AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Language Model Circuits Are Sparse in the Neuron Basis

Aryaman Arora, Zhengxuan Wu, Jacob Steinhardt, Sarah Schwettmann

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Sparse dictionary representations · 0 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
use pre-down-projection MLP activations instead of MLP outputs as the neuron basis (ablation)IG circuit tracing on Llama 3.1 8B · SVA (4 tasks avg)circuit size at given faithfulness: about 100x smaller circuitsImproved
use pre-down-projection MLP activations instead of MLP outputs (ablation)IG circuit tracing on Gemma-2-2B · SVA (4 tasks avg)circuit size: approximately 100x reductionImproved
replace Integrated Gradients with RelP attribution (ablation)MLP-activation neuron circuits · SVA pairedfaithfulness/completeness vs circuit size: near-perfect with ~200 neuronsImproved
replace IG with RelP attribution (ablation)residual-stream and MLP-output neuron bases · SVA pairedfaithfulnessImproved
replace IG with RelP attribution (ablation)SAE bases · SVA pairedfaithfulness/completenessMixed
use RelP instead of IG with a zero baseline (unpaired data) (ablation)MLP-activation neuron circuits · SVA unpaired, mean ablationfaithfulness/completenessImproved
use RelP with stop-gradients instead of IG-inp for edge attribution (ablation)MLP-activation neuron circuits · SVA edge-based circuitsfaithfulness: over 80% faithfulness with ~10^5 edgesImproved
remove stop-gradients on intermediate MLPs in RelP edge attribution (ablation)RelP edge attribution · SVA edge-based circuitsfaithfulness/completenessHurt
use RelP instead of IG for gradients in EAP/EAP-IG (ablation)MIB circuit localization, Llama 3.1 · MIB (5 tasks)CMD: best on three tasks, second-best on twoImproved
increase SAE width (16k to 65k) (ablation)SAE bases · SVA paired, Gemma-2-2Bcircuit sparsityNo clear effect
steer single neuron L23/N8079 (say a capital)Llama 3.1 8B Instruct · state-capitals multi-hop (50 questions)top output flipImproved
steer each of six neuron clusters in Texas circuitLlama 3.1 8B Instruct · Texas exampletop predictionMixed

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