paper serve to test the locality of knowledge within a model, they apply only to editing a single fact", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 464, 505, 476 ], "spans": [ { "bbox": [ 105, 464, 505, 476 ], "score": 1.0, "content": "at once. Adapting the approach to scale up to many more facts is the subject of other work such", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 475, 350, 487 ], "spans": [ { "bbox": [ 105, 475, 350, 487 ], "score": 1.0, "content": "as Meng, Sen Sharma, Andonian, Belinkov, and Bau (2022).", "type": "text" } ], "index": 33 } ], "index": 30, "bbox_fs": [ 105, 409, 506, 487 ] }, { "type": "text", "bbox": [ 107, 491, 503, 513 ], "lines": [ { "bbox": [ 105, 489, 505, 504 ], "spans": [ { "bbox": [ 105, 489, 505, 504 ], "score": 1.0, "content": "Code, interactive notebooks, dataset, benchmarks, and further visualizations are open-sourced at", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 501, 237, 514 ], "spans": [ { "bbox": [ 105, 501, 237, 514 ], "score": 1.0, "content": "https://rome.baulab.info.", "type": "text" } ], "index": 35 } ], "index": 34.5, "bbox_fs": [ 105, 489, 505, 514 ] }, { "type": "title", "bbox": [ 107, 529, 243, 542 ], "lines": [ { "bbox": [ 105, 527, 244, 544 ], "spans": [ { "bbox": [ 105, 527, 244, 544 ], "score": 1.0, "content": "6 Ethical Considerations", "type": "text" } ], "index": 36 } ], "index": 36 }, { "type": "text", "bbox": [ 107, 555, 505, 626 ], "lines": [ { "bbox": [ 105, 554, 506, 569 ], "spans": [ { "bbox": [ 105, 554, 506, 569 ], "score": 1.0, "content": "By explaining large autoregressive transformer language models’ internal organization and developing", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 565, 505, 578 ], "spans": [ { "bbox": [ 105, 565, 505, 578 ], "score": 1.0, "content": "a fast method for modifying stored knowledge, our work potentially improves the transparency of", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 575, 506, 588 ], "spans": [ { "bbox": [ 106, 575, 506, 588 ], "score": 1.0, "content": "these systems and reduces the energy consumed to correct their errors. However, the capability to", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 583, 507, 599 ], "spans": [ { "bbox": [ 105, 583, 507, 599 ], "score": 1.0, "content": "directly edit large models also has the potential for abuse, such as adding malicious misinformation,", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 595, 506, 608 ], "spans": [ { "bbox": [ 106, 595, 506, 608 ], "score": 1.0, "content": "bias, or other adversarial data to a model. Because of these concerns as well as our observations of", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 605, 506, 618 ], "spans": [ { "bbox": [ 105, 605, 506, 618 ], "score": 1.0, "content": "guessing behavior, we stress that large language models should not be used as an authoritative source", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 614, 268, 628 ], "spans": [ { "bbox": [ 105, 614, 268, 628 ], "score": 1.0, "content": "of factual knowledge in critical settings.", "type": "text" } ], "index": 43 } ], "index": 40, "bbox_fs": [ 105, 554, 507, 628 ] }, { "type": "title", "bbox": [ 108, 641, 207, 655 ], "lines": [ { "bbox": [ 105, 640, 208, 658 ], "spans": [ { "bbox": [ 105, 640, 208, 658 ], "score": 1.0, "content": "Acknowledgements", "type": "text" } ], "index": 44 } ], "index": 44 }, { "type": "text", "bbox": [ 107, 667, 505, 722 ], "lines": [ { "bbox": [ 105, 667, 506, 681 ], "spans": [ { "bbox": [ 105, 667, 506, 681 ], "score": 1.0, "content": "We are grateful to Antonio Torralba, Martin Wattenberg, and Bill Ferguson, whose insightful discussions,", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 678, 505, 690 ], "spans": [ { "bbox": [ 106, 678, 505, 690 ], "score": 1.0, "content": "financial support, and encouragement enabled this project. KM, DB and YB were supported by an AI Alignment", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 689, 506, 701 ], "spans": [ { "bbox": [ 105, 689, 506, 701 ], "score": 1.0, "content": "grant from Open Philanthropy. KM and DB were supported by DARPA SAIL-ON HR0011-20-C-0022 and XAI", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 700, 506, 713 ], "spans": [ { "bbox": [ 105, 700, 506, 713 ], "score": 1.0, "content": "FA8750-18-C-0004. YB was supported by the ISRAEL SCIENCE FOUNDATION (grant No. 448/20) and an", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 710, 297, 725 ], "spans": [ { "bbox": [ 105, 710, 297, 725 ], "score": 1.0, "content": "Azrieli Foundation Early Career Faculty Fellowship.", "type": "text" } ], "index": 49 } ], "index": 47, "bbox_fs": [ 105, 667, 506, 725 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 107, 72, 164, 84 ], "lines": [ { "bbox": [ 107, 72, 164, 84 ], "spans": [], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 104, 75, 507, 730 ], "lines": [ { "bbox": [ 106, 70, 165, 86 ], "spans": [ { "bbox": [ 106, 70, 165, 86 ], "score": 1.0, "content": "References", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 97, 507, 114 ], "spans": [ { "bbox": [ 105, 97, 507, 114 ], "score": 1.0, "content": "Adi, Y., Kermany, E., Belinkov, Y., Lavi, O., and Goldberg, Y. Fine-grained analysis of sentence em-", "type": "text" } ], "index": 2 }, { "bbox": [ 115, 110, 506, 124 ], "spans": [ { "bbox": [ 115, 110, 506, 124 ], "score": 1.0, "content": "beddings using auxiliary prediction tasks. In International Conference on Learning Representations", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 121, 198, 135 ], "spans": [ { "bbox": [ 115, 121, 198, 135 ], "score": 1.0, "content": "(ICLR), April 2017.", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 141, 507, 155 ], "spans": [ { "bbox": [ 105, 141, 507, 155 ], "score": 1.0, "content": "Anderson, J. A. A simple neural network generating an interactive memory. Mathematical biosciences,", "type": "text" } ], "index": 5 }, { "bbox": [ 114, 151, 213, 165 ], "spans": [ { "bbox": [ 114, 151, 213, 165 ], "score": 1.0, "content": "14(3-4):197–220, 1972.", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 171, 506, 185 ], "spans": [ { "bbox": [ 105, 171, 506, 185 ], "score": 1.0, "content": "Bau, D., Liu, S., Wang, T., Zhu, J.-Y., and Torralba, A. Rewriting a deep generative model. In", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 183, 426, 196 ], "spans": [ { "bbox": [ 115, 183, 426, 196 ], "score": 1.0, "content": "Proceedings of the European Conference on Computer Vision (ECCV), 2020.", "type": "text" } ], "index": 8 }, { "bbox": [ 104, 201, 507, 217 ], "spans": [ { "bbox": [ 104, 201, 507, 217 ], "score": 1.0, "content": "Belinkov, Y. Probing Classifiers: Promises, Shortcomings, and Advances. Computational Linguistics,", "type": "text" } ], "index": 9 }, { "bbox": [ 113, 212, 509, 228 ], "spans": [ { "bbox": [ 113, 212, 509, 228 ], "score": 1.0, "content": "pp. 1–13, 11 2021. ISSN 0891-2017. doi: 10.1162/coli a 00422. URL https://doi.org/10.", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 224, 210, 237 ], "spans": [ { "bbox": [ 116, 224, 210, 237 ], "score": 1.0, "content": "1162/coli_a_00422.", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 244, 507, 257 ], "spans": [ { "bbox": [ 105, 244, 507, 257 ], "score": 1.0, "content": "Belinkov, Y. and Glass, J. Analysis methods in neural language processing: A survey. Transactions of", "type": "text" } ], "index": 12 }, { "bbox": [ 115, 254, 507, 268 ], "spans": [ { "bbox": [ 115, 254, 507, 268 ], "score": 1.0, "content": "the Association for Computational Linguistics, 7:49–72, March 2019. doi: 10.1162/tacl a 00254.", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 265, 318, 281 ], "spans": [ { "bbox": [ 115, 265, 318, 281 ], "score": 1.0, "content": "URL https://aclanthology.org/Q19-1004.", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 286, 505, 298 ], "spans": [ { "bbox": [ 106, 286, 505, 298 ], "score": 1.0, "content": "Belinkov, Y., Durrani, N., Dalvi, F., Sajjad, H., and Glass, J. What do neural machine translation", "type": "text" } ], "index": 15 }, { "bbox": [ 115, 296, 506, 310 ], "spans": [ { "bbox": [ 115, 296, 506, 310 ], "score": 1.0, "content": "models learn about morphology? In Proceedings of the 55th Annual Meeting of the Association", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 308, 505, 321 ], "spans": [ { "bbox": [ 115, 308, 505, 321 ], "score": 1.0, "content": "for Computational Linguistics (Volume 1: Long Papers), pp. 861–872, Vancouver, Canada, July", "type": "text" } ], "index": 17 }, { "bbox": [ 114, 316, 507, 334 ], "spans": [ { "bbox": [ 114, 316, 507, 334 ], "score": 1.0, "content": "2017. Association for Computational Linguistics. doi: 10.18653/v1/P17-1080. URL https:", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 329, 263, 343 ], "spans": [ { "bbox": [ 115, 329, 263, 343 ], "score": 1.0, "content": "//aclanthology.org/P17-1080.", "type": "text" } ], "index": 19 }, { "bbox": [ 104, 347, 507, 364 ], "spans": [ { "bbox": [ 104, 347, 507, 364 ], "score": 1.0, "content": "Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam,", "type": "text" } ], "index": 20 }, { "bbox": [ 115, 360, 506, 374 ], "spans": [ { "bbox": [ 115, 360, 506, 374 ], "score": 1.0, "content": "P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R.,", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 371, 507, 385 ], "spans": [ { "bbox": [ 115, 371, 507, 385 ], "score": 1.0, "content": "Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray,", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 383, 506, 394 ], "spans": [ { "bbox": [ 116, 383, 506, 394 ], "score": 1.0, "content": "S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., and Amodei, D.", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 393, 507, 406 ], "spans": [ { "bbox": [ 115, 393, 507, 406 ], "score": 1.0, "content": "Language models are few-shot learners. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan,", "type": "text" } ], "index": 24 }, { "bbox": [ 113, 401, 507, 419 ], "spans": [ { "bbox": [ 113, 401, 507, 419 ], "score": 1.0, "content": "M. F., and Lin, H. (eds.), Advances in Neural Information Processing Systems, volume 33, pp.", "type": "text" } ], "index": 25 }, { "bbox": [ 113, 413, 505, 429 ], "spans": [ { "bbox": [ 113, 413, 505, 429 ], "score": 1.0, "content": "1877–1901. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper/", "type": "text" } ], "index": 26 }, { "bbox": [ 116, 426, 392, 439 ], "spans": [ { "bbox": [ 116, 426, 392, 439 ], "score": 1.0, "content": "2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf.", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 446, 505, 457 ], "spans": [ { "bbox": [ 106, 446, 505, 457 ], "score": 1.0, "content": "Conneau, A., Kruszewski, G., Lample, G., Barrault, L., and Baroni, M. What you can cram into a", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 455, 505, 471 ], "spans": [ { "bbox": [ 115, 455, 505, 471 ], "score": 1.0, "content": "single $&!#* vector: Probing sentence embeddings for linguistic properties. In Proceedings of the", "type": "text" } ], "index": 29 }, { "bbox": [ 115, 468, 506, 481 ], "spans": [ { "bbox": [ 115, 468, 506, 481 ], "score": 1.0, "content": "56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),", "type": "text" } ], "index": 30 }, { "bbox": [ 113, 476, 508, 493 ], "spans": [ { "bbox": [ 113, 476, 508, 493 ], "score": 1.0, "content": "pp. 2126–2136, Melbourne, Australia, July 2018. Association for Computational Linguistics. doi:", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 489, 415, 503 ], "spans": [ { "bbox": [ 115, 489, 415, 503 ], "score": 1.0, "content": "10.18653/v1/P18-1198. URL https://aclanthology.org/P18-1198.", "type": "text" } ], "index": 32 }, { "bbox": [ 104, 508, 506, 523 ], "spans": [ { "bbox": [ 104, 508, 506, 523 ], "score": 1.0, "content": "Dai, D., Dong, L., Hao, Y., Sui, Z., Chang, B., and Wei, F. Knowledge neurons in pretrained", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 520, 506, 534 ], "spans": [ { "bbox": [ 115, 520, 506, 534 ], "score": 1.0, "content": "transformers. In Proceedings of the 60th Annual Meeting of the Association for Computational", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 531, 358, 545 ], "spans": [ { "bbox": [ 115, 531, 358, 545 ], "score": 1.0, "content": "Linguistics (Volume 1: Long Papers), pp. 8493–8502, 2022.", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 550, 505, 564 ], "spans": [ { "bbox": [ 105, 550, 505, 564 ], "score": 1.0, "content": "De Cao, N., Aziz, W., and Titov, I. Editing factual knowledge in language models. In Proceedings", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 561, 505, 575 ], "spans": [ { "bbox": [ 115, 561, 505, 575 ], "score": 1.0, "content": "of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 6491–6506,", "type": "text" } ], "index": 37 }, { "bbox": [ 116, 573, 505, 585 ], "spans": [ { "bbox": [ 116, 573, 505, 585 ], "score": 1.0, "content": "Online and Punta Cana, Dominican Republic, November 2021. Association for Computational", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 583, 424, 597 ], "spans": [ { "bbox": [ 115, 583, 424, 597 ], "score": 1.0, "content": "Linguistics. URL https://aclanthology.org/2021.emnlp-main.522.", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 603, 506, 616 ], "spans": [ { "bbox": [ 105, 603, 506, 616 ], "score": 1.0, "content": "Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. BERT: Pre-training of deep bidirec-", "type": "text" } ], "index": 40 }, { "bbox": [ 115, 614, 506, 628 ], "spans": [ { "bbox": [ 115, 614, 506, 628 ], "score": 1.0, "content": "tional transformers for language understanding. In Proceedings of the 2019 Conference of", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 625, 506, 639 ], "spans": [ { "bbox": [ 115, 625, 506, 639 ], "score": 1.0, "content": "the North American Chapter of the Association for Computational Linguistics: Human Lan-", "type": "text" } ], "index": 42 }, { "bbox": [ 115, 635, 506, 651 ], "spans": [ { "bbox": [ 115, 635, 506, 651 ], "score": 1.0, "content": "guage Technologies, Volume 1 (Long and Short Papers), pp. 4171–4186, Minneapolis, Min-", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 646, 505, 660 ], "spans": [ { "bbox": [ 115, 646, 505, 660 ], "score": 1.0, "content": "nesota, June 2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1423. URL", "type": "text" } ], "index": 44 }, { "bbox": [ 114, 657, 295, 672 ], "spans": [ { "bbox": [ 114, 657, 295, 672 ], "score": 1.0, "content": "https://aclanthology.org/N19-1423.", "type": "text" } ], "index": 45 }, { "bbox": [ 106, 678, 506, 690 ], "spans": [ { "bbox": [ 106, 678, 506, 690 ], "score": 1.0, "content": "Elazar, Y., Kassner, N., Ravfogel, S., Ravichander, A., Hovy, E., Schutze, H., and Goldberg, Y. ¨", "type": "text" } ], "index": 46 }, { "bbox": [ 115, 688, 506, 702 ], "spans": [ { "bbox": [ 115, 688, 506, 702 ], "score": 1.0, "content": "Measuring and Improving Consistency in Pretrained Language Models. Transactions of the", "type": "text" } ], "index": 47 }, { "bbox": [ 114, 699, 507, 713 ], "spans": [ { "bbox": [ 114, 699, 507, 713 ], "score": 1.0, "content": "Association for Computational Linguistics, 9:1012–1031, 09 2021a. ISSN 2307-387X. doi:", "type": "text" } ], "index": 48 }, { "bbox": [ 115, 709, 424, 725 ], "spans": [ { "bbox": [ 115, 709, 424, 725 ], "score": 1.0, "content": "10.1162/tacl a 00410. URL https://doi.org/10.1162/tacl_a_00410.", "type": "text" } ], "index": 49 } ], "index": 25 } ], "page_idx": 10, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 741, 310, 750 ], "lines": [ { "bbox": [ 299, 740, 312, 755 ], "spans": [ { "bbox": [ 299, 740, 312, 755 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 107, 72, 164, 84 ], "lines": [ { "bbox": [ 107, 72, 164, 84 ], "spans": [], "index": 0 } ], "index": 0 }, { "type": "list", "bbox": [ 104, 75, 507, 730 ], "lines": [ { "bbox": [ 106, 70, 165, 86 ], "spans": [ { "bbox": [ 106, 70, 165, 86 ], "score": 1.0, "content": "References", "type": "text" } ], "index": 1, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 105, 97, 507, 114 ], "spans": [ { "bbox": [ 105, 97, 507, 114 ], "score": 1.0, "content": "Adi, Y., Kermany, E., Belinkov, Y., Lavi, O., and Goldberg, Y. Fine-grained analysis of sentence em-", "type": "text" } ], "index": 2, "is_list_start_line": true }, { "bbox": [ 115, 110, 506, 124 ], "spans": [ { "bbox": [ 115, 110, 506, 124 ], "score": 1.0, "content": "beddings using auxiliary prediction tasks. In International Conference on Learning Representations", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 121, 198, 135 ], "spans": [ { "bbox": [ 115, 121, 198, 135 ], "score": 1.0, "content": "(ICLR), April 2017.", "type": "text" } ], "index": 4, "is_list_end_line": true }, { "bbox": [ 105, 141, 507, 155 ], "spans": [ { "bbox": [ 105, 141, 507, 155 ], "score": 1.0, "content": "Anderson, J. A. A simple neural network generating an interactive memory. Mathematical biosciences,", "type": "text" } ], "index": 5, "is_list_start_line": true }, { "bbox": [ 114, 151, 213, 165 ], "spans": [ { "bbox": [ 114, 151, 213, 165 ], "score": 1.0, "content": "14(3-4):197–220, 1972.", "type": "text" } ], "index": 6, "is_list_end_line": true }, { "bbox": [ 105, 171, 506, 185 ], "spans": [ { "bbox": [ 105, 171, 506, 185 ], "score": 1.0, "content": "Bau, D., Liu, S., Wang, T., Zhu, J.-Y., and Torralba, A. Rewriting a deep generative model. In", "type": "text" } ], "index": 7, "is_list_start_line": true }, { "bbox": [ 115, 183, 426, 196 ], "spans": [ { "bbox": [ 115, 183, 426, 196 ], "score": 1.0, "content": "Proceedings of the European Conference on Computer Vision (ECCV), 2020.", "type": "text" } ], "index": 8, "is_list_end_line": true }, { "bbox": [ 104, 201, 507, 217 ], "spans": [ { "bbox": [ 104, 201, 507, 217 ], "score": 1.0, "content": "Belinkov, Y. Probing Classifiers: Promises, Shortcomings, and Advances. Computational Linguistics,", "type": "text" } ], "index": 9, "is_list_start_line": true }, { "bbox": [ 113, 212, 509, 228 ], "spans": [ { "bbox": [ 113, 212, 509, 228 ], "score": 1.0, "content": "pp. 1–13, 11 2021. ISSN 0891-2017. doi: 10.1162/coli a 00422. URL https://doi.org/10.", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 224, 210, 237 ], "spans": [ { "bbox": [ 116, 224, 210, 237 ], "score": 1.0, "content": "1162/coli_a_00422.", "type": "text" } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 105, 244, 507, 257 ], "spans": [ { "bbox": [ 105, 244, 507, 257 ], "score": 1.0, "content": "Belinkov, Y. and Glass, J. Analysis methods in neural language processing: A survey. Transactions of", "type": "text" } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 115, 254, 507, 268 ], "spans": [ { "bbox": [ 115, 254, 507, 268 ], "score": 1.0, "content": "the Association for Computational Linguistics, 7:49–72, March 2019. doi: 10.1162/tacl a 00254.", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 265, 318, 281 ], "spans": [ { "bbox": [ 115, 265, 318, 281 ], "score": 1.0, "content": "URL https://aclanthology.org/Q19-1004.", "type": "text" } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 106, 286, 505, 298 ], "spans": [ { "bbox": [ 106, 286, 505, 298 ], "score": 1.0, "content": "Belinkov, Y., Durrani, N., Dalvi, F., Sajjad, H., and Glass, J. What do neural machine translation", "type": "text" } ], "index": 15, "is_list_start_line": true }, { "bbox": [ 115, 296, 506, 310 ], "spans": [ { "bbox": [ 115, 296, 506, 310 ], "score": 1.0, "content": "models learn about morphology? In Proceedings of the 55th Annual Meeting of the Association", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 308, 505, 321 ], "spans": [ { "bbox": [ 115, 308, 505, 321 ], "score": 1.0, "content": "for Computational Linguistics (Volume 1: Long Papers), pp. 861–872, Vancouver, Canada, July", "type": "text" } ], "index": 17 }, { "bbox": [ 114, 316, 507, 334 ], "spans": [ { "bbox": [ 114, 316, 507, 334 ], "score": 1.0, "content": "2017. Association for Computational Linguistics. doi: 10.18653/v1/P17-1080. URL https:", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 329, 263, 343 ], "spans": [ { "bbox": [ 115, 329, 263, 343 ], "score": 1.0, "content": "//aclanthology.org/P17-1080.", "type": "text" } ], "index": 19, "is_list_end_line": true }, { "bbox": [ 104, 347, 507, 364 ], "spans": [ { "bbox": [ 104, 347, 507, 364 ], "score": 1.0, "content": "Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam,", "type": "text" } ], "index": 20, "is_list_start_line": true }, { "bbox": [ 115, 360, 506, 374 ], "spans": [ { "bbox": [ 115, 360, 506, 374 ], "score": 1.0, "content": "P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R.,", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 371, 507, 385 ], "spans": [ { "bbox": [ 115, 371, 507, 385 ], "score": 1.0, "content": "Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray,", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 383, 506, 394 ], "spans": [ { "bbox": [ 116, 383, 506, 394 ], "score": 1.0, "content": "S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., and Amodei, D.", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 393, 507, 406 ], "spans": [ { "bbox": [ 115, 393, 507, 406 ], "score": 1.0, "content": "Language models are few-shot learners. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan,", "type": "text" } ], "index": 24 }, { "bbox": [ 113, 401, 507, 419 ], "spans": [ { "bbox": [ 113, 401, 507, 419 ], "score": 1.0, "content": "M. F., and Lin, H. (eds.), Advances in Neural Information Processing Systems, volume 33, pp.", "type": "text" } ], "index": 25 }, { "bbox": [ 113, 413, 505, 429 ], "spans": [ { "bbox": [ 113, 413, 505, 429 ], "score": 1.0, "content": "1877–1901. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper/", "type": "text" } ], "index": 26 }, { "bbox": [ 116, 426, 392, 439 ], "spans": [ { "bbox": [ 116, 426, 392, 439 ], "score": 1.0, "content": "2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf.", "type": "text" } ], "index": 27, "is_list_end_line": true }, { "bbox": [ 106, 446, 505, 457 ], "spans": [ { "bbox": [ 106, 446, 505, 457 ], "score": 1.0, "content": "Conneau, A., Kruszewski, G., Lample, G., Barrault, L., and Baroni, M. What you can cram into a", "type": "text" } ], "index": 28, "is_list_start_line": true }, { "bbox": [ 115, 455, 505, 471 ], "spans": [ { "bbox": [ 115, 455, 505, 471 ], "score": 1.0, "content": "single $&!