AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Learning Self-Interpretation from Interpretability Artifacts: Training Lightweight Adapters on Vector-Label Pairs

Keenan Pepper, Alex McKenzie, Florin Pop, Stijn Servaes, Martin Leitgab, Michael Vaiana, Judd Rosenblatt, Michael S. A. Graziano, Diogo S de Lucena

ICML 2026 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes · 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
add the learned bias vector to a scale-only adapter (scalar affine) (ablation)self-interpretation adapter · Llama Scope SAE layer 19validation loss: 4.543 -> 1.787Improved
add a low-rank term to scalar affine adapter (rank 64) (ablation)scalar affine adapter · Llama Scope SAE layer 19validation loss: 1.787 -> 1.619Improved
use full-rank affine adapter instead of scalar affine on SAE features (ablation)adapter · Llama Scope SAE layer 19validation loss: 1.787 -> 1.743 (train loss 0.64)Hurt
keep scaled identity term (SA+LR instead of LR only) at rank 4 (ablation)low-rank adapter · Llama Scope SAE layer 19validation loss: 2.00 -> 1.69Improved
use full-rank instead of scalar affine on Wikipedia contrastive vectors (ablation)adapter · Wikipedia topics recall@1recall@1 (%): 79.4 -> 93.7Improved
use full-rank instead of scalar affine when generalizing to SAE latents (ablation)Wikipedia-trained adapter · Llama Scope SAE generation hit ratehit rate (%): 39.3 -> 22.1Hurt
train adapter vs untrained SelfIE on Goodfire SAE (70B)Llama-3.3-70B-Instruct · Goodfire SAE generation scoringhit rate (%): 48.1 -> 69.7Improved
use trained adapter instead of LoRA fine-tuning (rank 4)Llama-3.1-8B-Instruct · Goodfire SAE generation hit ratehit rate (%): 62.3 vs 59.2No clear effect
train adapter on contrastive topic vectorsQwen-2.5 7B-72B · topic recall@100 vs Taboo ceilingrecall@100Improved
trained adapter vs untrained SelfIE for bridge entity detectionLlama-3.1-8B-Instruct · TwoHopFactdetected bridge entities (%): 56.4 -> 91.0Improved
SA+LR vs SA adapter on SAE detection scoring (ablation)adapter · Llama Scope detection F1detection F1: 0.722 vs 0.706No clear effect
trained adapters vs repeating auto-interp training labelsGoodfire SAE (8B) · generation hit ratehit rate (%): 50.8 -> 59.2Improved

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