Research map / Large language models / Language, safety and interpretability
Do Sparse Autoencoders Identify Reasoning Features in Language Models?
George Ma, Zhongyuan Liang, Irene Y. Chen, Somayeh Sojoudi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| inject a few top-activating tokens into non-reasoning text (ablation) | top-100 contrastive SAE reasoning features · 22 configurations (models, layers, datasets) | share of features with significant activation increase: 45% to 90% of features activated | Improved |
| apply LLM-guided counterexample falsification to context-dependent features (ablation) | context-dependent SAE features · 248 features across configurations | genuine reasoning features: 0 of 248 genuine | No clear effect |
| steer feature 1053 with gamma=2 (ablation) | Gemma-3-12B-Instruct layer 22 · GPQA Diamond | accuracy: 37.9% -> 13.6% | Hurt |
| steer feature 1053 with gamma=2 (ablation) | Gemma-3-12B-Instruct layer 22 · AIME | accuracy: 26.7% -> 20.0% | Hurt |
| steer feature 0 with gamma=2 (ablation) | Gemma-3-12B-Instruct layer 22 · AIME | accuracy: 26.7% -> 10.0% | Hurt |
| steer feature 0 with gamma=2 (ablation) | Gemma-3-12B-Instruct layer 22 · GPQA Diamond | accuracy: 37.9% -> 20.2% | Hurt |
| steer feature 578 with gamma=2 (ablation) | Gemma-3-12B-Instruct layer 22 · AIME | accuracy: 26.7% -> 26.7% | No clear effect |
| steer feature 578 with gamma=2 (ablation) | Gemma-3-12B-Instruct layer 22 · GPQA Diamond | accuracy: 37.9% -> 33.3% | Hurt |
| use Gemma-3-12B layer 22 on s1K instead of other configurations | token injection analysis · s1K | context-dependent fraction: 55% context-dependent vs fewer than 12% in several Gemma configurations | Mixed |
Closest papers by meaning
- I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders (AAAI 2026 · similarity 0.94)
- Step-Level Sparse Autoencoder for Reasoning Process Interpretation (ICML 2026 · similarity 0.92)
- ActivationReasoning: Logical Reasoning in Latent Activation Spaces (ICLR 2026 · similarity 0.90)
- The Tell-Tale Norm: $\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models (ICML 2026 · similarity 0.90)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.89)
- Automatically Interpreting Millions of Features in Large Language Models (ICML 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
