MolLean Turns Chemical Reasoning into an Executable, Verifiable Process

Introduction

Large language models can produce fluent explanations of chemical reasoning without reliably carrying out the underlying structural operations. A model may miscount atoms during an edit, invent a functional group, or claim that a molecular property improved without actually measuring it. MolLean addresses this gap by representing reasoning as an executable trace tied to an explicit molecular state. Instead of asking readers to trust a plausible explanation, the framework makes each chemical commitment available for checking.

What the framework changes

• Molecular state is explicit. MolLean maintains a concrete state as reasoning proceeds. Graph edits, compound transformations, and assertions must be consistent with that state rather than existing only in generated prose.
• Verification is layered. A Lean-based kernel checks operations at multiple levels, from primitive graph edits to more complex transformations, as well as assertions over molecular readouts.
• Structural checks are separated from measurements. Its typed witness protocol distinguishes facts that the system can verify natively from measurements supplied by an external evaluator. A structural change and an improvement in a property therefore require different kinds of evidence.
• Interaction is stateful. MolLean supports one-pass execution and whole-trace refinement. MolLean-MCP adds an interactive tool interface through which the model receives feedback after each action. Rejected actions leave the molecular state unchanged, preventing an invalid step from silently contaminating later reasoning.

The paper’s examples include editing ethanol into 1-propanol, enforcing an exact 25-carbon constraint, optimizing QED with measured feedback, and recovering from a rejected phenyl attachment in an interactive episode. These demonstrations highlight an important distinction: verification is not merely a final score. It can constrain the next action while the trace is being generated.

Results and implications

The authors compare direct chain-of-thought with several MolLean modes on 8,800 matched examples spanning 12 tasks in molecular generation, editing, optimization, and reactions. With Qwen3.6-27B, whole-trace refinement outperformed direct chain-of-thought on all 12 tasks. MolLean-MCP complements this approach with stepwise feedback and achieved better results in reaction tasks according to the reported evaluation.

The broader contribution is a shift in what counts as a chemical answer. A useful system should not only state a plausible product or property; it should expose a process whose structural commitments can be executed, checked, rejected, and revised. That could make model-assisted molecule design and reaction reasoning easier to audit. The evidence should still be interpreted carefully: the supplied material reports the framework and benchmark results, while the project page says a code release is planned. Performance on broader datasets, real experimental loops, and other models remains an open question.

Hugging Face Daily Papers