Datalab-atom/Logos-1.5B

Logos: an evolvable reasoning engine for molecular science

Introduction

Meet Logos, an artificial intelligence model featuring an evolvable reasoning engine tailored for molecular science.

Logos comprehensively addresses a critical pain point in molecular design: it seamlessly integrates multi-step logical reasoning with strict chemical consistency. Previous models have often suffered from a severe imbalance, offering either high physical fidelity but lacking a transparent reasoning process (specialized models), or flexible linguistic reasoning capabilities without guarantees of chemical validity (general-purpose large language models). Logos breaks through this bottleneck, enabling artificial intelligence to truly serve as a reliable collaborator in real-world scientific design workflows.

To overcome the scarcity of reasoning data, we constructed a chain-of-thought (CoT) reasoning layer from scratch for existing molecule-caption pairs (e.g., ChEBI-20 and PCdes), bridging the gap between abstract descriptions and structural decisions.

Abandoning traditional unidirectional training, Logos employs an innovative three-cycle evolutionary strategy for closed-loop training: Cycle 1 (Self-Data Distillation), Cycle 2 (Supervised Fine-Tuning, SFT), and Cycle 3 (Molecule-focused Group Relative Policy Optimization, M-GRPO). By incorporating a bootstrapping mechanism within these cycles to enable continuous self-iteration of both data and the model, it not only delivers expert-level chemical design capabilities but also maintains exceptionally high deployment efficiency.

Core Capabilities:

• Transparent White-Box Reasoning: Before generating molecular structures, Logos explicitly exposes its intermediate logical reasoning steps within a dedicated <think> block. This allows researchers to directly inspect, evaluate, and trace the design logic behind every candidate molecule.

• Strict Guarantee of Chemical Validity: By directly incorporating chemical rules and invariants into the optimization objective during the reinforcement learning phase, Logos internalizes physical constraints. On mainstream benchmarks (such as ChEBI-20 and PCdes), its generated molecules achieve a chemical validity of approximately 99.9%, effectively eliminating the issue of invalid chemical structures frequently produced by general-purpose LLMs.

• Human-in-the-Loop Interactive Discovery: Moving away from "black-box" single-shot generation, Logos is tailored for iterative design. [cite_start]Users can input specific constraints and provide feedback (e.g., from experiments or simulations) after receiving the model's reasoning and molecules ; the model then leverages the reasoning in the <think> block to dynamically adjust its strategy for the next round.

• Complex Multi-Objective Optimization: When facing multiple, potentially conflicting real-world design constraints (e.g., simultaneously optimizing $log~D_{7.4}$, reducing solubility, and retaining a specific molecular scaffold), Logos exhibits high stability, precisely guiding the generated results into the target property space.

• Exceptional Efficiency at a Compact Scale: Powered by precise domain-aligned training and chemical reward mechanisms, the compact Logos model matches or even surpasses much larger general-purpose large language models (such as GPT-5, DeepSeek-14b, etc.) in structural accuracy, Exact Match, and chemical validity.

• Evolvable Self-Bootstrapping Mechanism: Logos possesses the ability to self-evolve from failed samples. The model can regenerate high-quality reasoning chains and correct molecular structures for previously failed tasks, continuously expanding the training data without manual annotation to achieve closed-loop, continuous self-improvement.

Pipeline:

1. Dual-Capability Fusion: Merges the chemical accuracy of specialized models with the reasoning of LLMs to achieve both structural validity and interpretability.

2. Self-Distillation & SFT: Generates chain-of-thought (CoT) data using a teacher model, followed by supervised fine-tuning specifically for molecular design.

3. Molecule-Focused GRPO: Employs reinforcement learning with chemical rewards to directly enforce physical constraints and structural validity.

Performance

Contact & Questions

For collaborations or inquiries, please contact haibin65535@gmail.com. You’re welcome to open an issue or join the discussion in this repository, we value your insights and contributions to Logos.

Stay tuned and join our community as we push the boundaries of AI-driven molecular science. Together, let’s make chemical AI more precise, smarter, and highly evolvable. 🤝