Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Logos: An evolvable reasoning engine for rational molecular design

Haibin Wen, Zhe Zhao, Fanfu Wang, Tianyi Xu, Hao Zhang, Chao Yang, Ye Wei

TL;DR

Logos is a compact molecular reasoning model that integrates multi-step logical reasoning with strict chemical consistency and achieves strong performance in both structural accuracy and chemical validity, matching or surpassing substantially larger general-purpose language models while operating with a fraction of their parameters.

Abstract

The discovery and design of functional molecules remain central challenges across chemistry,biology, and materials science. While recent advances in machine learning have accelerated molecular property prediction and candidate generation, existing models tend to excel either in physical fidelity without transparent reasoning, or in flexible reasoning without guarantees of chemical validity. This imbalance limits the reliability of artificial intelligence systems in real scientific design workflows.Here we present Logos, a compact molecular reasoning model that integrates multi-step logical reasoning with strict chemical consistency. Logos is trained using a staged strategy that first exposes the model to explicit reasoning examples linking molecular descriptions to structural decisions, and then progressively aligns these reasoning patterns with molecular representations. In a final training phase, chemical rules and invariants are incorporated directly into the optimization objective, guiding the model toward chemically valid outputs. Across multiple benchmark datasets, Logos achieves strong performance in both structural accuracy and chemical validity, matching or surpassing substantially larger general-purpose language models while operating with a fraction of their parameters. Beyond benchmark evaluation, the model exhibits stable behaviour in molecular optimization tasks involving multiple, potentially conflicting constraints. By explicitly exposing intermediate reasoning steps, Logos enables human inspection and assessment of the design logic underlying each generated structure. These results indicate that jointly optimizing for reasoning structure and physical consistency offers a practical pathway toward reliable and interpretable AI systems for molecular science, supporting closer integration of artificial intelligence into scientific discovery processes.

Logos: An evolvable reasoning engine for rational molecular design

TL;DR

Logos is a compact molecular reasoning model that integrates multi-step logical reasoning with strict chemical consistency and achieves strong performance in both structural accuracy and chemical validity, matching or surpassing substantially larger general-purpose language models while operating with a fraction of their parameters.

Abstract

The discovery and design of functional molecules remain central challenges across chemistry,biology, and materials science. While recent advances in machine learning have accelerated molecular property prediction and candidate generation, existing models tend to excel either in physical fidelity without transparent reasoning, or in flexible reasoning without guarantees of chemical validity. This imbalance limits the reliability of artificial intelligence systems in real scientific design workflows.Here we present Logos, a compact molecular reasoning model that integrates multi-step logical reasoning with strict chemical consistency. Logos is trained using a staged strategy that first exposes the model to explicit reasoning examples linking molecular descriptions to structural decisions, and then progressively aligns these reasoning patterns with molecular representations. In a final training phase, chemical rules and invariants are incorporated directly into the optimization objective, guiding the model toward chemically valid outputs. Across multiple benchmark datasets, Logos achieves strong performance in both structural accuracy and chemical validity, matching or surpassing substantially larger general-purpose language models while operating with a fraction of their parameters. Beyond benchmark evaluation, the model exhibits stable behaviour in molecular optimization tasks involving multiple, potentially conflicting constraints. By explicitly exposing intermediate reasoning steps, Logos enables human inspection and assessment of the design logic underlying each generated structure. These results indicate that jointly optimizing for reasoning structure and physical consistency offers a practical pathway toward reliable and interpretable AI systems for molecular science, supporting closer integration of artificial intelligence into scientific discovery processes.
Paper Structure (12 sections, 6 equations, 4 figures)

This paper contains 12 sections, 6 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Results
  3. Motivation and Three-Stage Training Framework
  4. Superior Performance across Multi-Dataset Benchmarks
  5. Evolutionary Training Mechanism, Ablation Analysis, and Structured Reasoning Format
  6. Interactive Application to Multi-Objective Molecular Optimization
  7. Discussion
  8. Trade-offs.
  9. Limitations.
  10. Scale versus reasoning.
  11. Future work.
  12. Methods

Figures (4)

  • Figure 1: Conceptual framework and training pipeline of Logos. a, Logos combines the chemical accuracy of specialized models (bottom lright) with the reasoning of general LLMs (top left), addressing both interpretability and structural validity. b, Three-stage pipeline: Cycle 1 (self-data distillation) builds chain-of-thought (CoT) data with a teacher model; Cycle 2 (supervised fine-tuning) trains on molecular design; Cycle 3 (molecule-focused GRPO) uses reinforcement learning with chemical rewards. c, Validity scores on ChEBI-20 and PCdes; Logos-1.5b (final) and Logos-4b approach $\sim$99.9%, outperforming larger baselines. d, Logos enables interactive molecular design through three paradigms
  • Figure 2: Benchmarking across Logos-1.5b versions (v1 to final), Logos-4b and general LLMs (e.g., DeepSeek-14b, GPT-5) on ChEBI-20 and PCdes. Logos-1.5b final reaches exact match 0.3406 (ChEBI-20) and validity $\sim$1.0. Structural similarity (MACCS, RDKit, Morgan) improves across versions; FCD decreases to 0.4795, indicating that generated molecules lie closer to real drug-like chemistry.
  • Figure 3: Evolutionary training and output format. a, Pipeline: teacher (LLM-14B) generates CoT for description--structure pairs (Cycle 1); student (LLM-1.5B) is fine-tuned on CoT data to become Logos-0 (Cycle 2); M-GRPO with chemical rewards yields Logos (Cycle 3). b, Ablations: full M-GRPO (solid blue) reaches EM $>$ 0.35; removing self-data distillation (w.o. SDD) lowers performance; M-GRPO stabilizes generation length. c, Output format: system prompt plus JSON; the model outputs reasoning in <think> then the molecule in JSON, e.g. for a trisaccharide task.
  • Figure 4: Interactive Application to Multi-Objective Molecular Optimization. a, Three exemplary paradigms for rational molecule design: (1) Translation, where the model converts a detailed textual description into an exact molecular representation through step-by-step structural assembly; (2) Optimization, which performs local edits on a given backbone to satisfy specific physicochemical property constraints (e.g., $\log D_{7.4}$ and solubility); and (3) Distillation, an iterative multi-turn process that refines candidate molecules from a fuzzy conceptual query, gradually converging on the desired molecule. b. Analysis of Logos' performance in real-world cases of multi-objectice optimization from velcicky2024discovery.