Retrosynthesis stops guessing reactants: an 8B model disconnects bonds first and lifts feasible template diversity from 2.02 to 3.43 on rare reactions

Retrosynthesis models begin to write the disconnection strategy into the reasoning chain instead of jumping from product to reactants. Earlier molecular reasoning models mostly stayed at generic product-level analysis such as functional-group enumeration and mechanism explanation, omitting the disconnection steps that link product analysis to a specific reactant choice. On the 500-instance ORDerly main test set, RetroReasoner (RL) reaches Exact@1 of 0.526 and Round-trip@1 of 0.826, against 0.486 and 0.802 for the same-architecture Prediction-Only (RL) that drops explicit reasoning.

The reasoning chain consists of four tagged structured steps joined by natural-language linking texts. SyntheticRetro starts from RXN SMILES, gathers directly readable, model-predicted and rule-derived information, and uses GPT-oss-20B to write linking texts that connect product analysis, candidate substructure, disconnection and synthetic equivalents into one trainable rationale. Generation is repeated 15 times per reaction instance to obtain multiple reasoning paths; training uses full-parameter BFloat16 tuning on eight NVIDIA H100 GPUs, and the RL stage rewards predictions whose reactants reconstruct the product through a forward model.

The round-trip reward gives positive signal to reactants beyond the single reference. With exact match to one reference reactant as the only reward, chemically valid but differently written reactants receive no reward. Replacing the round-trip reward with exact-match reward moves RetroReasoner (RL) from Exact@100 of 0.724 to 0.696 and Template Diversity from 3.186 to 2.430, while Feasible Ratio rises from 0.786 to 0.812.

Plugging the single-step model into parallelized Monte Carlo tree search cuts search time while raising route diversity. An LLM-based single-step model must be queried at every expansion, which becomes the main bottleneck of sequential multi-step planning. On the 500-target PaRoutes set-n1 and set-n5 subsets, average solving time drops nearly sixfold as workers increase, Solved Targets is largely maintained, and Routes Extracted and Number of Clusters rise gradually.

The demonstrated scope is the ORDerly-derived single-step evaluation set, two rare-reaction subsets, the external MHNreact and TempRe benchmarks, and multi-step planning on 500 targets each from PaRoutes set-n1 and set-n5; the model targets organic small molecules, and the authors position it as a strategic single-step reasoner that plugs into a search framework rather than an end-to-end route generator.

The authors note that the model ignores temperature, humidity and pressure of real synthesis, handles mainly organic small molecules rather than polymers and crystals, and reasons at the level of reaction mechanisms without extending to electron transfer; round-trip consistency remains a proxy for laboratory feasibility, and the model may propose hazardous or toxic compounds, so it is positioned for research use at this stage.