#* vector: Probing sentence embeddings for linguistic properties. In Proceedings of the", "type": "text" } ], "index": 29 }, { "bbox": [ 115, 468, 506, 481 ], "spans": [ { "bbox": [ 115, 468, 506, 481 ], "score": 1.0, "content": "56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),", "type": "text" } ], "index": 30 }, { "bbox": [ 113, 476, 508, 493 ], "spans": [ { "bbox": [ 113, 476, 508, 493 ], "score": 1.0, "content": "pp. 2126–2136, Melbourne, Australia, July 2018. Association for Computational Linguistics. doi:", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 489, 415, 503 ], "spans": [ { "bbox": [ 115, 489, 415, 503 ], "score": 1.0, "content": "10.18653/v1/P18-1198. URL https://aclanthology.org/P18-1198.", "type": "text" } ], "index": 32, "is_list_end_line": true }, { "bbox": [ 104, 508, 506, 523 ], "spans": [ { "bbox": [ 104, 508, 506, 523 ], "score": 1.0, "content": "Dai, D., Dong, L., Hao, Y., Sui, Z., Chang, B., and Wei, F. Knowledge neurons in pretrained", "type": "text" } ], "index": 33, "is_list_start_line": true }, { "bbox": [ 115, 520, 506, 534 ], "spans": [ { "bbox": [ 115, 520, 506, 534 ], "score": 1.0, "content": "transformers. In Proceedings of the 60th Annual Meeting of the Association for Computational", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 531, 358, 545 ], "spans": [ { "bbox": [ 115, 531, 358, 545 ], "score": 1.0, "content": "Linguistics (Volume 1: Long Papers), pp. 8493–8502, 2022.", "type": "text" } ], "index": 35, "is_list_end_line": true }, { "bbox": [ 105, 550, 505, 564 ], "spans": [ { "bbox": [ 105, 550, 505, 564 ], "score": 1.0, "content": "De Cao, N., Aziz, W., and Titov, I. Editing factual knowledge in language models. In Proceedings", "type": "text" } ], "index": 36, "is_list_start_line": true }, { "bbox": [ 115, 561, 505, 575 ], "spans": [ { "bbox": [ 115, 561, 505, 575 ], "score": 1.0, "content": "of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 6491–6506,", "type": "text" } ], "index": 37 }, { "bbox": [ 116, 573, 505, 585 ], "spans": [ { "bbox": [ 116, 573, 505, 585 ], "score": 1.0, "content": "Online and Punta Cana, Dominican Republic, November 2021. Association for Computational", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 583, 424, 597 ], "spans": [ { "bbox": [ 115, 583, 424, 597 ], "score": 1.0, "content": "Linguistics. URL https://aclanthology.org/2021.emnlp-main.522.", "type": "text" } ], "index": 39, "is_list_end_line": true }, { "bbox": [ 105, 603, 506, 616 ], "spans": [ { "bbox": [ 105, 603, 506, 616 ], "score": 1.0, "content": "Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. BERT: Pre-training of deep bidirec-", "type": "text" } ], "index": 40, "is_list_start_line": true }, { "bbox": [ 115, 614, 506, 628 ], "spans": [ { "bbox": [ 115, 614, 506, 628 ], "score": 1.0, "content": "tional transformers for language understanding. In Proceedings of the 2019 Conference of", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 625, 506, 639 ], "spans": [ { "bbox": [ 115, 625, 506, 639 ], "score": 1.0, "content": "the North American Chapter of the Association for Computational Linguistics: Human Lan-", "type": "text" } ], "index": 42 }, { "bbox": [ 115, 635, 506, 651 ], "spans": [ { "bbox": [ 115, 635, 506, 651 ], "score": 1.0, "content": "guage Technologies, Volume 1 (Long and Short Papers), pp. 4171–4186, Minneapolis, Min-", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 646, 505, 660 ], "spans": [ { "bbox": [ 115, 646, 505, 660 ], "score": 1.0, "content": "nesota, June 2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1423. URL", "type": "text" } ], "index": 44 }, { "bbox": [ 114, 657, 295, 672 ], "spans": [ { "bbox": [ 114, 657, 295, 672 ], "score": 1.0, "content": "https://aclanthology.org/N19-1423.", "type": "text" } ], "index": 45, "is_list_end_line": true }, { "bbox": [ 106, 678, 506, 690 ], "spans": [ { "bbox": [ 106, 678, 506, 690 ], "score": 1.0, "content": "Elazar, Y., Kassner, N., Ravfogel, S., Ravichander, A., Hovy, E., Schutze, H., and Goldberg, Y. ¨", "type": "text" } ], "index": 46, "is_list_start_line": true }, { "bbox": [ 115, 688, 506, 702 ], "spans": [ { "bbox": [ 115, 688, 506, 702 ], "score": 1.0, "content": "Measuring and Improving Consistency in Pretrained Language Models. Transactions of the", "type": "text" } ], "index": 47 }, { "bbox": [ 114, 699, 507, 713 ], "spans": [ { "bbox": [ 114, 699, 507, 713 ], "score": 1.0, "content": "Association for Computational Linguistics, 9:1012–1031, 09 2021a. ISSN 2307-387X. doi:", "type": "text" } ], "index": 48 }, { "bbox": [ 115, 709, 424, 725 ], "spans": [ { "bbox": [ 115, 709, 424, 725 ], "score": 1.0, "content": "10.1162/tacl a 00410. URL https://doi.org/10.1162/tacl_a_00410.", "type": "text" } ], "index": 49, "is_list_end_line": true }, { "bbox": [ 105, 72, 506, 85 ], "spans": [ { "bbox": [ 105, 72, 506, 85 ], "score": 1.0, "content": "Elazar, Y., Ravfogel, S., Jacovi, A., and Goldberg, Y. Amnesic probing: Behavioral explanation", "type": "text", "cross_page": true } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 115, 83, 507, 97 ], "spans": [ { "bbox": [ 115, 83, 507, 97 ], "score": 1.0, "content": "with amnesic counterfactuals. Transactions of the Association for Computational Linguistics, 9:", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 116, 95, 185, 106 ], "spans": [ { "bbox": [ 116, 95, 185, 106 ], "score": 1.0, "content": "160–175, 2021b.", "type": "text", "cross_page": true } ], "index": 2, "is_list_end_line": true }, { "bbox": [ 106, 113, 506, 126 ], "spans": [ { "bbox": [ 106, 113, 506, 126 ], "score": 1.0, "content": "Elhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., Askell, A., Bai, Y., Chen,", "type": "text", "cross_page": true } ], "index": 3, "is_list_start_line": true }, { "bbox": [ 114, 123, 506, 137 ], "spans": [ { "bbox": [ 114, 123, 506, 137 ], "score": 1.0, "content": "A., Conerly, T., DasSarma, N., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D.,", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 116, 136, 506, 147 ], "spans": [ { "bbox": [ 116, 136, 506, 147 ], "score": 1.0, "content": "Jones, A., Kernion, J., Lovitt, L., Ndousse, K., Amodei, D., Brown, T., Clark, J., Kaplan, J.,", "type": "text", "cross_page": true } ], "index": 5 }, { "bbox": [ 114, 144, 508, 159 ], "spans": [ { "bbox": [ 114, 144, 508, 159 ], "score": 1.0, "content": "McCandlish, S., and Olah, C. A mathematical framework for transformer circuits. https:", "type": "text", "cross_page": true } ], "index": 6 }, { "bbox": [ 116, 156, 461, 170 ], "spans": [ { "bbox": [ 116, 156, 461, 170 ], "score": 1.0, "content": "//transformer-circuits.pub/2021/framework/index.html, December 2021.", "type": "text", "cross_page": true } ], "index": 7, "is_list_end_line": true }, { "bbox": [ 106, 176, 506, 187 ], "spans": [ { "bbox": [ 106, 176, 506, 187 ], "score": 1.0, "content": "Ettinger, A., Elgohary, A., and Resnik, P. Probing for semantic evidence of composition by means of", "type": "text", "cross_page": true } ], "index": 8, "is_list_start_line": true }, { "bbox": [ 116, 186, 507, 200 ], "spans": [ { "bbox": [ 116, 186, 507, 200 ], "score": 1.0, "content": "simple classification tasks. In Proceedings of the 1st Workshop on Evaluating Vector-Space Repre-", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 114, 196, 506, 211 ], "spans": [ { "bbox": [ 114, 196, 506, 211 ], "score": 1.0, "content": "sentations for NLP, pp. 134–139, Berlin, Germany, August 2016. Association for Computational", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 114, 206, 488, 223 ], "spans": [ { "bbox": [ 114, 206, 488, 223 ], "score": 1.0, "content": "Linguistics. doi: 10.18653/v1/W16-2524. URL https://aclanthology.org/W16-2524.", "type": "text", "cross_page": true } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 104, 225, 506, 241 ], "spans": [ { "bbox": [ 104, 225, 506, 241 ], "score": 1.0, "content": "Feder, A., Oved, N., Shalit, U., and Reichart, R. CausaLM: Causal model explanation through", "type": "text", "cross_page": true } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 114, 238, 448, 251 ], "spans": [ { "bbox": [ 114, 238, 448, 251 ], "score": 1.0, "content": "counterfactual language models. Computational Linguistics, 47(2):333–386, 2021.", "type": "text", "cross_page": true } ], "index": 13, "is_list_end_line": true }, { "bbox": [ 105, 256, 506, 270 ], "spans": [ { "bbox": [ 105, 256, 506, 270 ], "score": 1.0, "content": "Finlayson, M., Mueller, A., Gehrmann, S., Shieber, S., Linzen, T., and Belinkov, Y. Causal analysis", "type": "text", "cross_page": true } ], "index": 14, "is_list_start_line": true }, { "bbox": [ 115, 267, 505, 280 ], "spans": [ { "bbox": [ 115, 267, 505, 280 ], "score": 1.0, "content": "of syntactic agreement mechanisms in neural language models. In Proceedings of the 59th", "type": "text", "cross_page": true } ], "index": 15 }, { "bbox": [ 115, 280, 505, 291 ], "spans": [ { "bbox": [ 115, 280, 505, 291 ], "score": 1.0, "content": "Annual Meeting of the Association for Computational Linguistics and the 11th International Joint", "type": "text", "cross_page": true } ], "index": 16 }, { "bbox": [ 116, 289, 506, 303 ], "spans": [ { "bbox": [ 116, 289, 506, 303 ], "score": 1.0, "content": "Conference on Natural Language Processing (Volume 1: Long Papers), pp. 1828–1843, Online,", "type": "text", "cross_page": true } ], "index": 17 }, { "bbox": [ 115, 300, 507, 314 ], "spans": [ { "bbox": [ 115, 300, 507, 314 ], "score": 1.0, "content": "August 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.acl-long.144.", "type": "text", "cross_page": true } ], "index": 18 }, { "bbox": [ 116, 311, 364, 325 ], "spans": [ { "bbox": [ 116, 311, 364, 325 ], "score": 1.0, "content": "URL https://aclanthology.org/2021.acl-long.144.", "type": "text", "cross_page": true } ], "index": 19, "is_list_end_line": true }, { "bbox": [ 105, 329, 507, 344 ], "spans": [ { "bbox": [ 105, 329, 507, 344 ], "score": 1.0, "content": "Geva, M., Schuster, R., Berant, J., and Levy, O. Transformer feed-forward layers are key-value memo-", "type": "text", "cross_page": true } ], "index": 20, "is_list_start_line": true }, { "bbox": [ 114, 340, 506, 355 ], "spans": [ { "bbox": [ 114, 340, 506, 355 ], "score": 1.0, "content": "ries. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,", "type": "text", "cross_page": true } ], "index": 21 }, { "bbox": [ 114, 351, 507, 366 ], "spans": [ { "bbox": [ 114, 351, 507, 366 ], "score": 1.0, "content": "pp. 5484–5495, Online and Punta Cana, Dominican Republic, November 2021. Association for", "type": "text", "cross_page": true } ], "index": 22 }, { "bbox": [ 115, 363, 487, 377 ], "spans": [ { "bbox": [ 115, 363, 487, 377 ], "score": 1.0, "content": "Computational Linguistics. URL https://aclanthology.org/2021.emnlp-main.446.", "type": "text", "cross_page": true } ], "index": 23, "is_list_end_line": true }, { "bbox": [ 105, 381, 506, 395 ], "spans": [ { "bbox": [ 105, 381, 506, 395 ], "score": 1.0, "content": "Hase, P., Diab, M., Celikyilmaz, A., Li, X., Kozareva, Z., Stoyanov, V., Bansal, M., and Iyer, S. Do", "type": "text", "cross_page": true } ], "index": 24, "is_list_start_line": true }, { "bbox": [ 115, 393, 506, 406 ], "spans": [ { "bbox": [ 115, 393, 506, 406 ], "score": 1.0, "content": "language models have beliefs? methods for detecting, updating, and visualizing model beliefs.", "type": "text", "cross_page": true } ], "index": 25 }, { "bbox": [ 115, 403, 281, 417 ], "spans": [ { "bbox": [ 115, 403, 281, 417 ], "score": 1.0, "content": "arXiv preprint arXiv:2111.13654, 2021.", "type": "text", "cross_page": true } ], "index": 26, "is_list_end_line": true }, { "bbox": [ 104, 421, 506, 436 ], "spans": [ { "bbox": [ 104, 421, 506, 436 ], "score": 1.0, "content": "Hupkes, D., Veldhoen, S., and Zuidema, W. Visualisation and ’diagnostic classifiers’ reveal how", "type": "text", "cross_page": true } ], "index": 27, "is_list_start_line": true }, { "bbox": [ 115, 433, 506, 447 ], "spans": [ { "bbox": [ 115, 433, 506, 447 ], "score": 1.0, "content": "recurrent and recursive neural networks process hierarchical structure. Journal of Artificial", "type": "text", "cross_page": true } ], "index": 28 }, { "bbox": [ 114, 444, 286, 457 ], "spans": [ { "bbox": [ 114, 444, 286, 457 ], "score": 1.0, "content": "Intelligence Research, 61:907–926, 2018.", "type": "text", "cross_page": true } ], "index": 29, "is_list_end_line": true }, { "bbox": [ 104, 462, 506, 478 ], "spans": [ { "bbox": [ 104, 462, 506, 478 ], "score": 1.0, "content": "Jiang, Z., Xu, F. F., Araki, J., and Neubig, G. How can we know what language models know?", "type": "text", "cross_page": true } ], "index": 30, "is_list_start_line": true }, { "bbox": [ 116, 473, 506, 487 ], "spans": [ { "bbox": [ 116, 473, 506, 487 ], "score": 1.0, "content": "Transactions of the Association for Computational Linguistics, 8:423–438, 2020. doi: 10.1162/", "type": "text", "cross_page": true } ], "index": 31 }, { "bbox": [ 114, 484, 406, 500 ], "spans": [ { "bbox": [ 114, 484, 406, 500 ], "score": 1.0, "content": "tacl a 00324. URL https://aclanthology.org/2020.tacl-1.28.", "type": "text", "cross_page": true } ], "index": 32, "is_list_end_line": true }, { "bbox": [ 106, 504, 506, 517 ], "spans": [ { "bbox": [ 106, 504, 506, 517 ], "score": 1.0, "content": "Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. In Bengio, Y. and LeCun,", "type": "text", "cross_page": true } ], "index": 33, "is_list_start_line": true }, { "bbox": [ 115, 514, 506, 528 ], "spans": [ { "bbox": [ 115, 514, 506, 528 ], "score": 1.0, "content": "Y. (eds.), 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA,", "type": "text", "cross_page": true } ], "index": 34 }, { "bbox": [ 116, 525, 506, 539 ], "spans": [ { "bbox": [ 116, 525, 506, 539 ], "score": 1.0, "content": "USA, May 7-9, 2015, Conference Track Proceedings, 2015. URL http://arxiv.org/abs/1412.", "type": "text", "cross_page": true } ], "index": 35 }, { "bbox": [ 114, 536, 144, 550 ], "spans": [ { "bbox": [ 114, 536, 144, 550 ], "score": 1.0, "content": "6980.", "type": "text", "cross_page": true } ], "index": 36, "is_list_end_line": true }, { "bbox": [ 106, 555, 506, 569 ], "spans": [ { "bbox": [ 106, 555, 506, 569 ], "score": 1.0, "content": "Kohonen, T. Correlation matrix memories. IEEE transactions on computers, 100(4):353–359, 1972.", "type": "text", "cross_page": true } ], "index": 37, "is_list_start_line": true }, { "bbox": [ 104, 572, 507, 588 ], "spans": [ { "bbox": [ 104, 572, 507, 588 ], "score": 1.0, "content": "Levy, O., Seo, M., Choi, E., and Zettlemoyer, L. Zero-shot relation extraction via reading compre-", "type": "text", "cross_page": true } ], "index": 38, "is_list_start_line": true }, { "bbox": [ 113, 583, 506, 600 ], "spans": [ { "bbox": [ 113, 583, 506, 600 ], "score": 1.0, "content": "hension. In Proceedings of the 21st Conference on Computational Natural Language Learning", "type": "text", "cross_page": true } ], "index": 39 }, { "bbox": [ 114, 595, 506, 611 ], "spans": [ { "bbox": [ 114, 595, 506, 611 ], "score": 1.0, "content": "(CoNLL 2017), pp. 333–342, Vancouver, Canada, August 2017. Association for Computational", "type": "text", "cross_page": true } ], "index": 40 }, { "bbox": [ 114, 605, 486, 622 ], "spans": [ { "bbox": [ 114, 605, 486, 622 ], "score": 1.0, "content": "Linguistics. doi: 10.18653/v1/K17-1034. URL https://aclanthology.org/K17-1034.", "type": "text", "cross_page": true } ], "index": 41, "is_list_end_line": true }, { "bbox": [ 105, 625, 506, 639 ], "spans": [ { "bbox": [ 105, 625, 506, 639 ], "score": 1.0, "content": "Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., and", "type": "text", "cross_page": true } ], "index": 42, "is_list_start_line": true }, { "bbox": [ 114, 635, 507, 652 ], "spans": [ { "bbox": [ 114, 635, 507, 652 ], "score": 1.0, "content": "Zettlemoyer, L. BART: Denoising sequence-to-sequence pre-training for natural language", "type": "text", "cross_page": true } ], "index": 43 }, { "bbox": [ 115, 648, 506, 661 ], "spans": [ { "bbox": [ 115, 648, 506, 661 ], "score": 1.0, "content": "generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of", "type": "text", "cross_page": true } ], "index": 44 }, { "bbox": [ 115, 658, 506, 672 ], "spans": [ { "bbox": [ 115, 658, 506, 672 ], "score": 1.0, "content": "the Association for Computational Linguistics, pp. 7871–7880, Online, July 2020. Associ-", "type": "text", "cross_page": true } ], "index": 45 }, { "bbox": [ 115, 669, 506, 683 ], "spans": [ { "bbox": [ 115, 669, 506, 683 ], "score": 1.0, "content": "ation for Computational Linguistics. doi: 10.18653/v1/2020.acl-main.703. URL https:", "type": "text", "cross_page": true } ], "index": 46 }, { "bbox": [ 116, 681, 310, 694 ], "spans": [ { "bbox": [ 116, 681, 310, 694 ], "score": 1.0, "content": "//aclanthology.org/2020.acl-main.703.", "type": "text", "cross_page": true } ], "index": 47, "is_list_end_line": true }, { "bbox": [ 104, 697, 506, 714 ], "spans": [ { "bbox": [ 104, 697, 506, 714 ], "score": 1.0, "content": "Meng, K., Sen Sharma, A., Andonian, A., Belinkov, Y., and Bau, D. Mass-editing memory in a", "type": "text", "cross_page": true } ], "index": 48, "is_list_start_line": true }, { "bbox": [ 115, 710, 331, 724 ], "spans": [ { "bbox": [ 115, 710, 331, 724 ], "score": 1.0, "content": "transformer. arXiv preprint arXiv:2210.07229, 2022.", "type": "text", "cross_page": true } ], "index": 49, "is_list_end_line": true }, { "bbox": [ 105, 72, 506, 86 ], "spans": [ { "bbox": [ 105, 72, 506, 86 ], "score": 1.0, "content": "Mitchell, E., Lin, C., Bosselut, A., Finn, C., and Manning, C. D. Fast model editing at scale. In", "type": "text", "cross_page": true } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 114, 83, 365, 96 ], "spans": [ { "bbox": [ 114, 83, 365, 96 ], "score": 1.0, "content": "International Conference on Learning Representations, 2021.", "type": "text", "cross_page": true } ], "index": 1, "is_list_end_line": true }, { "bbox": [ 105, 102, 506, 116 ], "spans": [ { "bbox": [ 105, 102, 506, 116 ], "score": 1.0, "content": "Pearl, J. Direct and indirect effects. In Proceedings of the Seventeenth conference on Uncertainty in", "type": "text", "cross_page": true } ], "index": 2, "is_list_start_line": true }, { "bbox": [ 114, 113, 285, 126 ], "spans": [ { "bbox": [ 114, 113, 285, 126 ], "score": 1.0, "content": "artificial intelligence, pp. 411–420, 2001.", "type": "text", "cross_page": true } ], "index": 3, "is_list_end_line": true }, { "bbox": [ 105, 131, 505, 145 ], "spans": [ { "bbox": [ 105, 131, 505, 145 ], "score": 1.0, "content": "Pearl, J. Causality: Models, Reasoning and Inference. Cambridge University Press, USA, 2nd", "type": "text", "cross_page": true } ], "index": 4, "is_list_start_line": true }, { "bbox": [ 115, 142, 258, 156 ], "spans": [ { "bbox": [ 115, 142, 258, 156 ], "score": 1.0, "content": "edition, 2009. ISBN 052189560X.", "type": "text", "cross_page": true } ], "index": 5, "is_list_end_line": true }, { "bbox": [ 104, 159, 507, 177 ], "spans": [ { "bbox": [ 104, 159, 507, 177 ], "score": 1.0, "content": "Petroni, F., Rocktaschel, T., Riedel, S., Lewis, P., Bakhtin, A., Wu, Y., and Miller, A. Language ¨", "type": "text", "cross_page": true } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 115, 172, 505, 186 ], "spans": [ { "bbox": [ 115, 172, 505, 186 ], "score": 1.0, "content": "models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 113, 182, 506, 199 ], "spans": [ { "bbox": [ 113, 182, 506, 199 ], "score": 1.0, "content": "Natural Language Processing and the 9th International Joint Conference on Natural Language", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 115, 194, 506, 208 ], "spans": [ { "bbox": [ 115, 194, 506, 208 ], "score": 1.0, "content": "Processing (EMNLP-IJCNLP), pp. 2463–2473, Hong Kong, China, November 2019. Association", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 113, 203, 507, 221 ], "spans": [ { "bbox": [ 113, 203, 507, 221 ], "score": 1.0, "content": "for Computational Linguistics. doi: 10.18653/v1/D19-1250. URL https://aclanthology.", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 115, 216, 186, 229 ], "spans": [ { "bbox": [ 115, 216, 186, 229 ], "score": 1.0, "content": "org/D19-1250.", "type": "text", "cross_page": true } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 105, 235, 506, 249 ], "spans": [ { "bbox": [ 105, 235, 506, 249 ], "score": 1.0, "content": "Petroni, F., Lewis, P., Piktus, A., Rocktaschel, T., Wu, Y., Miller, A. H., and Riedel, S. How context ¨", "type": "text", "cross_page": true } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 115, 246, 506, 260 ], "spans": [ { "bbox": [ 115, 246, 506, 260 ], "score": 1.0, "content": "affects language models’ factual predictions. In Automated Knowledge Base Construction, 2020.", "type": "text", "cross_page": true } ], "index": 13 }, { "bbox": [ 106, 265, 504, 277 ], "spans": [ { "bbox": [ 106, 265, 504, 277 ], "score": 1.0, "content": "Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. Language models are", "type": "text", "cross_page": true } ], "index": 14, "is_list_start_line": true }, { "bbox": [ 115, 276, 357, 290 ], "spans": [ { "bbox": [ 115, 276, 357, 290 ], "score": 1.0, "content": "unsupervised multitask learners. OpenAI blog, pp. 9, 2019.", "type": "text", "cross_page": true } ], "index": 15, "is_list_end_line": true }, { "bbox": [ 106, 295, 506, 307 ], "spans": [ { "bbox": [ 106, 295, 506, 307 ], "score": 1.0, "content": "Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., and Liu, P. J.", "type": "text", "cross_page": true } ], "index": 16, "is_list_start_line": true }, { "bbox": [ 115, 306, 506, 320 ], "spans": [ { "bbox": [ 115, 306, 506, 320 ], "score": 1.0, "content": "Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine", "type": "text", "cross_page": true } ], "index": 17 }, { "bbox": [ 115, 317, 282, 330 ], "spans": [ { "bbox": [ 115, 317, 282, 330 ], "score": 1.0, "content": "Learning Research, 21(140):1–67, 2020.", "type": "text", "cross_page": true } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 105, 334, 507, 349 ], "spans": [ { "bbox": [ 105, 334, 507, 349 ], "score": 1.0, "content": "Roberts, A., Raffel, C., and Shazeer, N. How much knowledge can you pack into the param-", "type": "text", "cross_page": true } ], "index": 19, "is_list_start_line": true }, { "bbox": [ 113, 345, 507, 361 ], "spans": [ { "bbox": [ 113, 345, 507, 361 ], "score": 1.0, "content": "eters of a language model? In Proceedings of the 2020 Conference on Empirical Methods", "type": "text", "cross_page": true } ], "index": 20 }, { "bbox": [ 113, 356, 508, 371 ], "spans": [ { "bbox": [ 113, 356, 508, 371 ], "score": 1.0, "content": "in Natural Language Processing (EMNLP), pp. 5418–5426, Online, November 2020. Associ-", "type": "text", "cross_page": true } ], "index": 21 }, { "bbox": [ 115, 368, 507, 382 ], "spans": [ { "bbox": [ 115, 368, 507, 382 ], "score": 1.0, "content": "ation for Computational Linguistics. doi: 10.18653/v1/2020.emnlp-main.437. URL https:", "type": "text", "cross_page": true } ], "index": 22 }, { "bbox": [ 115, 379, 320, 393 ], "spans": [ { "bbox": [ 115, 379, 320, 393 ], "score": 1.0, "content": "//aclanthology.org/2020.emnlp-main.437.", "type": "text", "cross_page": true } ], "index": 23, "is_list_end_line": true }, { "bbox": [ 104, 396, 507, 413 ], "spans": [ { "bbox": [ 104, 396, 507, 413 ], "score": 1.0, "content": "Sundararajan, M., Taly, A., and Yan, Q. Axiomatic attribution for deep networks. In International", "type": "text", "cross_page": true } ], "index": 24, "is_list_start_line": true }, { "bbox": [ 115, 409, 372, 423 ], "spans": [ { "bbox": [ 115, 409, 372, 423 ], "score": 1.0, "content": "conference on machine learning, pp. 3319–3328. PMLR, 2017.", "type": "text", "cross_page": true } ], "index": 25, "is_list_end_line": true }, { "bbox": [ 106, 427, 506, 441 ], "spans": [ { "bbox": [ 106, 427, 506, 441 ], "score": 1.0, "content": "Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., and", "type": "text", "cross_page": true } ], "index": 26, "is_list_start_line": true }, { "bbox": [ 114, 436, 507, 455 ], "spans": [ { "bbox": [ 114, 436, 507, 455 ], "score": 1.0, "content": "Polosukhin, I. Attention is all you need. In Advances in neural information processing systems, pp.", "type": "text", "cross_page": true } ], "index": 27 }, { "bbox": [ 115, 448, 192, 463 ], "spans": [ { "bbox": [ 115, 448, 192, 463 ], "score": 1.0, "content": "5998–6008, 2017.", "type": "text", "cross_page": true } ], "index": 28, "is_list_end_line": true }, { "bbox": [ 106, 468, 506, 482 ], "spans": [ { "bbox": [ 106, 468, 506, 482 ], "score": 1.0, "content": "Vig, J., Gehrmann, S., Belinkov, Y., Qian, S., Nevo, D., Sakenis, S., Huang, J., Singer, Y., and Shieber,", "type": "text", "cross_page": true } ], "index": 29, "is_list_start_line": true }, { "bbox": [ 115, 479, 506, 493 ], "spans": [ { "bbox": [ 115, 479, 506, 493 ], "score": 1.0, "content": "S. Causal mediation analysis for interpreting neural NLP: The case of gender bias. arXiv preprint", "type": "text", "cross_page": true } ], "index": 30 }, { "bbox": [ 115, 491, 224, 503 ], "spans": [ { "bbox": [ 115, 491, 224, 503 ], "score": 1.0, "content": "arXiv:2004.12265, 2020a.", "type": "text", "cross_page": true } ], "index": 31, "is_list_end_line": true }, { "bbox": [ 105, 507, 507, 524 ], "spans": [ { "bbox": [ 105, 507, 507, 524 ], "score": 1.0, "content": "Vig, J., Gehrmann, S., Belinkov, Y., Qian, S., Nevo, D., Singer, Y., and Shieber, S. M. Investigating", "type": "text", "cross_page": true } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 113, 520, 457, 533 ], "spans": [ { "bbox": [ 113, 520, 457, 533 ], "score": 1.0, "content": "gender bias in language models using causal mediation analysis. In NeurIPS, 2020b.", "type": "text", "cross_page": true } ], "index": 33, "is_list_end_line": true }, { "bbox": [ 105, 538, 507, 554 ], "spans": [ { "bbox": [ 105, 538, 507, 554 ], "score": 1.0, "content": "Wang, B. and Komatsuzaki, A. GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model.", "type": "text", "cross_page": true } ], "index": 34, "is_list_start_line": true }, { "bbox": [ 115, 550, 431, 564 ], "spans": [ { "bbox": [ 115, 550, 431, 564 ], "score": 1.0, "content": "https://github.com/kingoflolz/mesh-transformer-jax, May 2021.", "type": "text", "cross_page": true } ], "index": 35, "is_list_end_line": true }, { "bbox": [ 105, 567, 506, 582 ], "spans": [ { "bbox": [ 105, 567, 506, 582 ], "score": 1.0, "content": "Zhang, Y., Galley, M., Gao, J., Gan, Z., Li, X., Brockett, C., and Dolan, W. B. Generating informative", "type": "text", "cross_page": true } ], "index": 36, "is_list_start_line": true }, { "bbox": [ 115, 580, 507, 594 ], "spans": [ { "bbox": [ 115, 580, 507, 594 ], "score": 1.0, "content": "and diverse conversational responses via adversarial information maximization. In NeurIPS, 2018.", "type": "text", "cross_page": true } ], "index": 37 }, { "bbox": [ 104, 597, 507, 612 ], "spans": [ { "bbox": [ 104, 597, 507, 612 ], "score": 1.0, "content": "Zhao, S., Pascual, D., Brunner, G., and Wattenhofer, R. Of non-linearity and commutativity in BERT.", "type": "text", "cross_page": true } ], "index": 38, "is_list_start_line": true }, { "bbox": [ 115, 610, 484, 624 ], "spans": [ { "bbox": [ 115, 610, 484, 624 ], "score": 1.0, "content": "In 2021 International Joint Conference on Neural Networks (IJCNN), pp. 1–8. IEEE, 2021.", "type": "text", "cross_page": true } ], "index": 39, "is_list_end_line": true }, { "bbox": [ 105, 628, 506, 642 ], "spans": [ { "bbox": [ 105, 628, 506, 642 ], "score": 1.0, "content": "Zhong, Z., Friedman, D., and Chen, D. Factual probing is [MASK]: Learning vs. learning to", "type": "text", "cross_page": true } ], "index": 40, "is_list_start_line": true }, { "bbox": [ 115, 639, 505, 653 ], "spans": [ { "bbox": [ 115, 639, 505, 653 ], "score": 1.0, "content": "recall. In Proceedings of the 2021 Conference of the North American Chapter of the Association", "type": "text", "cross_page": true } ], "index": 41 }, { "bbox": [ 115, 650, 506, 664 ], "spans": [ { "bbox": [ 115, 650, 506, 664 ], "score": 1.0, "content": "for Computational Linguistics: Human Language Technologies, pp. 5017–5033, Online, June", "type": "text", "cross_page": true } ], "index": 42 }, { "bbox": [ 116, 662, 505, 673 ], "spans": [ { "bbox": [ 116, 662, 505, 673 ], "score": 1.0, "content": "2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.naacl-main.398. URL", "type": "text", "cross_page": true } ], "index": 43 }, { "bbox": [ 115, 672, 351, 686 ], "spans": [ { "bbox": [ 115, 672, 351, 686 ], "score": 1.0, "content": "https://aclanthology.org/2021.naacl-main.398.", "type": "text", "cross_page": true } ], "index": 44, "is_list_end_line": true }, { "bbox": [ 103, 689, 506, 706 ], "spans": [ { "bbox": [ 103, 689, 506, 706 ], "score": 1.0, "content": "Zhu, C., Rawat, A. S., Zaheer, M., Bhojanapalli, S., Li, D., Yu, F., and Kumar, S. Modifying", "type": "text", "cross_page": true } ], "index": 45, "is_list_start_line": true }, { "bbox": [ 115, 701, 415, 715 ], "spans": [ { "bbox": [ 115, 701, 415, 715 ], "score": 1.0, "content": "memories in transformer models. arXiv preprint arXiv:2012.00363, 2020.", "type": "text", "cross_page": true } ], "index": 46, "is_list_end_line": true } ], "index": 25, "bbox_fs": [ 104, 70, 509, 725 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 104, 72, 507, 727 ], "lines": [ { "bbox": [ 105, 72, 506, 85 ], "spans": [ { "bbox": [ 105, 72, 506, 85 ], "score": 1.0, "content": "Elazar, Y., Ravfogel, S., Jacovi, A., and Goldberg, Y. Amnesic probing: Behavioral explanation", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 83, 507, 97 ], "spans": [ { "bbox": [ 115, 83, 507, 97 ], "score": 1.0, "content": "with amnesic counterfactuals. Transactions of the Association for Computational Linguistics, 9:", "type": "text" } ], "index": 1 }, { "bbox": [ 116, 95, 185, 106 ], "spans": [ { "bbox": [ 116, 95, 185, 106 ], "score": 1.0, "content": "160–175, 2021b.", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 113, 506, 126 ], "spans": [ { "bbox": [ 106, 113, 506, 126 ], "score": 1.0, "content": "Elhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., Askell, A., Bai, Y., Chen,", "type": "text" } ], "index": 3 }, { "bbox": [ 114, 123, 506, 137 ], "spans": [ { "bbox": [ 114, 123, 506, 137 ], "score": 1.0, "content": "A., Conerly, T., DasSarma, N., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D.,", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 136, 506, 147 ], "spans": [ { "bbox": [ 116, 136, 506, 147 ], "score": 1.0, "content": "Jones, A., Kernion, J., Lovitt, L., Ndousse, K., Amodei, D., Brown, T., Clark, J., Kaplan, J.,", "type": "text" } ], "index": 5 }, { "bbox": [ 114, 144, 508, 159 ], "spans": [ { "bbox": [ 114, 144, 508, 159 ], "score": 1.0, "content": "McCandlish, S., and Olah, C. A mathematical framework for transformer circuits. https:", "type": "text" } ], "index": 6 }, { "bbox": [ 116, 156, 461, 170 ], "spans": [ { "bbox": [ 116, 156, 461, 170 ], "score": 1.0, "content": "//transformer-circuits.pub/2021/framework/index.html, December 2021.", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 176, 506, 187 ], "spans": [ { "bbox": [ 106, 176, 506, 187 ], "score": 1.0, "content": "Ettinger, A., Elgohary, A., and Resnik, P. Probing for semantic evidence of composition by means of", "type": "text" } ], "index": 8 }, { "bbox": [ 116, 186, 507, 200 ], "spans": [ { "bbox": [ 116, 186, 507, 200 ], "score": 1.0, "content": "simple classification tasks. In Proceedings of the 1st Workshop on Evaluating Vector-Space Repre-", "type": "text" } ], "index": 9 }, { "bbox": [ 114, 196, 506, 211 ], "spans": [ { "bbox": [ 114, 196, 506, 211 ], "score": 1.0, "content": "sentations for NLP, pp. 134–139, Berlin, Germany, August 2016. Association for Computational", "type": "text" } ], "index": 10 }, { "bbox": [ 114, 206, 488, 223 ], "spans": [ { "bbox": [ 114, 206, 488, 223 ], "score": 1.0, "content": "Linguistics. doi: 10.18653/v1/W16-2524. URL https://aclanthology.org/W16-2524.", "type": "text" } ], "index": 11 }, { "bbox": [ 104, 225, 506, 241 ], "spans": [ { "bbox": [ 104, 225, 506, 241 ], "score": 1.0, "content": "Feder, A., Oved, N., Shalit, U., and Reichart, R. CausaLM: Causal model explanation through", "type": "text" } ], "index": 12 }, { "bbox": [ 114, 238, 448, 251 ], "spans": [ { "bbox": [ 114, 238, 448, 251 ], "score": 1.0, "content": "counterfactual language models. Computational Linguistics, 47(2):333–386, 2021.", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 256, 506, 270 ], "spans": [ { "bbox": [ 105, 256, 506, 270 ], "score": 1.0, "content": "Finlayson, M., Mueller, A., Gehrmann, S., Shieber, S., Linzen, T., and Belinkov, Y. Causal analysis", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 267, 505, 280 ], "spans": [ { "bbox": [ 115, 267, 505, 280 ], "score": 1.0, "content": "of syntactic agreement mechanisms in neural language models. In Proceedings of the 59th", "type": "text" } ], "index": 15 }, { "bbox": [ 115, 280, 505, 291 ], "spans": [ { "bbox": [ 115, 280, 505, 291 ], "score": 1.0, "content": "Annual Meeting of the Association for Computational Linguistics and the 11th International Joint", "type": "text" } ], "index": 16 }, { "bbox": [ 116, 289, 506, 303 ], "spans": [ { "bbox": [ 116, 289, 506, 303 ], "score": 1.0, "content": "Conference on Natural Language Processing (Volume 1: Long Papers), pp. 1828–1843, Online,", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 300, 507, 314 ], "spans": [ { "bbox": [ 115, 300, 507, 314 ], "score": 1.0, "content": "August 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.acl-long.144.", "type": "text" } ], "index": 18 }, { "bbox": [ 116, 311, 364, 325 ], "spans": [ { "bbox": [ 116, 311, 364, 325 ], "score": 1.0, "content": "URL https://aclanthology.org/2021.acl-long.144.", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 329, 507, 344 ], "spans": [ { "bbox": [ 105, 329, 507, 344 ], "score": 1.0, "content": "Geva, M., Schuster, R., Berant, J., and Levy, O. Transformer feed-forward layers are key-value memo-", "type": "text" } ], "index": 20 }, { "bbox": [ 114, 340, 506, 355 ], "spans": [ { "bbox": [ 114, 340, 506, 355 ], "score": 1.0, "content": "ries. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,", "type": "text" } ], "index": 21 }, { "bbox": [ 114, 351, 507, 366 ], "spans": [ { "bbox": [ 114, 351, 507, 366 ], "score": 1.0, "content": "pp. 5484–5495, Online and Punta Cana, Dominican Republic, November 2021. Association for", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 363, 487, 377 ], "spans": [ { "bbox": [ 115, 363, 487, 377 ], "score": 1.0, "content": "Computational Linguistics. URL https://aclanthology.org/2021.emnlp-main.446.", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 381, 506, 395 ], "spans": [ { "bbox": [ 105, 381, 506, 395 ], "score": 1.0, "content": "Hase, P., Diab, M., Celikyilmaz, A., Li, X., Kozareva, Z., Stoyanov, V., Bansal, M., and Iyer, S. Do", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 393, 506, 406 ], "spans": [ { "bbox": [ 115, 393, 506, 406 ], "score": 1.0, "content": "language models have beliefs? methods for detecting, updating, and visualizing model beliefs.", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 403, 281, 417 ], "spans": [ { "bbox": [ 115, 403, 281, 417 ], "score": 1.0, "content": "arXiv preprint arXiv:2111.13654, 2021.", "type": "text" } ], "index": 26 }, { "bbox": [ 104, 421, 506, 436 ], "spans": [ { "bbox": [ 104, 421, 506, 436 ], "score": 1.0, "content": "Hupkes, D., Veldhoen, S., and Zuidema, W. Visualisation and ’diagnostic classifiers’ reveal how", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 433, 506, 447 ], "spans": [ { "bbox": [ 115, 433, 506, 447 ], "score": 1.0, "content": "recurrent and recursive neural networks process hierarchical structure. Journal of Artificial", "type": "text" } ], "index": 28 }, { "bbox": [ 114, 444, 286, 457 ], "spans": [ { "bbox": [ 114, 444, 286, 457 ], "score": 1.0, "content": "Intelligence Research, 61:907–926, 2018.", "type": "text" } ], "index": 29 }, { "bbox": [ 104, 462, 506, 478 ], "spans": [ { "bbox": [ 104, 462, 506, 478 ], "score": 1.0, "content": "Jiang, Z., Xu, F. F., Araki, J., and Neubig, G. How can we know what language models know?", "type": "text" } ], "index": 30 }, { "bbox": [ 116, 473, 506, 487 ], "spans": [ { "bbox": [ 116, 473, 506, 487 ], "score": 1.0, "content": "Transactions of the Association for Computational Linguistics, 8:423–438, 2020. doi: 10.1162/", "type": "text" } ], "index": 31 }, { "bbox": [ 114, 484, 406, 500 ], "spans": [ { "bbox": [ 114, 484, 406, 500 ], "score": 1.0, "content": "tacl a 00324. URL https://aclanthology.org/2020.tacl-1.28.", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 504, 506, 517 ], "spans": [ { "bbox": [ 106, 504, 506, 517 ], "score": 1.0, "content": "Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. In Bengio, Y. and LeCun,", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 514, 506, 528 ], "spans": [ { "bbox": [ 115, 514, 506, 528 ], "score": 1.0, "content": "Y. (eds.), 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA,", "type": "text" } ], "index": 34 }, { "bbox": [ 116, 525, 506, 539 ], "spans": [ { "bbox": [ 116, 525, 506, 539 ], "score": 1.0, "content": "USA, May 7-9, 2015, Conference Track Proceedings, 2015. URL http://arxiv.org/abs/1412.", "type": "text" } ], "index": 35 }, { "bbox": [ 114, 536, 144, 550 ], "spans": [ { "bbox": [ 114, 536, 144, 550 ], "score": 1.0, "content": "6980.", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 555, 506, 569 ], "spans": [ { "bbox": [ 106, 555, 506, 569 ], "score": 1.0, "content": "Kohonen, T. Correlation matrix memories. IEEE transactions on computers, 100(4):353–359, 1972.", "type": "text" } ], "index": 37 }, { "bbox": [ 104, 572, 507, 588 ], "spans": [ { "bbox": [ 104, 572, 507, 588 ], "score": 1.0, "content": "Levy, O., Seo, M., Choi, E., and Zettlemoyer, L. Zero-shot relation extraction via reading compre-", "type": "text" } ], "index": 38 }, { "bbox": [ 113, 583, 506, 600 ], "spans": [ { "bbox": [ 113, 583, 506, 600 ], "score": 1.0, "content": "hension. In Proceedings of the 21st Conference on Computational Natural Language Learning", "type": "text" } ], "index": 39 }, { "bbox": [ 114, 595, 506, 611 ], "spans": [ { "bbox": [ 114, 595, 506, 611 ], "score": 1.0, "content": "(CoNLL 2017), pp. 333–342, Vancouver, Canada, August 2017. Association for Computational", "type": "text" } ], "index": 40 }, { "bbox": [ 114, 605, 486, 622 ], "spans": [ { "bbox": [ 114, 605, 486, 622 ], "score": 1.0, "content": "Linguistics. doi: 10.18653/v1/K17-1034. URL https://aclanthology.org/K17-1034.", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 625, 506, 639 ], "spans": [ { "bbox": [ 105, 625, 506, 639 ], "score": 1.0, "content": "Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., and", "type": "text" } ], "index": 42 }, { "bbox": [ 114, 635, 507, 652 ], "spans": [ { "bbox": [ 114, 635, 507, 652 ], "score": 1.0, "content": "Zettlemoyer, L. BART: Denoising sequence-to-sequence pre-training for natural language", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 648, 506, 661 ], "spans": [ { "bbox": [ 115, 648, 506, 661 ], "score": 1.0, "content": "generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of", "type": "text" } ], "index": 44 }, { "bbox": [ 115, 658, 506, 672 ], "spans": [ { "bbox": [ 115, 658, 506, 672 ], "score": 1.0, "content": "the Association for Computational Linguistics, pp. 7871–7880, Online, July 2020. Associ-", "type": "text" } ], "index": 45 }, { "bbox": [ 115, 669, 506, 683 ], "spans": [ { "bbox": [ 115, 669, 506, 683 ], "score": 1.0, "content": "ation for Computational Linguistics. doi: 10.18653/v1/2020.acl-main.703. URL https:", "type": "text" } ], "index": 46 }, { "bbox": [ 116, 681, 310, 694 ], "spans": [ { "bbox": [ 116, 681, 310, 694 ], "score": 1.0, "content": "//aclanthology.org/2020.acl-main.703.", "type": "text" } ], "index": 47 }, { "bbox": [ 104, 697, 506, 714 ], "spans": [ { "bbox": [ 104, 697, 506, 714 ], "score": 1.0, "content": "Meng, K., Sen Sharma, A., Andonian, A., Belinkov, Y., and Bau, D. Mass-editing memory in a", "type": "text" } ], "index": 48 }, { "bbox": [ 115, 710, 331, 724 ], "spans": [ { "bbox": [ 115, 710, 331, 724 ], "score": 1.0, "content": "transformer. arXiv preprint arXiv:2210.07229, 2022.", "type": "text" } ], "index": 49 } ], "index": 24.5 } ], "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": [ 104, 72, 507, 727 ], "lines": [], "index": 24.5, "bbox_fs": [ 104, 72, 508, 724 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 103, 48, 507, 716 ], "lines": [ { "bbox": [ 105, 72, 506, 86 ], "spans": [ { "bbox": [ 105, 72, 506, 86 ], "score": 1.0, "content": "Mitchell, E., Lin, C., Bosselut, A., Finn, C., and Manning, C. D. Fast model editing at scale. In", "type": "text" } ], "index": 0 }, { "bbox": [ 114, 83, 365, 96 ], "spans": [ { "bbox": [ 114, 83, 365, 96 ], "score": 1.0, "content": "International Conference on Learning Representations, 2021.", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 102, 506, 116 ], "spans": [ { "bbox": [ 105, 102, 506, 116 ], "score": 1.0, "content": "Pearl, J. Direct and indirect effects. In Proceedings of the Seventeenth conference on Uncertainty in", "type": "text" } ], "index": 2 }, { "bbox": [ 114, 113, 285, 126 ], "spans": [ { "bbox": [ 114, 113, 285, 126 ], "score": 1.0, "content": "artificial intelligence, pp. 411–420, 2001.", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 131, 505, 145 ], "spans": [ { "bbox": [ 105, 131, 505, 145 ], "score": 1.0, "content": "Pearl, J. Causality: Models, Reasoning and Inference. Cambridge University Press, USA, 2nd", "type": "text" } ], "index": 4 }, { "bbox": [ 115, 142, 258, 156 ], "spans": [ { "bbox": [ 115, 142, 258, 156 ], "score": 1.0, "content": "edition, 2009. ISBN 052189560X.", "type": "text" } ], "index": 5 }, { "bbox": [ 104, 159, 507, 177 ], "spans": [ { "bbox": [ 104, 159, 507, 177 ], "score": 1.0, "content": "Petroni, F., Rocktaschel, T., Riedel, S., Lewis, P., Bakhtin, A., Wu, Y., and Miller, A. Language ¨", "type": "text" } ], "index": 6 }, { "bbox": [ 115, 172, 505, 186 ], "spans": [ { "bbox": [ 115, 172, 505, 186 ], "score": 1.0, "content": "models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in", "type": "text" } ], "index": 7 }, { "bbox": [ 113, 182, 506, 199 ], "spans": [ { "bbox": [ 113, 182, 506, 199 ], "score": 1.0, "content": "Natural Language Processing and the 9th International Joint Conference on Natural Language", "type": "text" } ], "index": 8 }, { "bbox": [ 115, 194, 506, 208 ], "spans": [ { "bbox": [ 115, 194, 506, 208 ], "score": 1.0, "content": "Processing (EMNLP-IJCNLP), pp. 2463–2473, Hong Kong, China, November 2019. Association", "type": "text" } ], "index": 9 }, { "bbox": [ 113, 203, 507, 221 ], "spans": [ { "bbox": [ 113, 203, 507, 221 ], "score": 1.0, "content": "for Computational Linguistics. doi: 10.18653/v1/D19-1250. URL https://aclanthology.", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 216, 186, 229 ], "spans": [ { "bbox": [ 115, 216, 186, 229 ], "score": 1.0, "content": "org/D19-1250.", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 235, 506, 249 ], "spans": [ { "bbox": [ 105, 235, 506, 249 ], "score": 1.0, "content": "Petroni, F., Lewis, P., Piktus, A., Rocktaschel, T., Wu, Y., Miller, A. H., and Riedel, S. How context ¨", "type": "text" } ], "index": 12 }, { "bbox": [ 115, 246, 506, 260 ], "spans": [ { "bbox": [ 115, 246, 506, 260 ], "score": 1.0, "content": "affects language models’ factual predictions. In Automated Knowledge Base Construction, 2020.", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 265, 504, 277 ], "spans": [ { "bbox": [ 106, 265, 504, 277 ], "score": 1.0, "content": "Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. Language models are", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 276, 357, 290 ], "spans": [ { "bbox": [ 115, 276, 357, 290 ], "score": 1.0, "content": "unsupervised multitask learners. OpenAI blog, pp. 9, 2019.", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 295, 506, 307 ], "spans": [ { "bbox": [ 106, 295, 506, 307 ], "score": 1.0, "content": "Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., and Liu, P. J.", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 306, 506, 320 ], "spans": [ { "bbox": [ 115, 306, 506, 320 ], "score": 1.0, "content": "Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 317, 282, 330 ], "spans": [ { "bbox": [ 115, 317, 282, 330 ], "score": 1.0, "content": "Learning Research, 21(140):1–67, 2020.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 334, 507, 349 ], "spans": [ { "bbox": [ 105, 334, 507, 349 ], "score": 1.0, "content": "Roberts, A., Raffel, C., and Shazeer, N. How much knowledge can you pack into the param-", "type": "text" } ], "index": 19 }, { "bbox": [ 113, 345, 507, 361 ], "spans": [ { "bbox": [ 113, 345, 507, 361 ], "score": 1.0, "content": "eters of a language model? In Proceedings of the 2020 Conference on Empirical Methods", "type": "text" } ], "index": 20 }, { "bbox": [ 113, 356, 508, 371 ], "spans": [ { "bbox": [ 113, 356, 508, 371 ], "score": 1.0, "content": "in Natural Language Processing (EMNLP), pp. 5418–5426, Online, November 2020. Associ-", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 368, 507, 382 ], "spans": [ { "bbox": [ 115, 368, 507, 382 ], "score": 1.0, "content": "ation for Computational Linguistics. doi: 10.18653/v1/2020.emnlp-main.437. URL https:", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 379, 320, 393 ], "spans": [ { "bbox": [ 115, 379, 320, 393 ], "score": 1.0, "content": "//aclanthology.org/2020.emnlp-main.437.", "type": "text" } ], "index": 23 }, { "bbox": [ 104, 396, 507, 413 ], "spans": [ { "bbox": [ 104, 396, 507, 413 ], "score": 1.0, "content": "Sundararajan, M., Taly, A., and Yan, Q. Axiomatic attribution for deep networks. In International", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 409, 372, 423 ], "spans": [ { "bbox": [ 115, 409, 372, 423 ], "score": 1.0, "content": "conference on machine learning, pp. 3319–3328. PMLR, 2017.", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 427, 506, 441 ], "spans": [ { "bbox": [ 106, 427, 506, 441 ], "score": 1.0, "content": "Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., and", "type": "text" } ], "index": 26 }, { "bbox": [ 114, 436, 507, 455 ], "spans": [ { "bbox": [ 114, 436, 507, 455 ], "score": 1.0, "content": "Polosukhin, I. Attention is all you need. In Advances in neural information processing systems, pp.", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 448, 192, 463 ], "spans": [ { "bbox": [ 115, 448, 192, 463 ], "score": 1.0, "content": "5998–6008, 2017.", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 468, 506, 482 ], "spans": [ { "bbox": [ 106, 468, 506, 482 ], "score": 1.0, "content": "Vig, J., Gehrmann, S., Belinkov, Y., Qian, S., Nevo, D., Sakenis, S., Huang, J., Singer, Y., and Shieber,", "type": "text" } ], "index": 29 }, { "bbox": [ 115, 479, 506, 493 ], "spans": [ { "bbox": [ 115, 479, 506, 493 ], "score": 1.0, "content": "S. Causal mediation analysis for interpreting neural NLP: The case of gender bias. arXiv preprint", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 491, 224, 503 ], "spans": [ { "bbox": [ 115, 491, 224, 503 ], "score": 1.0, "content": "arXiv:2004.12265, 2020a.", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 507, 507, 524 ], "spans": [ { "bbox": [ 105, 507, 507, 524 ], "score": 1.0, "content": "Vig, J., Gehrmann, S., Belinkov, Y., Qian, S., Nevo, D., Singer, Y., and Shieber, S. M. Investigating", "type": "text" } ], "index": 32 }, { "bbox": [ 113, 520, 457, 533 ], "spans": [ { "bbox": [ 113, 520, 457, 533 ], "score": 1.0, "content": "gender bias in language models using causal mediation analysis. In NeurIPS, 2020b.", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 538, 507, 554 ], "spans": [ { "bbox": [ 105, 538, 507, 554 ], "score": 1.0, "content": "Wang, B. and Komatsuzaki, A. GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model.", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 550, 431, 564 ], "spans": [ { "bbox": [ 115, 550, 431, 564 ], "score": 1.0, "content": "https://github.com/kingoflolz/mesh-transformer-jax, May 2021.", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 567, 506, 582 ], "spans": [ { "bbox": [ 105, 567, 506, 582 ], "score": 1.0, "content": "Zhang, Y., Galley, M., Gao, J., Gan, Z., Li, X., Brockett, C., and Dolan, W. B. Generating informative", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 580, 507, 594 ], "spans": [ { "bbox": [ 115, 580, 507, 594 ], "score": 1.0, "content": "and diverse conversational responses via adversarial information maximization. In NeurIPS, 2018.", "type": "text" } ], "index": 37 }, { "bbox": [ 104, 597, 507, 612 ], "spans": [ { "bbox": [ 104, 597, 507, 612 ], "score": 1.0, "content": "Zhao, S., Pascual, D., Brunner, G., and Wattenhofer, R. Of non-linearity and commutativity in BERT.", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 610, 484, 624 ], "spans": [ { "bbox": [ 115, 610, 484, 624 ], "score": 1.0, "content": "In 2021 International Joint Conference on Neural Networks (IJCNN), pp. 1–8. IEEE, 2021.", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 628, 506, 642 ], "spans": [ { "bbox": [ 105, 628, 506, 642 ], "score": 1.0, "content": "Zhong, Z., Friedman, D., and Chen, D. Factual probing is [MASK]: Learning vs. learning to", "type": "text" } ], "index": 40 }, { "bbox": [ 115, 639, 505, 653 ], "spans": [ { "bbox": [ 115, 639, 505, 653 ], "score": 1.0, "content": "recall. In Proceedings of the 2021 Conference of the North American Chapter of the Association", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 650, 506, 664 ], "spans": [ { "bbox": [ 115, 650, 506, 664 ], "score": 1.0, "content": "for Computational Linguistics: Human Language Technologies, pp. 5017–5033, Online, June", "type": "text" } ], "index": 42 }, { "bbox": [ 116, 662, 505, 673 ], "spans": [ { "bbox": [ 116, 662, 505, 673 ], "score": 1.0, "content": "2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.naacl-main.398. URL", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 672, 351, 686 ], "spans": [ { "bbox": [ 115, 672, 351, 686 ], "score": 1.0, "content": "https://aclanthology.org/2021.naacl-main.398.", "type": "text" } ], "index": 44 }, { "bbox": [ 103, 689, 506, 706 ], "spans": [ { "bbox": [ 103, 689, 506, 706 ], "score": 1.0, "content": "Zhu, C., Rawat, A. S., Zaheer, M., Bhojanapalli, S., Li, D., Yu, F., and Kumar, S. Modifying", "type": "text" } ], "index": 45 }, { "bbox": [ 115, 701, 415, 715 ], "spans": [ { "bbox": [ 115, 701, 415, 715 ], "score": 1.0, "content": "memories in transformer models. arXiv preprint arXiv:2012.00363, 2020.", "type": "text" } ], "index": 46 } ], "index": 23 } ], "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": [ 103, 48, 507, 716 ], "lines": [], "index": 23, "bbox_fs": [ 103, 72, 508, 715 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 107, 71, 156, 84 ], "lines": [ { "bbox": [ 106, 70, 158, 85 ], "spans": [ { "bbox": [ 106, 70, 158, 85 ], "score": 1.0, "content": "Checklist", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 131, 99, 208, 110 ], "lines": [ { "bbox": [ 129, 97, 210, 111 ], "spans": [ { "bbox": [ 129, 97, 210, 111 ], "score": 1.0, "content": "1. For all authors...", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 146, 114, 505, 187 ], "lines": [ { "bbox": [ 145, 113, 505, 127 ], "spans": [ { "bbox": [ 145, 113, 505, 127 ], "score": 1.0, "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", "type": "text" } ], "index": 2 }, { "bbox": [ 161, 124, 288, 138 ], "spans": [ { "bbox": [ 161, 124, 288, 138 ], "score": 1.0, "content": "contributions and scope? [Yes]", "type": "text" } ], "index": 3 }, { "bbox": [ 145, 137, 376, 150 ], "spans": [ { "bbox": [ 145, 137, 376, 150 ], "score": 1.0, "content": "(b) Did you describe the limitations of your work? [Yes]", "type": "text" } ], "index": 4 }, { "bbox": [ 145, 150, 468, 163 ], "spans": [ { "bbox": [ 145, 150, 468, 163 ], "score": 1.0, "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes]", "type": "text" } ], "index": 5 }, { "bbox": [ 145, 163, 505, 176 ], "spans": [ { "bbox": [ 145, 163, 505, 176 ], "score": 1.0, "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", "type": "text" } ], "index": 6 }, { "bbox": [ 161, 174, 214, 188 ], "spans": [ { "bbox": [ 161, 174, 214, 188 ], "score": 1.0, "content": "them? [Yes]", "type": "text" } ], "index": 7 } ], "index": 4.5 }, { "type": "text", "bbox": [ 131, 190, 302, 201 ], "lines": [ { "bbox": [ 129, 189, 304, 203 ], "spans": [ { "bbox": [ 129, 189, 304, 203 ], "score": 1.0, "content": "2. If you are including theoretical results...", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 145, 204, 500, 229 ], "lines": [ { "bbox": [ 145, 203, 500, 217 ], "spans": [ { "bbox": [ 145, 203, 500, 217 ], "score": 1.0, "content": "(a) Did you state the full set of assumptions of all theoretical results? [Yes] In appendix", "type": "text" } ], "index": 9 }, { "bbox": [ 144, 216, 471, 230 ], "spans": [ { "bbox": [ 144, 216, 471, 230 ], "score": 1.0, "content": "(b) Did you include complete proofs of all theoretical results? [Yes] In appendix", "type": "text" } ], "index": 10 } ], "index": 9.5 }, { "type": "text", "bbox": [ 131, 232, 241, 244 ], "lines": [ { "bbox": [ 128, 230, 243, 246 ], "spans": [ { "bbox": [ 128, 230, 243, 246 ], "score": 1.0, "content": "3. If you ran experiments...", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 146, 247, 505, 352 ], "lines": [ { "bbox": [ 145, 246, 506, 260 ], "spans": [ { "bbox": [ 145, 246, 506, 260 ], "score": 1.0, "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-", "type": "text" } ], "index": 12 }, { "bbox": [ 161, 258, 505, 270 ], "spans": [ { "bbox": [ 161, 258, 505, 270 ], "score": 1.0, "content": "mental results (either in the supplemental material or as a URL)? [Yes] In supplemental", "type": "text" } ], "index": 13 }, { "bbox": [ 161, 269, 201, 279 ], "spans": [ { "bbox": [ 161, 269, 201, 279 ], "score": 1.0, "content": "materials", "type": "text" } ], "index": 14 }, { "bbox": [ 144, 281, 506, 296 ], "spans": [ { "bbox": [ 144, 281, 506, 296 ], "score": 1.0, "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", "type": "text" } ], "index": 15 }, { "bbox": [ 162, 293, 295, 304 ], "spans": [ { "bbox": [ 162, 293, 295, 304 ], "score": 1.0, "content": "were chosen)? [Yes] In appendix", "type": "text" } ], "index": 16 }, { "bbox": [ 146, 304, 507, 318 ], "spans": [ { "bbox": [ 146, 304, 507, 318 ], "score": 1.0, "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-", "type": "text" } ], "index": 17 }, { "bbox": [ 162, 317, 280, 329 ], "spans": [ { "bbox": [ 162, 317, 280, 329 ], "score": 1.0, "content": "ments multiple times)? [Yes]", "type": "text" } ], "index": 18 }, { "bbox": [ 146, 328, 505, 343 ], "spans": [ { "bbox": [ 146, 328, 505, 343 ], "score": 1.0, "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type", "type": "text" } ], "index": 19 }, { "bbox": [ 162, 340, 420, 353 ], "spans": [ { "bbox": [ 162, 340, 420, 353 ], "score": 1.0, "content": "of GPUs, internal cluster, or cloud provider)? [Yes] In appendix", "type": "text" } ], "index": 20 } ], "index": 16 }, { "type": "text", "bbox": [ 129, 355, 504, 367 ], "lines": [ { "bbox": [ 129, 353, 506, 369 ], "spans": [ { "bbox": [ 129, 353, 506, 369 ], "score": 1.0, "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 146, 370, 505, 466 ], "lines": [ { "bbox": [ 145, 370, 424, 383 ], "spans": [ { "bbox": [ 145, 370, 424, 383 ], "score": 1.0, "content": "(a) If your work uses existing assets, did you cite the creators? [Yes]", "type": "text" } ], "index": 22 }, { "bbox": [ 145, 383, 407, 396 ], "spans": [ { "bbox": [ 145, 383, 407, 396 ], "score": 1.0, "content": "(b) Did you mention the license of the assets? [Yes] In appendix", "type": "text" } ], "index": 23 }, { "bbox": [ 145, 395, 505, 409 ], "spans": [ { "bbox": [ 145, 395, 505, 409 ], "score": 1.0, "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [Yes]", "type": "text" } ], "index": 24 }, { "bbox": [ 162, 407, 259, 419 ], "spans": [ { "bbox": [ 162, 407, 259, 419 ], "score": 