PaperlayerView exact revision on arXiv

Source-linked AI summary

Chem-R: Learning to Reason as a Chemist

Weida Wang, Benteng Chen, Di Zhang, Wanhao Liu, Shuchen Pu, Ben Gao, Jin Zeng, Xiaoyong Wei, Tianshu Yu, Shuzhou Sun, Tianfan Fu, Wanli Ouyang, Lei Bai, Jiatong Li, Zifu Wang, Yuqiang Li, Shufei Zhang

arXiv:2510.16880v2cs.CE

TL;DR

Current LLMs lack core chemical knowledge and reliable, structured reasoning for complex chemical problems. Chem-R combines chemical foundation training, reasoning-protocol distillation, and multi-task optimization; it reports state-of-the-art performance across chemical benchmarks, including a 32% gain over Gemini-2.5-Pro on Name Prediction (Exact Match).

  • Problem

    Current LLMs lack chemical fundamentals and structured, reliable reasoning processes for complex chemical problems.

    Source: §1 · p. 1

  • Method

    Chem-R trains through chemical foundation training, protocol-guided reasoning distillation, and multi-task optimization across molecular- and reaction-level tasks.

    Source: §1 · p. 2, §3.3 · p. 5

  • Results

    Across nine macro-tasks and 25 sub-tasks, Chem-R consistently achieves state-of-the-art performance; it improves 32% over Gemini-2.5-Pro on Name Prediction (Exact Match).

    Source: §1 · p. 3

  • Takeaways & Limitations

    Chem-R produces interpretable, logically coherent, chemically sound reasoning chains, a quality validated by human expert evaluation.

    Source: §5 · p. 9

Abstract

from arXiv · show

Although large language models (LLMs) have significant potential to advance chemical discovery, current LLMs lack core chemical knowledge, produce unreliable reasoning trajectories, and exhibit suboptimal performance across diverse chemical tasks. To address these challenges, we propose Chem-R, a generalizable Chemical Reasoning model designed to emulate the deliberative processes of chemists. Chem-R is trained through a three-phase framework that progressively builds advanced reasoning capabilities, including: 1) Chemical Foundation Training, which establishes core chemical knowledge. 2) Chemical Reasoning Protocol Distillation, incorporating structured, expert-like reasoning traces to guide systematic and reliable problem solving. 3) Multi-task Group Relative Policy Optimization that optimizes the model for balanced performance across diverse molecular- and reaction-level tasks. This structured pipeline enables Chem-R to achieve state-of-the-art performance on comprehensive benchmarks, surpassing leading large language models, including Gemini-2.5-Pro and DeepSeek-R1, by up to 32% on molecular tasks and 48% on reaction tasks. Meanwhile, Chem-R also consistently outperforms the existing chemical foundation models across both molecular and reaction level tasks. These results highlight Chem-R's robust generalization, interpretability, and potential as a foundation for next-generation AI-driven chemical discovery. The code and model are available at https://github.com/davidweidawang/Chem-R.

1 INTRODUCTION

Chem-R addresses chemical LLMs’ gaps in foundational knowledge, structured reasoning, and balanced performance with a three-phase training framework. It reports state-of-the-art results across molecular- and reaction-level tasks.

1 INTRODUCTION

  • Current LLMs may lack chemical fundamentals and structured, reliable reasoning, limiting their performance and interpretability on complex chemical problems.The introduction identifies errors in molecular representations and reaction rules, unsystematic reasoning, and imbalanced performance across tasks.Source: §1 · pp. 1–2
  • Chem-R combines chemical foundation training, protocol-distilled reasoning, and multi-task optimization to support structured, generalizable reasoning across molecular and reaction tasks.The framework uses large-scale non-reasoning corpora, modular protocols distilled from a teacher, and curriculum-like task weighting.Source: §1 · p. 2
  • Across nine macro-tasks and 25 sub-tasks, Chem-R consistently achieves state-of-the-art performance, with gains on both Name Prediction and Yield Prediction.Against Gemini-2.5-Pro, it improves 32% on Name Prediction (Exact Match) and 50% on Yield Prediction (Accuracy).Source: §1 · p. 3

2 RELATED WORK

Prior work explores reasoning supervision and chemistry-specific language models, while highlighting challenges in reliable reasoning and specialized scientific data. Chem-R is positioned against those gaps.

