Explorations of Self-Repair in Language Models

Cody Rushing, Neel Nanda
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:42836-42855, 2024.

Abstract

Prior interpretability research studying narrow distributions has preliminarily identified self-repair, a phenomena where if components in large language models are ablated, later components will change their behavior to compensate. Our work builds off this past literature, demonstrating that self-repair exists on a variety of models families and sizes when ablating individual attention heads on the full training distribution. We further show that on the full training distribution self-repair is imperfect, as the original direct effect of the head is not fully restored, and noisy, since the degree of self-repair varies significantly across different prompts (sometimes overcorrecting beyond the original effect). We highlight two different mechanisms that contribute to self-repair, including changes in the final LayerNorm scaling factor and sparse sets of neurons implementing Anti-Erasure. We additionally discuss the implications of these results for interpretability practitioners and close with a more speculative discussion on the mystery of why self-repair occurs in these models at all, highlighting evidence for the Iterative Inference hypothesis in language models, a framework that predicts self-repair.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-rushing24a, title = {Explorations of Self-Repair in Language Models}, author = {Rushing, Cody and Nanda, Neel}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {42836--42855}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/rushing24a/rushing24a.pdf}, url = {https://proceedings.mlr.press/v235/rushing24a.html}, abstract = {Prior interpretability research studying narrow distributions has preliminarily identified self-repair, a phenomena where if components in large language models are ablated, later components will change their behavior to compensate. Our work builds off this past literature, demonstrating that self-repair exists on a variety of models families and sizes when ablating individual attention heads on the full training distribution. We further show that on the full training distribution self-repair is imperfect, as the original direct effect of the head is not fully restored, and noisy, since the degree of self-repair varies significantly across different prompts (sometimes overcorrecting beyond the original effect). We highlight two different mechanisms that contribute to self-repair, including changes in the final LayerNorm scaling factor and sparse sets of neurons implementing Anti-Erasure. We additionally discuss the implications of these results for interpretability practitioners and close with a more speculative discussion on the mystery of why self-repair occurs in these models at all, highlighting evidence for the Iterative Inference hypothesis in language models, a framework that predicts self-repair.} }
Endnote
%0 Conference Paper %T Explorations of Self-Repair in Language Models %A Cody Rushing %A Neel Nanda %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-rushing24a %I PMLR %P 42836--42855 %U https://proceedings.mlr.press/v235/rushing24a.html %V 235 %X Prior interpretability research studying narrow distributions has preliminarily identified self-repair, a phenomena where if components in large language models are ablated, later components will change their behavior to compensate. Our work builds off this past literature, demonstrating that self-repair exists on a variety of models families and sizes when ablating individual attention heads on the full training distribution. We further show that on the full training distribution self-repair is imperfect, as the original direct effect of the head is not fully restored, and noisy, since the degree of self-repair varies significantly across different prompts (sometimes overcorrecting beyond the original effect). We highlight two different mechanisms that contribute to self-repair, including changes in the final LayerNorm scaling factor and sparse sets of neurons implementing Anti-Erasure. We additionally discuss the implications of these results for interpretability practitioners and close with a more speculative discussion on the mystery of why self-repair occurs in these models at all, highlighting evidence for the Iterative Inference hypothesis in language models, a framework that predicts self-repair.
APA
Rushing, C. & Nanda, N.. (2024). Explorations of Self-Repair in Language Models. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:42836-42855 Available from https://proceedings.mlr.press/v235/rushing24a.html.

Related Material