1.0, "content": "Supplemental materials", "type": "text" } ], "index": 25 }, { "bbox": [ 146, 419, 505, 432 ], "spans": [ { "bbox": [ 146, 419, 505, 432 ], "score": 1.0, "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", "type": "text" } ], "index": 26 }, { "bbox": [ 161, 431, 300, 443 ], "spans": [ { "bbox": [ 161, 431, 300, 443 ], "score": 1.0, "content": "using/curating? [Yes] In appendix", "type": "text" } ], "index": 27 }, { "bbox": [ 146, 443, 505, 456 ], "spans": [ { "bbox": [ 146, 443, 505, 456 ], "score": 1.0, "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", "type": "text" } ], "index": 28 }, { "bbox": [ 162, 453, 371, 467 ], "spans": [ { "bbox": [ 162, 453, 371, 467 ], "score": 1.0, "content": "information or offensive content? [Yes] In appendix", "type": "text" } ], "index": 29 } ], "index": 25.5 }, { "type": "text", "bbox": [ 132, 469, 431, 480 ], "lines": [ { "bbox": [ 128, 468, 433, 483 ], "spans": [ { "bbox": [ 128, 468, 433, 483 ], "score": 1.0, "content": "5. If you used crowdsourcing or conducted research with human subjects...", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 146, 484, 505, 555 ], "lines": [ { "bbox": [ 145, 483, 506, 497 ], "spans": [ { "bbox": [ 145, 483, 506, 497 ], "score": 1.0, "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", "type": "text" } ], "index": 31 }, { "bbox": [ 162, 495, 284, 507 ], "spans": [ { "bbox": [ 162, 495, 284, 507 ], "score": 1.0, "content": "applicable? [Yes] In appendix", "type": "text" } ], "index": 32 }, { "bbox": [ 145, 508, 505, 520 ], "spans": [ { "bbox": [ 145, 508, 505, 520 ], "score": 1.0, "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", "type": "text" } ], "index": 33 }, { "bbox": [ 161, 519, 389, 532 ], "spans": [ { "bbox": [ 161, 519, 389, 532 ], "score": 1.0, "content": "Board (IRB) approvals, if applicable? [Yes] In appendix", "type": "text" } ], "index": 34 }, { "bbox": [ 147, 532, 506, 544 ], "spans": [ { "bbox": [ 147, 532, 506, 544 ], "score": 1.0, "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", "type": "text" } ], "index": 35 }, { "bbox": [ 162, 543, 379, 555 ], "spans": [ { "bbox": [ 162, 543, 379, 555 ], "score": 1.0, "content": "spent on participant compensation? [Yes] In appendix", "type": "text" } ], "index": 36 } ], "index": 33.5 } ], "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": "title", "bbox": [ 107, 71, 156, 84 ], "lines": [ { "bbox": [ 106, 70, 158, 85 ], "spans": [ { "bbox": [ 106, 70, 158, 85 ], "score": 1.0, "content": "Checklist", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 131, 99, 208, 110 ], "lines": [ { "bbox": [ 129, 97, 210, 111 ], "spans": [ { "bbox": [ 129, 97, 210, 111 ], "score": 1.0, "content": "1. For all authors...", "type": "text" } ], "index": 1 } ], "index": 1, "bbox_fs": [ 129, 97, 210, 111 ] }, { "type": "list", "bbox": [ 146, 114, 505, 187 ], "lines": [ { "bbox": [ 145, 113, 505, 127 ], "spans": [ { "bbox": [ 145, 113, 505, 127 ], "score": 1.0, "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", "type": "text" } ], "index": 2, "is_list_start_line": true }, { "bbox": [ 161, 124, 288, 138 ], "spans": [ { "bbox": [ 161, 124, 288, 138 ], "score": 1.0, "content": "contributions and scope? [Yes]", "type": "text" } ], "index": 3, "is_list_end_line": true }, { "bbox": [ 145, 137, 376, 150 ], "spans": [ { "bbox": [ 145, 137, 376, 150 ], "score": 1.0, "content": "(b) Did you describe the limitations of your work? [Yes]", "type": "text" } ], "index": 4, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 145, 150, 468, 163 ], "spans": [ { "bbox": [ 145, 150, 468, 163 ], "score": 1.0, "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes]", "type": "text" } ], "index": 5, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 145, 163, 505, 176 ], "spans": [ { "bbox": [ 145, 163, 505, 176 ], "score": 1.0, "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", "type": "text" } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 161, 174, 214, 188 ], "spans": [ { "bbox": [ 161, 174, 214, 188 ], "score": 1.0, "content": "them? [Yes]", "type": "text" } ], "index": 7, "is_list_end_line": true } ], "index": 4.5, "bbox_fs": [ 145, 113, 505, 188 ] }, { "type": "text", "bbox": [ 131, 190, 302, 201 ], "lines": [ { "bbox": [ 129, 189, 304, 203 ], "spans": [ { "bbox": [ 129, 189, 304, 203 ], "score": 1.0, "content": "2. If you are including theoretical results...", "type": "text" } ], "index": 8 } ], "index": 8, "bbox_fs": [ 129, 189, 304, 203 ] }, { "type": "text", "bbox": [ 145, 204, 500, 229 ], "lines": [ { "bbox": [ 145, 203, 500, 217 ], "spans": [ { "bbox": [ 145, 203, 500, 217 ], "score": 1.0, "content": "(a) Did you state the full set of assumptions of all theoretical results? [Yes] In appendix", "type": "text" } ], "index": 9 }, { "bbox": [ 144, 216, 471, 230 ], "spans": [ { "bbox": [ 144, 216, 471, 230 ], "score": 1.0, "content": "(b) Did you include complete proofs of all theoretical results? [Yes] In appendix", "type": "text" } ], "index": 10 } ], "index": 9.5, "bbox_fs": [ 144, 203, 500, 230 ] }, { "type": "text", "bbox": [ 131, 232, 241, 244 ], "lines": [ { "bbox": [ 128, 230, 243, 246 ], "spans": [ { "bbox": [ 128, 230, 243, 246 ], "score": 1.0, "content": "3. If you ran experiments...", "type": "text" } ], "index": 11 } ], "index": 11, "bbox_fs": [ 128, 230, 243, 246 ] }, { "type": "list", "bbox": [ 146, 247, 505, 352 ], "lines": [ { "bbox": [ 145, 246, 506, 260 ], "spans": [ { "bbox": [ 145, 246, 506, 260 ], "score": 1.0, "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-", "type": "text" } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 161, 258, 505, 270 ], "spans": [ { "bbox": [ 161, 258, 505, 270 ], "score": 1.0, "content": "mental results (either in the supplemental material or as a URL)? [Yes] In supplemental", "type": "text" } ], "index": 13 }, { "bbox": [ 161, 269, 201, 279 ], "spans": [ { "bbox": [ 161, 269, 201, 279 ], "score": 1.0, "content": "materials", "type": "text" } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 144, 281, 506, 296 ], "spans": [ { "bbox": [ 144, 281, 506, 296 ], "score": 1.0, "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", "type": "text" } ], "index": 15, "is_list_start_line": true }, { "bbox": [ 162, 293, 295, 304 ], "spans": [ { "bbox": [ 162, 293, 295, 304 ], "score": 1.0, "content": "were chosen)? [Yes] In appendix", "type": "text" } ], "index": 16, "is_list_end_line": true }, { "bbox": [ 146, 304, 507, 318 ], "spans": [ { "bbox": [ 146, 304, 507, 318 ], "score": 1.0, "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-", "type": "text" } ], "index": 17, "is_list_start_line": true }, { "bbox": [ 162, 317, 280, 329 ], "spans": [ { "bbox": [ 162, 317, 280, 329 ], "score": 1.0, "content": "ments multiple times)? [Yes]", "type": "text" } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 146, 328, 505, 343 ], "spans": [ { "bbox": [ 146, 328, 505, 343 ], "score": 1.0, "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type", "type": "text" } ], "index": 19, "is_list_start_line": true }, { "bbox": [ 162, 340, 420, 353 ], "spans": [ { "bbox": [ 162, 340, 420, 353 ], "score": 1.0, "content": "of GPUs, internal cluster, or cloud provider)? [Yes] In appendix", "type": "text" } ], "index": 20, "is_list_end_line": true } ], "index": 16, "bbox_fs": [ 144, 246, 507, 353 ] }, { "type": "text", "bbox": [ 129, 355, 504, 367 ], "lines": [ { "bbox": [ 129, 353, 506, 369 ], "spans": [ { "bbox": [ 129, 353, 506, 369 ], "score": 1.0, "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", "type": "text" } ], "index": 21 } ], "index": 21, "bbox_fs": [ 129, 353, 506, 369 ] }, { "type": "list", "bbox": [ 146, 370, 505, 466 ], "lines": [ { "bbox": [ 145, 370, 424, 383 ], "spans": [ { "bbox": [ 145, 370, 424, 383 ], "score": 1.0, "content": "(a) If your work uses existing assets, did you cite the creators? [Yes]", "type": "text" } ], "index": 22, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 145, 383, 407, 396 ], "spans": [ { "bbox": [ 145, 383, 407, 396 ], "score": 1.0, "content": "(b) Did you mention the license of the assets? [Yes] In appendix", "type": "text" } ], "index": 23, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 145, 395, 505, 409 ], "spans": [ { "bbox": [ 145, 395, 505, 409 ], "score": 1.0, "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [Yes]", "type": "text" } ], "index": 24, "is_list_start_line": true }, { "bbox": [ 162, 407, 259, 419 ], "spans": [ { "bbox": [ 162, 407, 259, 419 ], "score": 1.0, "content": "Supplemental materials", "type": "text" } ], "index": 25, "is_list_end_line": true }, { "bbox": [ 146, 419, 505, 432 ], "spans": [ { "bbox": [ 146, 419, 505, 432 ], "score": 1.0, "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", "type": "text" } ], "index": 26, "is_list_start_line": true }, { "bbox": [ 161, 431, 300, 443 ], "spans": [ { "bbox": [ 161, 431, 300, 443 ], "score": 1.0, "content": "using/curating? [Yes] In appendix", "type": "text" } ], "index": 27, "is_list_end_line": true }, { "bbox": [ 146, 443, 505, 456 ], "spans": [ { "bbox": [ 146, 443, 505, 456 ], "score": 1.0, "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", "type": "text" } ], "index": 28, "is_list_start_line": true }, { "bbox": [ 162, 453, 371, 467 ], "spans": [ { "bbox": [ 162, 453, 371, 467 ], "score": 1.0, "content": "information or offensive content? [Yes] In appendix", "type": "text" } ], "index": 29, "is_list_end_line": true } ], "index": 25.5, "bbox_fs": [ 145, 370, 505, 467 ] }, { "type": "text", "bbox": [ 132, 469, 431, 480 ], "lines": [ { "bbox": [ 128, 468, 433, 483 ], "spans": [ { "bbox": [ 128, 468, 433, 483 ], "score": 1.0, "content": "5. If you used crowdsourcing or conducted research with human subjects...", "type": "text" } ], "index": 30 } ], "index": 30, "bbox_fs": [ 128, 468, 433, 483 ] }, { "type": "list", "bbox": [ 146, 484, 505, 555 ], "lines": [ { "bbox": [ 145, 483, 506, 497 ], "spans": [ { "bbox": [ 145, 483, 506, 497 ], "score": 1.0, "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", "type": "text" } ], "index": 31, "is_list_start_line": true }, { "bbox": [ 162, 495, 284, 507 ], "spans": [ { "bbox": [ 162, 495, 284, 507 ], "score": 1.0, "content": "applicable? [Yes] In appendix", "type": "text" } ], "index": 32, "is_list_end_line": true }, { "bbox": [ 145, 508, 505, 520 ], "spans": [ { "bbox": [ 145, 508, 505, 520 ], "score": 1.0, "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", "type": "text" } ], "index": 33, "is_list_start_line": true }, { "bbox": [ 161, 519, 389, 532 ], "spans": [ { "bbox": [ 161, 519, 389, 532 ], "score": 1.0, "content": "Board (IRB) approvals, if applicable? [Yes] In appendix", "type": "text" } ], "index": 34, "is_list_end_line": true }, { "bbox": [ 147, 532, 506, 544 ], "spans": [ { "bbox": [ 147, 532, 506, 544 ], "score": 1.0, "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", "type": "text" } ], "index": 35, "is_list_start_line": true }, { "bbox": [ 162, 543, 379, 555 ], "spans": [ { "bbox": [ 162, 543, 379, 555 ], "score": 1.0, "content": "spent on participant compensation? [Yes] In appendix", "type": "text" } ], "index": 36, "is_list_end_line": true } ], "index": 33.5, "bbox_fs": [ 145, 483, 506, 555 ] } ] } ], "_backend": "pipeline", "_version_name": "2.2.2" }