2.1 REASONING FOR LLMS

  • Outcome-based supervision can produce unstructured, inconsistent reasoning, while process-level feedback and verifier models offer alternatives.The literature also describes chain-of-thought prompting, rejection sampling, reward modeling, preference learning, and rule-based reinforcement learning.Source: §2.1 · p. 3

2.2 LLMS FOR CHEMISTRY

  • Chemistry LLMs cover tasks from molecular captioning to reaction analysis, but specialized scientific training data remains comparatively scarce.The related work presents Chem-R’s pipeline for synthesizing high-quality chain-of-thought data as a response to this data bottleneck.Source: §2.2 · p. 3

3 METHOD

Chem-R uses three training phases to establish chemical knowledge, distill structured reasoning, and improve performance balance across molecular- and reaction-level tasks.

3 METHOD

  • The three-phase pipeline moves from chemical foundation training to protocol distillation and multi-task reinforcement learning.Its stated aim is to transform ad-hoc reasoning traces into chemically sound, interpretable, and broadly generalizable reasoning.Source: §3 · pp. 3–4, Figure 2 · p. 4

3.1 PHASE 1: CHEMICAL FOUNDATION TRAINING

  • Phase 1 fine-tunes on paired chemical inputs and labels, aligning molecular descriptors and encoding reaction transformations, reagents, and conditions.The corpus covers molecular- and reaction-level supervision, grounding the model in chemically valid input-output behavior.Source: §3.1 · p. 4

3.2 PHASE 2: CHEMICAL REASONING PROTOCOL DISTILLATION

  • Phase 2 distills reusable, stepwise Chemical Reasoning Protocols into student-model training data through protocol-guided chain-of-thought synthesis.The teacher derives protocols from correct and incorrect reasoning paths, adds cautionary guidance, and synthesizes traces with correct information; only traces whose regenerated answer matches the original are retained for fine-tuning.Source: §3.2 · p. 5

3.3 PHASE 3: MULTI-TASK GRPO

  • Phase 3 weights task sampling toward weaker tasks, then applies multi-task GRPO to refine protocol-guided reasoning across heterogeneous chemical tasks.The sampling probabilities use validation accuracy after Phase 2; token updates use a clipped-ratio objective with KL regularization and task-specific accuracy rewards, without format-based rewards.Source: §3.3 · pp. 5–6

4 EXPERIMENTS

The experiments compare Chem-R with general, reasoning, chemical-foundation, and specialist models across four benchmarks, nine macro-tasks, and 25 subtasks. Ablations examine the three training phases, protocol components, sampling choices, and out-of-distribution and expert evaluations.

4.2 MAIN RESULTS

  • Chem-R’s 8B model sets a new state of the art across diverse benchmarks; Yield Prediction reaches 0.85, and Retrosynthesis reaches 0.39 versus the strongest baseline’s 0.15.ChemDFM-v1.5-8B scores higher on Molecule Design, while Chem-R provides interpretable, step-by-step reasoning chains.Source: Table 1 · p. 7, §4.2 · p. 7

4.3 ABLATION STUDY

  • Removing Foundation Training drops Name Prediction from 0.49 to 0.14; removing Phases 1 and 2 reduces Reaction Prediction to zero, while Multi-task GRPO raises Reagent Selection from 0.50 to 0.69.The study identifies Phase 1 with chemical fundamentals, Phase 2 with reasoning, and Phase 3 with refinement.Source: §4.3 · p. 7
  • Removing the Phase 2 CRP lowers Retrosynthesis from 0.28 to 0.20, while removing Correct Information drops Reaction Prediction from 0.69 to 0.13.The authors describe the CRP as providing reasoning structure and Correct Information as grounding the reasoning-chain training data.Source: §4.3 · pp. 7–8
  • Phase 2’s Multi-task model exceeds Single-task models on Reagent Selection (0.50 vs. 0.46) and Retrosynthesis (0.28 vs. 0.26), while a Single-task model scores 0.75 on Reaction Prediction.Source: §4.3 · p. 8
  • Name Prediction and Molecule Design gains diminish after sample size k reaches approximately 10, while cross-task mixing achieves high performance with k=5.The authors note this can help on complex tasks where successful reasoning paths are scarce.Source: §4.3 · p. 8, Figure 3 · p. 8
  • With the same training steps, weighted sampling raises Retrosynthesis from 0.33 to 0.39 and Reagent Selection from 0.63 to 0.69 versus uniform sampling.The weighting allocates more training focus to tasks the model finds more difficult.Source: §4.3 · p. 8
  • On four out-of-distribution molecule-optimization tasks absent from training, Chem-R raises Solubility success rate from 10% to 83%.Source: §4.3 · pp. 8–9, Figure 3 · p. 8
  • Chemistry PhD reviewers rated Chem-R highest across six reasoning dimensions, including Chemical Soundness, Logical Coherence, and Expert-level Insight.The evaluation averaged ratings of 50 randomly sampled responses from Chem-R and three baselines on a 1-to-5 scale.Source: §4.3 · p. 9

