RetroAgent is a 4B-parameter LLM agent for multi-step retrosynthesis planning. It decomposes a target molecule into commercially available building blocks by searching over an AND-OR graph of molecules and reactions, driven entirely by tool calls. The policy is trained end-to-end with reinforcement learning (GSPO) on top of the Qwen/Qwen3-4B-Instruct-2507 backbone, using the slime framework (Megatron + SGLang) with a reward shaped around search-budget efficiency rather than final success alone.

Paper: RetroAgent: Harnessing LLMs to Search Over Structured Memory for Agentic Retrosynthesis Planning (COLM 2026)
Code: https://github.com/SXKDZ/RetroAgent
Model: https://huggingface.co/SXKDZ/RetroAgent

Results

Pass@1 and success rate (%) at different search budgets N on USPTO-190 and ChEMBL-1000, reported as means over multiple runs at temperature 0. N is the number of single-step model calls; pass@1 is the fraction of targets solved on the first rollout. Numbers below match the paper's main table; the strongest LLM baseline, Retro-R1 (V1), is shown for comparison.

USPTO-190

Method	pass@1	N=50	N=100	N=200	N=300	N=400	N=500
Retro-R1 (V1, 7B)	50.00	64.63	71.32	78.11	80.89	82.95	84.32
RetroAgent (4B)	53.26	58.42	70.53	78.63	81.58	83.68	85.05

ChEMBL-1000

Method	pass@1	N=50	N=100	N=200	N=300	N=400	N=500
Retro-R1 (V1, 7B)	68.52	73.92	78.00	80.36	81.36	81.98	82.42
RetroAgent (4B)	73.82	76.36	79.62	81.96	83.22	83.94	84.40

With a 4B model and no reranking by model probabilities, RetroAgent outperforms the 7B Retro-R1 on both benchmarks. The advantage is largest on the out-of-distribution ChEMBL-1000 set (+5.30 pass@1, +1.98 at N=500), indicating that the structured-memory interface yields transferable planning strategies rather than dataset-specific memorization.

Intended Use

RetroAgent is an agent policy, not a standalone chat model. It perceives and acts only through tool calls served by the RetroAgent tool server (AND-OR graph memory, single-step template model, building-block lookup, molecular scoring). To reproduce the reported results, run the model behind the tool server and evaluation harness from the code repository; prompting it as a plain chat model outside the harness will not perform retrosynthesis.

The weights load with transformers as a standard Qwen3ForCausalLM (no trust_remote_code required):

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("SXKDZ/RetroAgent", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("SXKDZ/RetroAgent")

Training

Backbone: Qwen3-4B-Instruct-2507 (non-thinking mode).
Algorithm: GSPO (sequence-level clipping) under the slime RL framework, on a single 8xH200 node.
Reward: base +1 for a complete route, reduced by soft penalties on search budget and route depth, plus small per-turn penalties for invalid or failed tool calls.

Citation

bibtex

@inproceedings{zhu2026retroagent,
  title     = {RetroAgent: Harnessing {LLMs} to Search Over Structured Memory for Agentic Retrosynthesis Planning},
  author    = {Zhu, Yanqiao and Gan, Jingru and Sun, Xiaoqi and Sun, Fang and Shi, Yidan and Islam, Md Mofijul and Shang, Chao and Gao, Wenhao and Coley, Connor W. and Sun, Yizhou and Wang, Wei},
  booktitle = {Conference on Language Modeling (COLM)},
  year      = {2026}
}

License

Released under the Apache License 2.0, inherited from the Qwen3-4B-Instruct-2507 base model. The RetroAgent code is released separately under the MIT license (see the code repository).

Iambackup/RetroAgent

RetroAgent