Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

TL;DR

Chem-R introduces a principled three-phase framework to endow LLMs with structured and generalizable chemical reasoning. Phase 1 grounds the model in chemical fundamentals, Phase 2 distills expert reasoning protocols into reusable workflows, and Phase 3 balances performance across diverse molecular and reaction tasks via a curriculum-like reinforcement strategy. Across four comprehensive benchmarks, Chem-R achieves state-of-the-art results and demonstrates strong interpretability through explicit reasoning traces, outperforming leading chemistry-focused models and generalists. The work provides a practical pathway to reliable AI-assisted chemical discovery with scalable reasoning and broad generalization.

Abstract

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.

Chem-R: Learning to Reason as a Chemist

TL;DR

Chem-R introduces a principled three-phase framework to endow LLMs with structured and generalizable chemical reasoning. Phase 1 grounds the model in chemical fundamentals, Phase 2 distills expert reasoning protocols into reusable workflows, and Phase 3 balances performance across diverse molecular and reaction tasks via a curriculum-like reinforcement strategy. Across four comprehensive benchmarks, Chem-R achieves state-of-the-art results and demonstrates strong interpretability through explicit reasoning traces, outperforming leading chemistry-focused models and generalists. The work provides a practical pathway to reliable AI-assisted chemical discovery with scalable reasoning and broad generalization.

Abstract

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.
Paper Structure (58 sections, 10 equations, 5 figures, 14 tables)

This paper contains 58 sections, 10 equations, 5 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Reasoning for LLMs
  4. LLMs for Chemistry
  5. Method
  6. Phase 1: Chemical Foundation Training
  7. Phase 2: Chemical Reasoning Protocol Distillation
  8. I. Chemical Reasoning Protocol Generation.
  9. II. Protocol-Guided Synthesis and Finetuning.
  10. Phase 3: Multi-Task GRPO
  11. Experiments
  12. Experimental Setup
  13. Benchmarks.
  14. Evaluation Metrics.
  15. Baselines.
  16. ...and 43 more sections

Figures (5)

  • Figure 1: Challenges and the proposed Chem-R solution. The left panel highlights three key deficiencies observed in current reasoning models. To overcome these limitations, we introduce a three-phase training framework, illustrated on the right. This strategy is designed to first build a solid chemical foundation (Phase 1), then instill correct, step-by-step reasoning pathways (Phase 2), and finally, optimize for balanced, multi-task proficiency (Phase 3).
  • Figure 2: The overall pipeline of Chem-R. The model is trained in three phases. 1) Chemical Foundation Training: Instills basic chemical knowledge using question-answer pairs. 2) Chemical Reasoning Protocol Distillation: Teaches structured reasoning by fine-tuning on protocol-guided CoT. 3) Multi-task GRPO: Refines reasoning skills across all tasks using reinforcement learning.
  • Figure 3: Comprehensive evaluation of Chem-R. (a) Molecule task performance in different phases. (b) Reaction task performance in different phases. (c) Effect of sample size ($k$) on performance. (d) Molecule task performance during phase 3. (e) Reaction task performance during phase 3. (f) Model performance comparison across OOD tasks in ChemCoTBench li2025beyond.
  • Figure 4: Radar chart of the human expert evaluation.
  • Figure 5: Human evaluation rubric for Chain-of-Thought quality. Experts are to score the generated reasoning on a 0-5 scale (0=worst, 5=best) across the six metrics provided: Chemical Soundness, Logical Coherence, Completeness, Justification, Clarity, and Expert-level Insight.