5 CONCLUSION

Chem-R combines a three-phase training framework with interpretable reasoning to establish state-of-the-art chemical reasoning, particularly on complex reaction prediction and retrosynthesis. The appendix overview maps where task descriptions, implementation details, results, and further examples appear.

5 CONCLUSION

  • The human expert evaluation is visualized as a radar chart.Figure 4 presents the evaluation chart.Source: Figure 4 · p. 9, §5 · p. 9
  • Chem-R establishes state-of-the-art chemical reasoning, with particularly strong gains in complex reaction prediction and retrosynthesis.The conclusion also reports interpretable, logically coherent, and chemically sound reasoning chains, validated by human expert evaluation.Source: §5 · p. 9
  • Chem-R’s three-phase framework builds foundational knowledge, distills structured reasoning protocols, and optimizes balanced performance across tasks.The conclusion describes this framework as emulating expert chemists’ deliberative thought processes.Source: §5 · p. 9

A OVERVIEW OF THE APPENDIX

  • The appendix overview directs readers to task descriptions, implementation details, quantitative results, additional examples, and the stated role of LLMs.It also notes that the results section covers the described tasks and out-of-distribution tasks.Source: A OVERVIEW OF THE APPENDIX · p. 15

B TASK DESCRIPTION

The experiments evaluate molecular- and reaction-level capabilities using tasks selected from four chemistry benchmarks for practical relevance and diversity.

B TASK DESCRIPTION

  • The task suite draws on ChEBI-20, ChemLLMBench, USPTO, and TOMG-Bench.The benchmarks were selected for practical relevance and diversity in evaluating molecular- and reaction-level capabilities.Source: B TASK DESCRIPTION · p. 15
  • Molecular tasks span molecule generation and property prediction.The section frames these as applications for evaluating molecular-level capabilities.Source: B TASK DESCRIPTION · p. 15
  • Reaction tasks span reaction prediction and retrosynthesis.Together with the molecular tasks, these applications are used for a comprehensive evaluation of chemical reasoning performance.Source: B TASK DESCRIPTION · p. 15

B.1 MOLECULAR TASKS

The molecular task suite tests representations, predictions, generation, and descriptions across individual molecules. It also includes text-based open molecule generation, editing, and optimization.

B.1 MOLECULAR TASKS

  • Name prediction converts between IUPAC names and SMILES, requiring interpretation of molecular structures and chemical conventions.The task covers conversion in either direction.Source: §B.1 · p. 15
  • Property prediction classifies molecules by biological activity or toxicity, including BBBP, HIV, Tox21, ClinTox, and BACE.The task tests reasoning about structure–property relationships and generalization across molecular structures.Source: §B.1 · p. 15

Molecule Design

  • Molecule design maps a textual description to a corresponding SMILES representation.Descriptions may specify features such as functional groups or molecular size.Source: §B.1 · p. 16

Molecule Captioning

  • Molecule captioning translates a SMILES string into a natural-language description of key molecular features.The description should capture structural features such as functional groups and chemical bonding.Source: Molecule Design · p. 16

B.1 MOLECULAR TASKS

  • Text-based open molecule generation evaluates creating, editing, and optimizing molecular structures against textual criteria.Optimization can target LogP, QED, or MR while navigating trade-offs between properties.Source: Molecule Captioning · p. 16

B.2 REACTION TASKS

The reaction task suite evaluates yield prediction, reagent selection, product prediction, and retrosynthesis. These tasks require reasoning about reaction context, conditions, transformations, and pathways.

