MolReasoner

Language models applied to chemistry have mostly taken one of two routes, and both have a characteristic failure mode. Prompting a general-purpose model works without training but leaves the model without domain-specific molecular semantics, so it describes compounds in plausible-sounding generalities. Supervised fine-tuning on molecule-text pairs fixes the vocabulary but rewards memorization over understanding: the model learns to reproduce descriptions it has seen and, when pushed off distribution, invents functional groups that are not present or emits strings that do not correspond to a real molecule.

MolReasoner, developed by DP Technology with collaborators at the AI for Science Institute, Shanghai Jiao Tong University, and Fudan University, attacks that problem by making the intermediate reasoning explicit. Built on Qwen2.5-7B-Instruct, it is trained in two stages. Mol-SFT teaches the model to produce a structured chain of thought about a molecule before answering, using reasoning traces distilled from GPT-4o and enriched with chemical structure knowledge. Mol-RL then refines that reasoning with reinforcement learning against a task-adaptive reward that scores the answer on several axes at once rather than as a single correct/incorrect signal.

The two tasks it targets are the standard pair for molecular language modeling: molecule captioning, which turns a structure into a natural-language description, and text-based de novo molecule generation, which goes the other way. This places it alongside text-to-molecule models such as BioT5+, with the distinguishing feature that its reasoning is visible and auditable.

#Key Features

• Explicit chemical reasoning traces: The model works through base structure, functional groups, and connectivity in readable steps before committing to an answer, so a chemist can inspect where a wrong output went wrong instead of treating it as a black box.
• Knowledge-enhanced chain-of-thought distillation: The supervised stage uses CoT data generated by GPT-4o whose prompts were enriched with structural statistics (ring counts, aromaticity, molecular weight) and extended functional group annotations, rather than raw molecule-text pairs.
• Task-adaptive multi-level rewards: Captioning is scored on a blend of language metrics, while generation is scored on a composite of fingerprint similarity, sequence-level similarity, fragment similarity, and functional group matching, addressing the hallucination modes specific to each task.
• SELFIES rather than SMILES: Molecules are represented in SELFIES so that generated strings decode to chemically valid structures by construction, an ablation the authors show outperforms an otherwise identical SMILES-based model.

#Technical Details

Mol-SFT trains on roughly 42,000 curated chain-of-thought samples derived from the ChEBI-20 training set: about 24,000 for text-based de novo generation and 18,000 for captioning, retained after filtering for structural validity. Mol-RL uses Group Relative Policy Optimization with the composite rewards described above, run for 15 epochs at batch size 256 and a sampling temperature of 0.7, consuming roughly 1,200 GPU hours.

On the ChEBI-20 test set the gains are large. For captioning, MolReasoner reaches BLEU-2 of 0.4394 and ROUGE-L of 0.4827, against 0.2792 and 0.3428 for the strongest fine-tuned baseline, and far ahead of GPT-4o and of Qwen2.5-72B-Instruct prompted directly. For text-based generation it reaches BLEU of 0.7832 versus 0.3382 for the best baseline, MACCS fingerprint similarity of 0.6762, functional-group match of 0.5390, and 96.55% chemical validity. Ablations attribute much of the gain to reward design: binary correct/incorrect rewards are insufficient for these tasks, and the chain-of-thought annotations in Mol-SFT contribute independently of the RL stage.

#Applications

The model suits work at the boundary between chemical structure and text: annotating compound libraries with natural-language descriptions, curating or checking descriptions drawn from the literature, and drafting candidate structures from a written specification of desired properties. Because the reasoning trace is part of the output, it is also usable as a teaching and auditing aid, letting a medicinal chemist see which structural features the model believed it was reasoning about before accepting or rejecting a suggestion.

#Impact

MolReasoner is evidence that reasoning-style post-training, which has reshaped general-purpose language models, transfers to chemistry when the reward is designed around the domain rather than borrowed wholesale. The authors' error analysis shows the model shifts its failures away from confident hallucination and toward local, diagnosable mistakes, which is the more useful failure mode for a scientific tool. The work remains a preprint, evaluation is confined to ChEBI-20, and the release consists of task-specific checkpoints for captioning and generation rather than a single general model. Weights are published on HuggingFace under an MIT license.