Summary of Explorations Of Self-repair in Language Models, by Cody Rushing et al.
Explorations of Self-Repair in Language Models
by Cody Rushing, Neel Nanda
First submitted to arxiv on: 23 Feb 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This research paper investigates self-repair, a phenomenon where components in large language models adapt to compensate when individual parts are removed. The study demonstrates that self-repair exists across various model families and sizes when ablating attention heads on the full training dataset. However, it is imperfect and noisy, with the original effect of the head not fully restored and varying degrees of compensation across different prompts. Two mechanisms contributing to self-repair are identified: changes in LayerNorm scaling factors and sparse neuron sets implementing Anti-Erasure. The findings have implications for interpretability practitioners and offer insights into why self-repair occurs in these models, with evidence pointing towards the Iterative Inference hypothesis. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This study shows that when we remove parts of a big language model, other parts can change to make up for it. This is called self-repair. The researchers looked at many different types and sizes of language models and found that this happens in all of them. But it’s not perfect – sometimes the original effect is not fully restored, and sometimes it even gets worse than before. They also discovered two ways that self-repair works: by changing how certain parts of the model are scaled, or by using special neurons to “undo” what was done earlier. This has important implications for people who try to understand language models, and it might help us figure out why these models can sometimes compensate for missing parts. |
Keywords
* Artificial intelligence * Attention * Inference * Language model