B.2 REACTION TASKS

  • Yield prediction classifies Buchwald–Hartwig and Suzuki-coupling reactions as high- or low-yielding from reactants and conditions.For the supplied Buchwald–Hartwig example, high yield is defined as greater than 70%.Source: §B.2 · p. 17

Reagent Selection

  • Reagent selection chooses reactants, solvents, and ligands from a predefined list for a given reaction.The task was chosen over USPTO-Condition because options have associated yields, making reasoning more verifiable and practically constrained.Source: §B.2 · p. 17

Reaction Prediction

  • Reaction prediction infers products from reactants and conditions using knowledge of reactivity, mechanisms, and stereochemistry.The task requires applying chemical principles beyond pattern recognition.Source: Reagent Selection · p. 18

Retrosynthesis

  • Retrosynthesis predicts starting materials for a target by deconstructing it into simpler components and considering plausible reaction pathways.The task also involves identifying suitable bond-breaking reactions and reagents.Source: Reaction Prediction · p. 18

C IMPLEMENTATION DETAILS

Table 3 reports hyper-parameters for each training phase and inference, including compute, optimization, batch-size, and epoch settings.

C.1 HYPER-PARAMETERS

  • The listed settings include GPU counts, learning rates, batch sizes, and epoch counts.Source: §C.1 · p. 19
  • Table 3 covers hyper-parameters for Phases 1, 2, and 3, as well as inference.Source: Table 3 · p. 19, §C.1 · p. 19

C.2 DATA STATISTICS

This section outlines phase-specific training-data allocations, test-set separation, and TOMG exclusion, then presents structured distillation prompts for stepwise reasoning across chemical prediction, molecule translation and modification, and reaction tasks. The prompts use structural or constraint analysis, task-specific evidence where available, and validation of outputs.

C.2 DATA STATISTICS

  • Phase 1 uses large question-answer datasets (920k for Name Prediction), Phase 2 targets about 100k examples per major task, and Phase 3 allocates data based on post-Phase-2 performance.Phase 2 uses higher volumes for especially difficult tasks such as Name Prediction.Source: §C.2 · p. 20
  • No test-set molecules appear in training, and TOMG tasks are excluded from Phase 3 because of long evaluation times and better performance.Source: §C.2 · p. 20

C.3 DISTILLATION PROMPTS

  • For SMILES-to-IUPAC naming, prompts guide stepwise structure analysis, parent selection, numbering and substituent naming, then assembly of the name with stereochemical descriptors.The prompt supplies functional-group hints and directs reasoning to proceed as if deriving the answer from scratch.Source: §C.3 · p. 22
  • For IUPAC-to-SMILES generation, prompts have the model decompose the name, formulate molecular connectivity, translate components into SMILES, and assemble a string with branches, ring closures, and stereochemistry.Source: §C.3 · p. 23
  • Property-prediction prompts elicit judgments using molecular features and activity relationships, then require a holistic assessment and a definitive prediction.For BACE1, the prompt considers pharmacophores, physicochemical properties, structure-activity relationships, and possible binding liabilities; the BBBP prompt assesses penetration-relevant properties and active efflux.Source: §C.3 · pp. 24–25
  • Clinical-trial toxicity prompts combine structural alerts and physicochemical risks with FDA approval status, treating approval as evidence that risks were clinically acceptable.They ask for a definitive Yes/No prediction while balancing structural risks against approval status.Source: §C.3 · p. 26
  • HIV-inhibition prompts assess inhibitor-like scaffolds, target interactions, and molecular properties, then use CA, CM, or CI activity results to confirm or override structural predictions.CA or CM strongly supports Yes, whereas CI strongly supports No, regardless of favorable structural features.Source: §C.3 · p. 27
  • Tox21 toxicity prompts evaluate reactive or metabolically activated motifs, physicochemical risks, and possible cellular damage, while treating alerts as risk indicators rather than proof.The holistic assessment weighs multiple alerts against molecular context that may mitigate toxicophore reactivity.Source: §C.3 · p. 28
  • Molecule-design prompts translate descriptions and functional-group hints into SMILES by identifying the scaffold, mapping substituents and locants, and encoding specified stereochemistry.They direct the model to note when stereochemistry is absent and to construct syntax with branches, ring closures, and stereochemical markers as appropriate.Source: §C.3 · p. 29
  • Structure-to-description prompts parse SMILES backbones, branching, heteroatoms, functional groups, and cyclic systems to produce a detailed chemical description.The instructions emphasize accounting for every specified functional group.Source: §C.3 · p. 30
  • Molecular prompts guide description, generation, editing, and optimization through structured analysis, scaffold construction or targeted modification, and checks of constraint satisfaction and chemical validity.Description prompts synthesize scaffold, functional-group, and stereochemical findings into a cohesive narrative; generation prompts validate the final SMILES, while editing and optimization prompts check the intended change and its rationale.Source: §C.3 · pp. 31–33

