Single-step Retrosynthesis via Reinforcement Learning from Forward Feedback
We propose Retrosynthesis via Reinforcement Learning from Forward Feedback (RetroRLFF), which optimizes reactant generation using both reference agreement and forward consistency with the target product.
ProofPaper ↗
Key points
- Single-step retrosynthesis is a one-to-many inverse problem, yet most models are trained and evaluated against a single recorded reactant set.
- Starting from a supervised decoder-only Transformer, RetroRLFF generates complete reactant candidates and uses a frozen forward reaction model to predict products.
- A sequence-level reward combines reference agreement and forward consistency while penalizing invalid outputs.
- Direct Advantage-Based Policy Optimization updates the policy without human preference annotations or a preference-trained reward model.
Sources (1)
- [1]Single-step Retrosynthesis via Reinforcement Learning from Forward FeedbackarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 10, 07:37 PM
We propose Retrosynthesis via Reinforcement Learning from Forward Feedback (RetroRLFF), which optimizes reactant generation using both reference agreement and forward consistency with the target product.
Single-step retrosynthesis is a one-to-many inverse problem, yet most models are trained and evaluated against a single recorded reactant set.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 10, 2026Getting Space Right: Agentic Rewards for Image Generation
- Oct 10, 2026vllm-sr/d3
- Oct 10, 2026Co-adaptive Rule Routing and Constraining for Structured Knowledge Retrieval
- Oct 9, 2026LiquidAI/LFM2.5-24B-A2B
- Oct 9, 2026Rethinking Relative Policy Differences Transfer of On-Policy Distillation
- Oct 9, 2026Learning Representations to Guide Pixel Diffusion Transformers