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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add the learned bias vector to a scale-only adapter (scalar affine) (ablation) | self-interpretation adapter · Llama Scope SAE layer 19 | validation loss: 4.543 -> 1.787 | Improved |
| add a low-rank term to scalar affine adapter (rank 64) (ablation) | scalar affine adapter · Llama Scope SAE layer 19 | validation loss: 1.787 -> 1.619 | Improved |
| use full-rank affine adapter instead of scalar affine on SAE features (ablation) | adapter · Llama Scope SAE layer 19 | validation 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 19 | validation loss: 2.00 -> 1.69 | Improved |
| use full-rank instead of scalar affine on Wikipedia contrastive vectors (ablation) | adapter · Wikipedia topics recall@1 | recall@1 (%): 79.4 -> 93.7 | Improved |
| use full-rank instead of scalar affine when generalizing to SAE latents (ablation) | Wikipedia-trained adapter · Llama Scope SAE generation hit rate | hit rate (%): 39.3 -> 22.1 | Hurt |
| train adapter vs untrained SelfIE on Goodfire SAE (70B) | Llama-3.3-70B-Instruct · Goodfire SAE generation scoring | hit rate (%): 48.1 -> 69.7 | Improved |
| use trained adapter instead of LoRA fine-tuning (rank 4) | Llama-3.1-8B-Instruct · Goodfire SAE generation hit rate | hit rate (%): 62.3 vs 59.2 | No clear effect |
| train adapter on contrastive topic vectors | Qwen-2.5 7B-72B · topic recall@100 vs Taboo ceiling | recall@100 | Improved |
| trained adapter vs untrained SelfIE for bridge entity detection | Llama-3.1-8B-Instruct · TwoHopFact | detected bridge entities (%): 56.4 -> 91.0 | Improved |
| SA+LR vs SA adapter on SAE detection scoring (ablation) | adapter · Llama Scope detection F1 | detection F1: 0.722 vs 0.706 | No clear effect |
| trained adapters vs repeating auto-interp training labels | Goodfire SAE (8B) · generation hit rate | hit rate (%): 50.8 -> 59.2 | Improved |
Closest papers by meaning
- Training Language Models to Explain Their Own Computations (NeurIPS 2026 · similarity 0.87)
- Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants (NeurIPS 2026 · similarity 0.87)
- SelfIE: Self-Interpretation of Large Language Model Embeddings (ICML 2024 · similarity 0.87)
- Think in Latent, Explain in Language: Self-Explainable Latent Reasoning (ICML 2026 · similarity 0.86)
- I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders (AAAI 2026 · similarity 0.86)
- SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures (NeurIPS 2024 · similarity 0.86)
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.
