Research map / Probabilistic and scientific modeling / Molecules and proteins
Induction Meets Biology: Mechanisms of Repeat Detection in Protein Language Models
Gal Pomerants, Yaniv Nikankin, Anja Reusch, Tomer Tsaban, Ora Schueler-Furman, Yonatan Belinkov
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 |
|---|---|---|---|
| use approximate-repeat circuit on identical/synthetic tasks instead of their own circuits | ESM-3 circuits · Cross-task faithfulness | faithfulness: faithfulness over 1.0 on other tasks | Improved |
| apply synthetic or identical repeat circuits to approximate repeats | ESM-3 circuits · Approximate repeats | cross-task faithfulness | Hurt |
| switch task from identical/synthetic to approximate repeats | ESM-3 sufficient circuit size · Circuit discovery | fraction of components for 85% faithfulness: 14% -> 17% | Hurt |
| keep only top ~3% neurons per MLP layer (ablation) | ESM-3 approximate-repeat circuit · Approximate repeats | faithfulness: 3% of neurons recover 80/85% faithfulness | No clear effect |
| keep only ~7% neurons per MLP layer (ablation) | ESM-C approximate-repeat circuit · Approximate repeats | faithfulness: 7% of neurons achieves 80% faithfulness | No clear effect |
| increase repeat length from 1-10 to 26+ (ESM-3, substitution 25-50%) | ESM-3 · Approximate repeats | accuracy: 0.45 -> 0.96 | Improved |
| increase substitution rate from 0-25% to 25-50% (length 1-10, ESM-3) | ESM-3 · Approximate repeats | accuracy: 0.77 -> 0.45 | Hurt |
| increase substitution rate from 0-25% to 25-50% (length 1-10, ESM-C) | ESM-C · Approximate repeats | accuracy: 0.75 -> 0.51 | Hurt |
| add indel mutations to approximate repeats instead of substitutions only | ESM-3 · Approximate repeats | accuracy: 79.0% -> 42.9% | Hurt |
| add indel mutations to approximate repeats instead of substitutions only | ESM-C · Approximate repeats | accuracy: 79.7% -> 48.1% | Hurt |
| raise substitution rate of mutated repeat copy for induction-head attribution | induction heads, length 10 repeats · Identical natural repeats mutated | induction-head attribution: drops sharply beyond 30-40% substitution rate for length 10 vs approximately 70% for length 20 | Hurt |
Closest papers by meaning
- From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models (ICML 2025 · similarity 0.87)
- Steering Protein Language Models (ICML 2025 · similarity 0.86)
- LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling (NeurIPS 2026 · similarity 0.86)
- Training Compute-Optimal Protein Language Models (NeurIPS 2024 · similarity 0.86)
- Concept Bottleneck Language Models For Protein Design (ICLR 2025 · similarity 0.86)
- MutaPLM: Protein Language Modeling for Mutation Explanation and Engineering (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.