Reaction Prediction

  • Reaction-prediction prompts analyze reactant structures and functional groups, infer a likely reaction pathway, then construct product SMILES with regioselectivity and stereoselectivity considered.If multiple products are predicted, the prompt requires separating them with periods rather than commas.Source: Reaction Prediction · p. 34

Retrosynthesis

  • Retrosynthesis prompts analyze product structure and functional groups, select a plausible disconnection and forward reaction, then derive stable, charge-neutral reactant SMILES.Multiple predicted reactants are to be separated with periods rather than commas.Source: Retrosynthesis · p. 35

C.4 HUMAN EVALUATION

The human evaluation scores generated reasoning across six dimensions, while the following results subsections specify evaluation setups for name prediction, property prediction, and molecule design. The evaluation also notes that Chem-R’s protocol-based training may make its reasoning stylistically distinct, so experts rated anonymized responses.

C.4 HUMAN EVALUATION

  • Experts score Chem-R and baseline reasoning from 0–5 on six dimensions, with Table 5 highlighting the best and second-best model in each column.The dimensions are Chemical Soundness, Logical Coherence, Step-by-Step Completeness, Justification of the Conclusion, Clarity and Conciseness, and Expert-level Insight.Source: §C.4 · p. 36, Figure 5 · p. 36, Table 5 · p. 36
  • Blind, fully anonymized ratings were used to mitigate potential bias from Chem-R’s often more structured, protocol-based reasoning style.Experts did not know which model generated each response.Source: §C.4 · p. 36

D.1 NAME PREDICTION

  • Name prediction evaluates SMILES-to-IUPAC and IUPAC-to-SMILES translation using exact-match accuracy.The results section introduces the task-specific evaluations; Table 6 is captioned as reporting accuracy scores.Source: §D.1 · p. 37, Table 6 · p. 37

D.2 PROPERTY PREDICTION

  • Property prediction covers BACE, BBBP, ClinTox, HIV, and Tox21 as binary classification tasks, using classification accuracy as the sole metric.Tables 7 and 8 are captioned as reporting AUC-ROC and accuracy scores, respectively.Source: §D.2 · p. 37, Table 7 · p. 37, Table 8 · p. 37

D.3 MOLECULE DESIGN

  • Molecule design evaluates text-to-SMILES generation on ChEBI-20 using validity, string-fidelity metrics, and fingerprint similarity to assess structural resemblance.String metrics include exact match, Levenshtein distance, and BLEU; structural similarity uses MACCS, RDKit, and Morgan fingerprints.Source: §D.3 · p. 38, Table 9 · p. 38

D.4 MOLECULE CAPTIONING

Molecule captioning evaluates how accurately and fluently a model describes ChEBI-20 molecular structures in natural language. The assessment combines precision- and recall-oriented text-overlap measures with METEOR.

D.4 MOLECULE CAPTIONING

  • Captioning on ChEBI-20 is assessed with BLEU-2 and BLEU-4, ROUGE-1, ROUGE-2, and ROUGE-L F-scores, and METEOR.BLEU measures n-gram precision; ROUGE measures unigram, bigram, or longest-common-subsequence overlap, while METEOR incorporates synonymy and stemming.Source: §D.4 · pp. 38–39

D.5 TEXT-BASED OPEN MOLECULE GENERATION

The section defines task-specific evaluation for text-based molecule generation and four reaction-level tasks. Text-based generation combines success with similarity or novelty, while the reaction tasks use classification or exact-match measures.

