Research map / Reinforcement learning / MDPs and planning
RetrOrchestrator: A Multi-Step Retrosynthesis Agent Dynamically Orchestrating Single-Step Transition Models
Liao Chang, Luotian Yuan, Yiping Ke, Ying Wei
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 scaffold-aware step-level advantage (SA-GRPO) instead of vanilla GRPO (ablation) | RetrOrchestrator · Retro*-190 | success rate: 61.9565% -> 94.2105% | Improved |
| use SA-GRPO instead of GiGPO (ablation) | RetrOrchestrator · Retro*-190 | success rate: 74.2105% -> 94.2105% | Improved |
| use GiGPO (discounted rewards) instead of vanilla GRPO (ablation) | RetrOrchestrator · Retro*-190 | success rate: 61.9565% -> 74.2105% | Improved |
| add curriculum stage 2 after stage 1 (two-stage curriculum vs single stage) (ablation) | RetrOrchestrator · Retro*-190 | success rate: 65.4255% / 72.2222% -> 94.2105% | Improved |
| stage 1 training only (chemical rule adherence) (ablation) | Qwen2.5-1.5B-Instruct agent · Retro*-190 | action validity: 12.1% -> 61.9% | Improved |
| RL training of the LLM agent vs off-the-shelf (ablation) | Qwen2.5-1.5B · Retro*-190 | success rate: 9.5% -> 94.2% | Improved |
| scaffold-level grouping instead of molecule-level grouping (ablation) | SA-GRPO · Retro*-190 | success rate: 74.21 -> 94.21 | Improved |
| scaffold-level grouping instead of search-graph-level grouping (ablation) | SA-GRPO · Retro*-190 | success rate: 88.42 -> 94.21 | Improved |
| too loose or too strict scaffold similarity threshold (ablation) | SA-GRPO · Retro*-190 | success rate: 0.95: 94.21; 0.80: 71.58; 0.999: 73.68 | Hurt |
| learned LLM router over molecule, SSR and prediction count vs non-LLM state-dependent router | RetrOrchestrator · Retro*-190 | success rate: 82.63 -> 94.21 | Improved |
| dynamically orchestrate SSR models vs static ensemble of all SSRs | RetrOrchestrator · Retro*-190 | success rate: 77.8947% -> 94.2105% | Improved |
| use scaffold clustering instead of exact molecule matching (ablation) | SA-GRPO · training credit assignment | singleton cluster rate: 20% -> under 6% | Improved |
Closest papers by meaning
- RetroBridge: Modeling Retrosynthesis with Markov Bridges (ICLR 2024 · similarity 0.91)
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.90)
- Advancing Retrosynthesis with Retrieval-Augmented Graph Generation (AAAI 2025 · similarity 0.89)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.89)
- Retrosynthesis Planning via Worst-path Policy Optimisation in Tree-structured MDPs (NeurIPS 2025 · similarity 0.89)
- RetroOOD: Understanding Out-of-Distribution Generalization in Retrosynthesis Prediction (AAAI 2024 · similarity 0.89)
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.
