Timezone: »
We analyze the storage and recall of factual associations in autoregressive transformer language models, finding evidence that these associations correspond to localized, directly-editable computations. We first develop a causal intervention for identifying neuron activations that are decisive in a model's factual predictions. This reveals a distinct set of steps in middle-layer feed-forward modules that mediate factual predictions while processing subject tokens. To test our hypothesis that these computations correspond to factual association recall, we modify feed-forward weights to update specific factual associations using Rank-One Model Editing (ROME). We find that ROME is effective on a standard zero-shot relation extraction (zsRE) model-editing task, comparable to existing methods. To perform a more sensitive evaluation, we also evaluate ROME on a new dataset of counterfactual assertions, on which it simultaneously maintains both specificity and generalization, whereas other methods sacrifice one or another. Our results confirm an important role for mid-layer feed-forward modules in storing factual associations and suggest that direct manipulation of computational mechanisms may be a feasible approach for model editing. The code, dataset, visualizations, and an interactive demo notebook are available in the supplemental materials.
Author Information
Kevin Meng (MIT)
EECS @ MIT. Interested in language models, robotics, and computational biology.
David Bau (Northeastern University)
Alex Andonian (MIT)
Yonatan Belinkov (Technion)
More from the Same Authors
-
2020 : Horses With Blue Jeans - Creating New Worlds by Rewriting a GAN »
David Bau -
2022 Poster: Measures of Information Reflect Memorization Patterns »
Rachit Bansal · Danish Pruthi · Yonatan Belinkov -
2021 Poster: Editing a classifier by rewriting its prediction rules »
Shibani Santurkar · Dimitris Tsipras · Mahalaxmi Elango · David Bau · Antonio Torralba · Aleksander Madry -
2018 : Coffee break + posters 2 »
Jan Kremer · Erik McDermott · Brandon Carter · Albert Zeyer · Andreas Krug · Paul Pu Liang · Katherine Lee · Dominika Basaj · Abelino Jimenez · Lisa Fan · Gautam Bhattacharya · Tzeviya S Fuchs · David Gifford · Loren Lugosch · Orhan Firat · Benjamin Baer · JAHANGIR ALAM · Jamin Shin · Mirco Ravanelli · Paul Smolensky · Zining Zhu · Hamid Eghbal-zadeh · Skyler Seto · Imran Sheikh · Joao Felipe Santos · Yonatan Belinkov · Nadir Durrani · Oiwi Parker Jones · Shuai Tang · André Merboldt · Titouan Parcollet · Wei-Ning Hsu · Krishna Pillutla · Ehsan Hosseini-Asl · Monica Dinculescu · Alexander Amini · Ying Zhang · Taoli Cheng · Alain Tapp -
2018 : Coffee break + posters 1 »
Samuel Myer · Wei-Ning Hsu · Jialu Li · Monica Dinculescu · Lea Schönherr · Ehsan Hosseini-Asl · Skyler Seto · Oiwi Parker Jones · Imran Sheikh · Thomas Manzini · Yonatan Belinkov · Nadir Durrani · Alexander Amini · Johanna Hansen · Gabi Shalev · Jamin Shin · Paul Smolensky · Lisa Fan · Zining Zhu · Hamid Eghbal-zadeh · Benjamin Baer · Abelino Jimenez · Joao Felipe Santos · Jan Kremer · Erik McDermott · Andreas Krug · Tzeviya S Fuchs · Shuai Tang · Brandon Carter · David Gifford · Albert Zeyer · André Merboldt · Krishna Pillutla · Katherine Lee · Titouan Parcollet · Orhan Firat · Gautam Bhattacharya · JAHANGIR ALAM · Mirco Ravanelli -
2017 Poster: Analyzing Hidden Representations in End-to-End Automatic Speech Recognition Systems »
Yonatan Belinkov · Jim Glass