Research map / Large language models / Language, safety and interpretability
Extrapolative Controlled Sequence Generation via Iterative Refinement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Scorer-guided selection (top-k sampling of 5, pick best by scorer) added to iterative editing (ablation) | ICE editor · Sentiment control (Yelp) | extrapolation-region success rate average: 0.376 -> 0.610 | Helped |
| Scorer guidance added to mask-and-infill iterative sampling | Iterative Sampling · Sentiment control (Yelp) | extrapolation-region success rate average: 0.056 (Sampling) -> 0.324 | Helped |
| Replacing trained ICE editor with MLM mask-and-infill (iterative sampling) (ablation) | Iterative editing pipeline · Sentiment control (Yelp) | extrapolation-region success rate average: 0.610 -> 0.324 | Hurt |
| Trained local editor (ICE) instead of Genhance latent perturbation, scorer-free | Sentiment control · Sentiment control (Yelp) | extrapolation-region success rate average: 0.376 vs 0.387 | No clear effect |
| Masking span size variant (Small vs Super Large) for synthetic pair creation (ablation) | ICE editor · Sentiment control (Yelp) | extrapolation-region success rate average: 0.429 (Small) -> 0.610 (Super Large) | Helped |
| Increasing Genhance samples n from 1 to 100 scored by f_s (ablation) | Genhance · Sentiment control (Yelp) | extrapolation-region success rate average: 0.044 (n=1) -> 0.550 (n=100), still below ICE 0.610 | Helped |
| Scorer guidance added to ICE editor (ablation) | ACE2 stability · ACE2 ddG extrapolation to -5 | success rate: 0.062 -> 0.361 | Helped |
| ICE editor instead of Genhance (both with scorer) | ACE2 stability · ACE2 ddG extrapolation to -5 | success rate: 0.159 -> 0.361 | Helped |
| Scorer guidance added to ICE editor (ablation) | AAV fitness · AAV fitness extrapolation to fitness 0 | success rate: 0.188 -> 0.223 | Helped |
| Scorer-free beam size varied (3, 5, 10) at iteration 10 (ablation) | ICE Scorer-Free on ACE2 · ACE2 stability | success rate at ddG -5: 0.059 to 0.062 range | No clear effect |
| Top-k varied (5, 10, 15) for scorer-guided sampling at iteration 10 (ablation) | ICE w/ Scorer on ACE2 · ACE2 stability | success rate at ddG -5: 0.249 to 0.362 | Mixed |
| Scorer guidance on AAV for Iterative Sampling | Iterative Sampling · AAV fitness | success rate at fitness 0: 0.018 (Sampling) -> 0.064 | Helped |
Closest papers by meaning
- Learning Extrapolative Sequence Transformations from Markov Chains (ICML 2025 · similarity 0.88)
- Controllable Protein Sequence Generation with LLM Preference Optimization (AAAI 2025 · similarity 0.87)
- Diffusion on Language Model Encodings for Protein Sequence Generation (ICML 2025 · similarity 0.86)
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.86)
- Steering Protein Language Models (ICML 2025 · similarity 0.85)
- ProSpero: Active Learning for Robust Protein Design Beyond Wild-Type Neighborhoods (NeurIPS 2025 · similarity 0.85)
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.