AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Towards Automated Circuit Discovery for Mechanistic Interpretability

Arthur Conmy, Augustine N. Mavor-Parker, Aengus Lynch, Stefan Heimersheim, Adrià Garriga-Alonso

NeurIPS 2023 · 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
Minimize task-specific metric instead of KL divergence (ablation)ACDC · Docstring edge ROCAUC: 0.982 -> 0.972No clear effect
Minimize task-specific metric instead of KL divergence (ablation)ACDC · Greater-Than edge ROCAUC: 0.853 -> 0.461Hurt
Minimize task-specific metric instead of KL divergence (ablation)ACDC · IOI edge ROCAUC: 0.869 -> 0.589Hurt
Patch with zero activations instead of corrupted activations (ablation)ACDC · Tracr-Reverse edge ROCAUC: 0.200 -> 1.000Improved
Patch with zero activations instead of corrupted activations (ablation)ACDC · Tracr-Proportion edge ROCAUC: 0.679 -> 1.000Improved
Patch with zero activations instead of corrupted activations (ablation)ACDC · IOI edge ROC, KLAUC: 0.869 -> 0.539Hurt
Patch with zero activations instead of corrupted activations (ablation)Subnetwork Probing · Docstring edge ROC, KLAUC: 0.937 -> 0.428Hurt
Use ACDC instead of SP and HISPACDC · IOI edge ROC, KL, corruptedAUC: 0.823 -> 0.869Improved
Use ACDC instead of HISP when optimizing task lossACDC · Tracr-Reverse edge ROC, corruptedAUC: 0.577 -> 0.200Hurt
Use ACDC instead of SP/HISP, number of edges above 20ACDC · Induction, KL on held-out setKL divergence per edgesImproved
Use zero activations instead of corrupted activations (ablation)ACDC, SP, HISP · InductionKL divergenceImproved
Evaluate on reset (permuted weight) network versus trained network (ablation)HISP and SP · Induction reset networkKL divergenceHurt

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