Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MiST: Understanding the Role of Mid-Stage Scientific Training in Developing Chemical Reasoning Models

Andres M Bran, Tong Xie, Shai Pranesh, Jeffrey Meng, Xuan Vu Nguyen, Jeremy Goumaz, David Ming Segura, Ruizhi Xu, Dongzhan Zhou, Wenjie Zhang, Bram Hoex, Philippe Schwaller

TL;DR

This work identifies latent solvability as a prerequisite for RL-based chemical reasoning in LLMs and introduces MiST, a mid-stage training pipeline combining in-domain continued pretraining and supervised fine-tuning. By jointly improving Symbolic Competence and Latent Chemical Knowledge, MiST elevates the model’s latent solvability, enabling RL to achieve substantial gains on diverse chemical tasks and to produce interpretable reasoning traces. The authors formalize the Symbolic Competence Score (SCS) and Chemical Competence Score (CCS) as predictive diagnostics, showing that pre-RL SCS strongly forecasts post-RL performance with correlations up to $ ho \,\approx\,0.64$, and establishing a practical threshold of SCS > 1.5 for RL effectiveness. The findings imply that domain-specific mid-stage training can unlock reasoning in smaller, compute-efficient models, with potential applicability beyond chemistry to other domains with symbolic data and verifiable rewards.

Abstract

Large Language Models can develop reasoning capabilities through online fine-tuning with rule-based rewards. However, recent studies reveal a critical constraint: reinforcement learning succeeds only when the base model already assigns non-negligible probability to correct answers -- a property we term 'latent solvability'. This work investigates the emergence of chemical reasoning capabilities and what these prerequisites mean for chemistry. We identify two necessary conditions for RL-based chemical reasoning: 1) Symbolic competence, and 2) Latent chemical knowledge. We propose mid-stage scientific training (MiST): a set of mid-stage training techniques to satisfy these, including data-mixing with SMILES/CIF-aware pre-processing, continued pre-training on 2.9B tokens, and supervised fine-tuning on 1B tokens. These steps raise the latent-solvability score on 3B and 7B models by up to 1.8x, and enable RL to lift top-1 accuracy from 10.9 to 63.9% on organic reaction naming, and from 40.6 to 67.4% on inorganic material generation. Similar results are observed for other challenging chemical tasks, while producing interpretable reasoning traces. Our results define clear prerequisites for chemical reasoning training and highlight the broader role of mid-stage training in unlocking reasoning capabilities.

MiST: Understanding the Role of Mid-Stage Scientific Training in Developing Chemical Reasoning Models

TL;DR

This work identifies latent solvability as a prerequisite for RL-based chemical reasoning in LLMs and introduces MiST, a mid-stage training pipeline combining in-domain continued pretraining and supervised fine-tuning. By jointly improving Symbolic Competence and Latent Chemical Knowledge, MiST elevates the model’s latent solvability, enabling RL to achieve substantial gains on diverse chemical tasks and to produce interpretable reasoning traces. The authors formalize the Symbolic Competence Score (SCS) and Chemical Competence Score (CCS) as predictive diagnostics, showing that pre-RL SCS strongly forecasts post-RL performance with correlations up to , and establishing a practical threshold of SCS > 1.5 for RL effectiveness. The findings imply that domain-specific mid-stage training can unlock reasoning in smaller, compute-efficient models, with potential applicability beyond chemistry to other domains with symbolic data and verifiable rewards.

Abstract

Large Language Models can develop reasoning capabilities through online fine-tuning with rule-based rewards. However, recent studies reveal a critical constraint: reinforcement learning succeeds only when the base model already assigns non-negligible probability to correct answers -- a property we term 'latent solvability'. This work investigates the emergence of chemical reasoning capabilities and what these prerequisites mean for chemistry. We identify two necessary conditions for RL-based chemical reasoning: 1) Symbolic competence, and 2) Latent chemical knowledge. We propose mid-stage scientific training (MiST): a set of mid-stage training techniques to satisfy these, including data-mixing with SMILES/CIF-aware pre-processing, continued pre-training on 2.9B tokens, and supervised fine-tuning on 1B tokens. These steps raise the latent-solvability score on 3B and 7B models by up to 1.8x, and enable RL to lift top-1 accuracy from 10.9 to 63.9% on organic reaction naming, and from 40.6 to 67.4% on inorganic material generation. Similar results are observed for other challenging chemical tasks, while producing interpretable reasoning traces. Our results define clear prerequisites for chemical reasoning training and highlight the broader role of mid-stage training in unlocking reasoning capabilities.
Paper Structure (56 sections, 8 equations, 13 figures, 22 tables, 3 algorithms)

This paper contains 56 sections, 8 equations, 13 figures, 22 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Post-training methods for reasoning
  4. Chemical language modeling
  5. Mid-stage domain adaptation
  6. LLM capability diagnostics
  7. Preliminaries
  8. Prerequisite 1: Symbolic Competence
  9. Token log-likelihood extraction
  10. Symbolic Competence Score
  11. Prerequisite 2: Latent Chemical Knowledge
  12. Post-training methods
  13. Supervised fine-tuning on reasoning traces
  14. Reinforcement learning from Scientific Feedback
  15. Downstream reasoning tasks
  16. ...and 41 more sections

Figures (13)

  • Figure 1: Multi-stage pipeline for training a chemical-reasoning language model. Step1 (MiST, 3.9 B tokens) Continued Pretraining exposes a general-purpose base model to a chemistry-centric corpus that interleaves plain text with compound & synthesis information. A subsequent 1 B-token supervised fine-tuning phase teaches three formats: (i) symbol-level molecular or material understanding, (ii) structure-aware question & answers, and (iii) chemical chain-of-thought (CoT). In Step2 the MiST backbone is further specialized with either RLSF (reinforcement learning from scientific feedback) or rSFT (reasoning-style supervised fine-tuning). A pool of candidate answers $(o_{1}, \ldots, o_{n})$ generated by the MiST model is scored by a task-specific reward model $(r_{1}, \ldots, r_{n})$; a group-computation module aggregates these signals to update the policy, iteratively refining the model into a Chemical Reasoning Model.
  • Figure 2: ChemBench sub-domain Accuracy (%). Results obtained based on Qwen-2.5-3B
  • Figure 3: Selection of SCS corruption rate. SCS measured at varying corruption rates for Qwen-3B (left) and Qwen-7B (right) across base, MiST, and MiST+RL model variants. A corruption rate of 0.2 is selected to balance discriminative power between model treatments while maintaining task difficulty.
  • Figure 4: Post-RL task performance vs. pre-RL SCS(0.2). Markers indicate model size (circles = 3B, squares = 7B) and prompt type (filled = no-cot (S1), stars = cot (S2)). Reference markers: 'x' = base Qwen, '+' = MiST+FT. Red markers with '$\star$' indicate the model trained specifically on that task. Colors: darker = 3B, brighter = 7B (magma palette).
  • Figure 5: SCS across larger base LLMs.
  • ...and 8 more figures