AI Paper Atlas

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

ICML 2026 · Reinforcement learning › MDPs and planning · approach: Robust MDP policy gradient

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 changedWhereMeasuredEffect
add scaffold-aware step-level advantage (SA-GRPO) instead of vanilla GRPO (ablation)RetrOrchestrator · Retro*-190success rate: 61.9565% -> 94.2105%Improved
use SA-GRPO instead of GiGPO (ablation)RetrOrchestrator · Retro*-190success rate: 74.2105% -> 94.2105%Improved
use GiGPO (discounted rewards) instead of vanilla GRPO (ablation)RetrOrchestrator · Retro*-190success rate: 61.9565% -> 74.2105%Improved
add curriculum stage 2 after stage 1 (two-stage curriculum vs single stage) (ablation)RetrOrchestrator · Retro*-190success rate: 65.4255% / 72.2222% -> 94.2105%Improved
stage 1 training only (chemical rule adherence) (ablation)Qwen2.5-1.5B-Instruct agent · Retro*-190action validity: 12.1% -> 61.9%Improved
RL training of the LLM agent vs off-the-shelf (ablation)Qwen2.5-1.5B · Retro*-190success rate: 9.5% -> 94.2%Improved
scaffold-level grouping instead of molecule-level grouping (ablation)SA-GRPO · Retro*-190success rate: 74.21 -> 94.21Improved
scaffold-level grouping instead of search-graph-level grouping (ablation)SA-GRPO · Retro*-190success rate: 88.42 -> 94.21Improved
too loose or too strict scaffold similarity threshold (ablation)SA-GRPO · Retro*-190success rate: 0.95: 94.21; 0.80: 71.58; 0.999: 73.68Hurt
learned LLM router over molecule, SSR and prediction count vs non-LLM state-dependent routerRetrOrchestrator · Retro*-190success rate: 82.63 -> 94.21Improved
dynamically orchestrate SSR models vs static ensemble of all SSRsRetrOrchestrator · Retro*-190success rate: 77.8947% -> 94.2105%Improved
use scaffold clustering instead of exact molecule matching (ablation)SA-GRPO · training credit assignmentsingleton cluster rate: 20% -> under 6%Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy