AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Circuit Insights: Towards Interpretability Beyond Activations

Elena Golimblevskaia, Aakriti Jain, Bruno Puri, Ammar Ibrahim, Wojciech Samek, Sebastian Lapuschkin

ICLR 2026 · Architectures and efficiency › RNNs and normalization · approach: Mechanistic circuit interpretability

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
use WeightLens token descriptions instead of activation-maximization descriptionsGemma-2-2B transcoder features · ~250 features per layerClarity and ResponsivenessImproved
use WeightLens instead of MaxAct* (ablation)Gemma-2-2B transcoder features · ~250 features per layerPurityHurt
add LLM refinement (gpt-4o-mini) to WeightLens+Out (ablation)WeightLens · Gemma-2-2B transcodersFADE scoresNo clear effect
add output-token (unembedding) information to WeightLens descriptions (ablation)WeightLens · Gemma-2-2B transcodersFADE scoresNo clear effect
add validation step keeping only tokens that activate the feature (ablation)WeightLens · Gemma-2-2B transcodersdescription qualityImproved
use circuit-based input masking and clustering instead of activation-only explainer inputsCircuitLens · Gemma-2-2B transcodersFADE scoresImproved
use the larger 2.3B-token dataset instead of 24M tokens (ablation)CircuitLens · Gemma-2-2B transcodersinput-centric FADE metricsImproved
combine WeightLens with CircuitLens-Full (ablation)CircuitLens · Gemma-2-2B transcodersrobustness to dataset sizeImproved
use circuit-based clusteringCircuitLens · Gemma-2-2B transcodersnumber of very low-clarity featuresImproved
use WeightLens or MaxAct* descriptions (ablation)all methods · Gemma-2-2B, GPT-2, Llama-3.2-1BFaithfulnessNo clear effect
apply WeightLens to middle layers (ablation)WeightLens · Gemma-2-2B and Llama-3.2-1Bfraction of token-based featuresHurt

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