D.5 TEXT-BASED OPEN MOLECULE GENERATION

  • TOMG-Bench evaluates molecule editing, optimization, and customized generation using task-specific success, similarity or novelty, weighted success rate, and validity measures.Editing and optimization use similarity, while customized generation uses novelty; weighted success rate averages performance across nine subtasks.Source: §D.5 · pp. 39–40

D.6 YIELD PREDICTION

  • Yield prediction classifies reactions as high- or low-yield on Buchwald-Hartwig and Suzuki-Miyaura datasets, using accuracy and reporting the average across datasets.Table 12 is captioned as reporting accuracy scores.Source: §D.6 · p. 41, Table 12 · p. 41

D.7 REAGENT SELECTION

  • Reagent selection uses Suzuki HTE data to predict reactants, solvents, and ligands, with top-1 accuracy for reactants and solvents and top-5 accuracy for ligands.The table’s average combines the three accuracy scores.Source: §D.7 · p. 41, Table 13 · p. 41

D.8 REACTION PREDICTION

  • Reaction prediction evaluates major-product prediction on USPTO Mixed using exact-match accuracy between predicted and reference product SMILES.Table 13 includes this task alongside reagent selection and retrosynthesis.Source: §D.8 · p. 41, Table 13 · p. 41

D.9 RETROSYNTHESIS

  • Retrosynthesis evaluates reactant prediction on USPTO-50k, counting a prediction as correct only when the predicted reactant SMILES set exactly matches the reference set.Table 13 includes this task alongside reagent selection and reaction prediction.Source: Table 13 · p. 41, §D.9 · p. 42

D.10 OUT OF DOMAIN

The out-of-domain evaluation tests molecular property optimization across four targets. Chem-R improves both reported measures over its base model and is reported to significantly outperform a much larger LLM.

D.10 OUT OF DOMAIN

  • Across four optimization targets, Chem-R improves SR% and ∆ over Llama-3.1-8B-Instruct and significantly outperforms the larger Llama-3.3-70B-Instruct.The targets are Solubility, DRD2, JNK3, and GSK3-β; SR% measures successful increases in the target property, while ∆ denotes mean property improvement.Source: Table 14 · p. 42, §D.10 · p. 42

E MORE CASES

The examples contrast chemically grounded reasoning with trajectories that misread structures or invoke invalid mechanisms, sometimes reaching a correct answer despite flawed reasoning. They also include successful molecule-design and reaction-prediction examples and an LLM-use disclosure.

E MORE CASES

  • A reaction-prediction example correctly identifies tert-butoxide and the acyl chloride and uses nucleophilic acyl substitution to explain the product.It also dismisses the ether solvent as a primary participant.Source: E MORE CASES · p. 43
  • A benzyne mechanism is chemically incorrect here, even though the reasoning reaches the correct product; tert-butoxide instead attacks the carbonyl directly.The critique identifies nucleophilic acyl substitution as the actual pathway.Source: E MORE CASES · p. 44
  • Misreading an acyl chloride as an aryl chloride leads to an inapplicable benzyne mechanism and an incorrect reaction product.The critique identifies nucleophilic acyl substitution as the actual reaction.Source: E MORE CASES · p. 45
  • The molecule-design example correctly builds the N-carbamoylaspartic acid structure and omits explicit stereochemistry because the description does not specify it.The critique describes the reasoning as systematic and chemically accurate.Source: E MORE CASES · p. 46
  • A second molecule-design response reaches the correct SMILES only by coincidence, relying on a fictitious rule to merge separate substituents.The critique says the carbamoyl group is attached to the amino nitrogen, not separately to the same carbon.Source: E MORE CASES · p. 47
  • One molecule-design response reaches the correct connectivity despite mislabeling the carbamoyl group; the critique distinguishes it from the final N-carbamoylamino group.The critique identifies the final group as an N-carbamoylamino or ureido group.Source: E MORE CASES · p. 48

F USE OF LLMS

  • The authors report using LLMs to improve language and readability, then reviewing and editing the content and taking full responsibility.Source: F USE OF LLMS · p. 48

Related papers

Loading 2510.16880v2…