Research map / Large language models / Agents, code and math
Training a Scientific Reasoning Model for Chemistry
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Trains ether0, a 24B chemistry reasoning model from Mistral-Small-24B via alternating SFT, specialist GRPO RL, distillation, and generalist GRPO on 640k verifiable chemistry tasks, outperforming frontier LLMs, experts, and chemistry models.
Problem
Whether reasoning models trained with RL can generalize beyond math and code to scientific domains, specifically chemical reasoning and molecular design tasks, with data efficiency.
Approach
Post-train a general instruction model for chemistry using RL with verifiable, experimentally grounded rewards across 375 tasks, interleaving specialist GRPO with distillation into a generalist, plus an advantage-based curriculum buffer, LLM-based problem rewriting, and a molecule-plausibility bloom-filter reward.
Key results
- ether0 reaches 70% accuracy on reaction prediction after 46,000 examples, versus Molecular Transformer's 64.1% on full USPTO; retrained MT stays below 30%.
- Zero-shot MCQ accuracy of ether0 is 50.1%, comparable to o1-2024-12-17 at 52.2%; with one-shot prompting ether0 surpasses all evaluated frontier models.
- Safety alignment yields refusal of 80% of unsafe questions without meaningfully degrading measured task capability.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with reinforcement learning on reaction prediction problems | ether0 · Reaction prediction held-out test set | accuracy | Helped |
| Train on the paper's reaction dataset instead of the full USPTO training data | Molecular Transformer · Reaction prediction | accuracy: failed to exceed 30% accuracy | Hurt |
| Provide one distractor as an in-context example | ether0 · Multiple-choice questions | accuracy | Helped |
| Remove reasoning thoughts from the distillation sequences (ablation) | ether0 · Evaluated chemistry tasks | performance | Hurt |
| Filter BAD-quality reasoning before distillation (ablation) | ether0 · Test set | performance: a few percentage points | Helped |
| Add explicit reasoning to the model (ablation) | ether0 · Solubility editing and SMILES completion | performance | No clear effect |
| Add explicit reasoning to the model (ablation) | ether0 · Molecular formula and functional group | performance | Helped |
Surprising findings
- Reasoning quality degraded during RL with emergent made-up chemicals, non-English characters, and nonsense formatting, requiring language-based filtering.
- Filtering non-English reasoning led the model to switch to Cyrillic characters, showing that language filters alone shift rather than remove the artifact.
Limitations
- Weak generalization outside training distribution, e.g., inorganic chemistry and crystal structures, since training focused on organic SMILES.
- RL reduced general instruction-following and multi-turn chat ability; no tool calling was included in training.
Experimental setup
Datasets: ChEMBL COCONUT ORD LlaSMol PubChem pyFUME BBB EveBio Fang ADME AqSolDB IUPAC pKa Photoswitches
Benchmarks: ether0 holdout chemistry evaluation set (18 tasks) Molecular Transformer USPTO reaction prediction
Metrics: accuracy data efficiency vs. training examples behavior counts refusal rate on unsafe questions
Baselines: o1 o3 DeepSeek-R1 Claude Gemini 2.5 Pro GPT-4.5 ChemDFM TxGemma Molecular Transformer Mistral-Small-24B-Instruct-2501
Builds on: GRPO DAPO Mistral-Small-24B-Instruct-2501 DeepSeek-R1 DeepSpeed ZeRO
Closest papers by meaning
- LatentChem: From Textual CoT to Latent Thinking in Chemical Reasoning (ICML 2026 · similarity 0.92)
- Beyond Chemical QA: Evaluating LLM's Chemical Reasoning with Modular Chemical Operations (NeurIPS 2025 · similarity 0.91)
- Structured Chemistry Reasoning with Large Language Models (ICML 2024 · similarity 0.90)
- SciInstruct: a Self-Reflective Instruction Annotated Dataset for Training Scientific Language Models (NeurIPS 2024 · similarity 0.90)
- From What to Why: A Multi-Agent System for Evidence-based Chemical Reaction Condition Reasoning (ICLR 2026 · similarity 0.88)
- General-Reasoner: Advancing LLM Reasoning Across All Domains (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.