Research map / Reinforcement learning / RL from feedback
Iterative Distillation for Reward-Guided Fine-Tuning of Diffusion Models in Biomolecular Design
Xingyu Su, Xiner Li, Masatoshi Uehara, Sunwoo Kim, Yulai Zhao, Gabriele Scalia, Ehsan Hajiramezanali, Tommaso Biancalani, Degui Zhi, Shuiwang Ji
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 VIDD instead of DDPO (PPO-style) | diffusion fine-tuning · IFNAR2 binder design | ipTM: 0.2403 -> 0.5090 | Improved |
| Use VIDD instead of DDPO | EvoDiff · Protein SS-match | pLDDT: 0.55 -> 0.82 | Improved |
| Use VIDD instead of DRAKES (reward backprop) | DNA diffusion · DNA Enhancer HepG2 | Pred-Activity: 6.44 -> 8.28 | Improved |
| Add Best-of-N sampling to VIDD (ablation) | VIDD · Molecule Docking Parp1 | docking score: 9.4 -> 12.1 | Improved |
| Add Best-of-N (N=128) to VIDD (ablation) | VIDD · IFNAR2 binder design | ipTM: 0.51 -> 0.66 | Improved |
| Update roll-out policy every step (K=1) rather than K=5 (ablation) | VIDD · DNA Enhancer HepG2 | Pred-Activity: 8.28 -> 7.06 | Hurt |
| Lazy update interval K=50 rather than 5 (ablation) | VIDD · DNA Enhancer HepG2 | Pred-Activity: 8.28 -> 7.23 | Hurt |
| Increase regularization alpha from 1.0 to 2.0 (ablation) | VIDD · DNA Enhancer HepG2 | Pred-Activity: 8.28 -> 7.26 | Hurt |
| Decrease alpha from 1.0 to 0.8 (ablation) | VIDD · DNA Enhancer HepG2 | Pred-Activity: 8.28 -> 8.21 | No clear effect |
| Use VIDD instead of DDPO | diffusion fine-tuning · DNA Enhancer HepG2 | 3-mer Corr: 0.398 -> 0.162 | Hurt |
| Use VIDD instead of DDPO | diffusion fine-tuning · Protein SS-match | Diversity: 0.52 -> 0.52 | No clear effect |
Closest papers by meaning
- Reward-Guided Iterative Refinement in Diffusion Models at Test-Time with Applications to Protein and DNA Design (ICML 2025 · similarity 0.92)
- Bridging Model-Based Optimization and Generative Modeling via Conservative Fine-Tuning of Diffusion Models (NeurIPS 2024 · similarity 0.91)
- Fine-Tuning Discrete Diffusion Models via Reward Optimization with Applications to DNA and Protein Design (ICLR 2025 · similarity 0.91)
- Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding (NeurIPS 2025 · similarity 0.90)
- On the Design Space of Discrete Diffusion Online Adaptation for Molecular Optimization (NeurIPS 2026 · similarity 0.89)
- Feedback Efficient Online Fine-Tuning of Diffusion Models (ICML 2024 · similarity 0.88)
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.
