Research map / Large language models / Instruction and fine-tuning
STRIDE: Post-Training LLMs to Reason and Refine Bio-Sequences via Edit Trajectories
Daiheng Zhang, Shiyang Zhang, Sizhuang He, Yangtian Zhang, Syed A Rizvi, David van Dijk
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 edit-trajectory supervision (full STRIDE) over direct final-sequence generation (ablation) | Qwen3-14B · GFP replace-only | novelty: 23 -> 53 | Improved |
| add structured edit tokens without rationale over direct generation (ablation) | Qwen3-14B · GFP replace-only | novelty: 23 -> 40 | Improved |
| add free-form rationale on top of structured edit tokens (ablation) | STRIDE · GFP replace-only | success: 48 -> 61 | Improved |
| use STRIDE instead of vanilla SFT, full action space | Qwen3-14B · GFP full action | success (of 100): 42 -> 89 | Improved |
| use STRIDE instead of vanilla SFT, replace-only | Qwen3-14B · GFP replace-only | unique improved: 39 -> 59 | Improved |
| use STRIDE instead of Edit Flow on AAV | Edit Flow · AAV full action | success: 52 -> 73 | Improved |
| add GSPO post-training to STRIDE-SFT (ablation) | STRIDE-SFT · DrugAssist | strict success: 0.579 -> 0.676 | Improved |
| add GSPO post-training to STRIDE-SFT (ablation) | STRIDE-SFT · DrugAssist | shift rate: 0.983 -> 0.755 | Improved |
| add edit trajectories to GSPO RL (ablation) | Vanilla GSPO · DrugAssist | shift rate: 0.865 -> 0.755 | Improved |
| add explicit trajectory supervision in SFT without RL (ablation) | Vanilla SFT · DrugAssist | strict success: 0.629 -> 0.579 | Hurt |
| use GRPO or CISPO instead of GSPO for STRIDE RL (ablation) | STRIDE · DrugAssist | shift rate: 0.755 -> 0.956 | Hurt |
| closed-loop 3x30 refinement instead of one-shot 1x90 | STRIDE · GFP replace-only | novelty: 84.3% -> 97.0% | Improved |
Closest papers by meaning
- pCoMole: Pareto-Constrained Molecule Editing with Discrete Flows (NeurIPS 2026 · similarity 0.87)
- Steering Protein Language Models (ICML 2025 · similarity 0.86)
- Generalists vs. Specialists: Evaluating LLMs on Highly-Constrained Biophysical Sequence Optimization Tasks (ICML 2025 · similarity 0.86)
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.85)
- Fine-Tuning Discrete Diffusion Models via Reward Optimization with Applications to DNA and Protein Design (ICLR 2025 · similarity 0.84)
- Improved Off-policy Reinforcement Learning in Biological Sequence Design (ICML 2025 · similarity 0.84)
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.
