ChemDiffAgent Explores Chemical Reasoning with a Diffusion Language Model

Instead of building its response only from left to right, the agent refines a complete trajectory. It is evaluated on six chemistry tasks with external tools.

ChemDiffAgent reformulates reasoning and tool use as a succession of corrections to a partially masked trajectory. This approach aims at better coordination of tool choices and bidirectional reading of molecular information.

Two-step training strengthens the model's capacity for action and preferences. In six tasks and at comparable budgets, the authors report an overall gain of about 20% compared to the autoregressive agents evaluated. Their model of 8 billion parameters competes with some larger models on these tests, without demonstrating a general superiority in chemistry.

Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.