Researchers extract circuits from language models using Attention routing
The authors propose a method to extract circuits from language models by treating attention as a routing map from a single forward pass. They evaluate on GPT-2 Small, GPT-2 Medium, and Pythia-410M using 100 prompts per cell. Ablation of extracted edges hurts performance more than random edges. The extraction step costs one forward pass, whereas a head‑by‑head patch sweep costs about two orders…
1 source primary source
Key points
- Method extracts circuits using attention routing from a single forward pass
- Evaluation on GPT-2 Small, GPT-2 Medium, Pythia-410M with 100 prompts per cell
- Extraction costs one forward pass, patch sweep costs ~100x more
Read the original at arXiv cs.AI · by Ash Manvi, Samreena Tajreen primary sourceOpen source ↗
Topics · follow one to build your own front page
GPT-2 SmallGPT-2 MediumPythia-410M
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Semantic Routing Calibration mitigates LLM over-refusal · 1 src
- AIBuildAI-2.5 ranks first on MLE-Bench with 73.3% medal rate · 1 src
- Researchers fine-tune 406M model for meeting summaries with retrieved text spans · 1 src
- Researchers test how language models handle numerical formats in word problems · 1 src
- Author pretrains language model End-to-End in Rust for $164 · 1 src
Comments
via GitHub Discussions