Research map / Architectures and efficiency / RNNs and normalization
Does Higher Interpretability Imply Better Utility? A Pairwise Analysis on Sparse Autoencoders
Xu Wang, Yan Hu, Benyou Wang, Difan Zou
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 |
|---|---|---|---|
| add Delta Token Confidence feature selection | SAE steering on Gemma-2-2B · AxBench Concept100 | steering score: 0.133 -> 0.328 | Improved |
| add Delta Token Confidence feature selection | SAE steering on Qwen-2.5-3B · AxBench Concept100 | steering score: 0.171 -> 0.399 | Improved |
| add Delta Token Confidence feature selection | SAE steering on Gemma-2-9B · AxBench Concept100 | steering score: 0.142 -> 0.289 | Improved |
| select by Delta Token Confidence instead of output score | SAE steering on Gemma-2-2B · AxBench Concept100 | steering score: 0.233 -> 0.328 | Improved |
| select by Delta Token Confidence instead of output score | SAE steering on Qwen-2.5-3B · AxBench Concept100 | steering score: 0.292 -> 0.399 | Improved |
| select by Delta Token Confidence instead of output score | SAE steering on Gemma-2-9B · AxBench Concept100 | steering score: 0.255 -> 0.289 | Improved |
| add output-score feature selection | SAE steering on Gemma-2-2B · AxBench Concept100 | steering score: 0.133 -> 0.233 | Improved |
| increase SAE sparsity level L0 from 50 to 820 | SAE pool · interpretability-steering rank agreement | Kendall tau_b: 0.5429 -> -0.0476 | Hurt |
| use Gated architecture instead of other architectures | SAE pool · interpretability-steering rank agreement | Kendall tau_b: -0.2026 | Hurt |
| select high-steering-gain features before correlating | SAE pool · interpretability vs steering | Kendall tau_b: 0.2979 -> -0.0692 | Hurt |
Closest papers by meaning
- SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability (ICML 2025 · similarity 0.90)
- Transformer Key-Value Memories Are Nearly as Interpretable as Sparse Autoencoders (NeurIPS 2025 · similarity 0.89)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.89)
- Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words (ICLR 2025 · similarity 0.88)
- Automatically Interpreting Millions of Features in Large Language Models (ICML 2025 · similarity 0.88)
- Scaling and evaluating sparse autoencoders (ICLR 2025 · similarity 0.88)
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.
