AI Paper Atlas

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

ICLR 2026 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models · 5 citations (OpenAlex)

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 changedWhereMeasuredEffect
add Delta Token Confidence feature selectionSAE steering on Gemma-2-2B · AxBench Concept100steering score: 0.133 -> 0.328Improved
add Delta Token Confidence feature selectionSAE steering on Qwen-2.5-3B · AxBench Concept100steering score: 0.171 -> 0.399Improved
add Delta Token Confidence feature selectionSAE steering on Gemma-2-9B · AxBench Concept100steering score: 0.142 -> 0.289Improved
select by Delta Token Confidence instead of output scoreSAE steering on Gemma-2-2B · AxBench Concept100steering score: 0.233 -> 0.328Improved
select by Delta Token Confidence instead of output scoreSAE steering on Qwen-2.5-3B · AxBench Concept100steering score: 0.292 -> 0.399Improved
select by Delta Token Confidence instead of output scoreSAE steering on Gemma-2-9B · AxBench Concept100steering score: 0.255 -> 0.289Improved
add output-score feature selectionSAE steering on Gemma-2-2B · AxBench Concept100steering score: 0.133 -> 0.233Improved
increase SAE sparsity level L0 from 50 to 820SAE pool · interpretability-steering rank agreementKendall tau_b: 0.5429 -> -0.0476Hurt
use Gated architecture instead of other architecturesSAE pool · interpretability-steering rank agreementKendall tau_b: -0.2026Hurt
select high-steering-gain features before correlatingSAE pool · interpretability vs steeringKendall tau_b: 0.2979 -> -0.0692Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy