sxkdz/RetroAgent
RetroAgent is a 4 billion parameter LLM agent developed by SXKDZ for multi-step retrosynthesis planning, built on the Qwen3-4B-Instruct-2507 backbone. It excels at decomposing target molecules into commercially available building blocks by navigating an AND-OR graph using tool calls. The model is trained with reinforcement learning (GSPO) to optimize for search-budget efficiency, outperforming larger models like Retro-R1 (7B) on retrosynthesis benchmarks like USPTO-190 and ChEMBL-1000.
Loading preview...
RetroAgent: LLM for Retrosynthesis Planning
RetroAgent is a 4 billion parameter language model agent specifically designed for multi-step retrosynthesis planning. Developed by SXKDZ, it leverages the Qwen3-4B-Instruct-2507 backbone and is trained using reinforcement learning (GSPO) to efficiently search over an AND-OR graph of molecules and reactions.
Key Capabilities & Differentiators
- Agentic Retrosynthesis: Decomposes complex target molecules into commercially available building blocks through a series of tool calls.
- Search-Budget Efficiency: The policy is trained with a reward function that prioritizes efficient search, not just final success, leading to more optimized planning.
- Performance: Despite being a 4B model, RetroAgent outperforms the 7B Retro-R1 baseline on both USPTO-190 and ChEMBL-1000 benchmarks, particularly showing a significant advantage on the out-of-distribution ChEMBL-1000 set (+5.30 pass@1).
- Structured Memory Interface: Its design with a structured-memory interface yields transferable planning strategies rather than dataset-specific memorization.
Intended Use
RetroAgent functions as an agent policy and is not a standalone chat model. It requires a dedicated tool server (providing AND-OR graph memory, single-step template model, building-block lookup, and molecular scoring) and an evaluation harness to reproduce its reported performance. It is ideal for research and applications in chemical synthesis planning where efficient and accurate retrosynthesis is critical.