Designing reasoning agents that can integrate external tools for complex chemical tasks remains a fundamental challenge. Recent chemical agents have shown promise by attaching tools, yet they rely on autoregressive architectures constrained by causal attention and left‑to‑right generation, which hampers bidirectional molecular understanding, global tool planning, and multi‑source evidence integration.
To overcome these issues we introduce ChemDiffAgent, the first masked diffusion language agent for tool‑integrated chemical reasoning. The approach reformulates reasoning as iterative denoising over interleaved reasoning and tool‑use trajectories, allowing tool‑use decisions to be jointly determined under full‑trajectory context and aligning the decision boundary with the knowledge boundary.
We analyze how this paradigm benefits training through harder infilling sub‑problems and enables flexible arbitrary‑order decoding during inference. A two‑stage post‑training pipeline is then built, consisting of agentic supervised fine‑tuning and variance‑reduced preference optimization, to boost both tool usage and reasoning capabilities.
A comprehensive benchmark covering six chemistry tasks, with both single‑turn and multi‑turn tool scenarios, is constructed. Under matched budgets and evaluation protocols ChemDiffAgent consistently outperforms autoregressive agents, delivering roughly a 20% overall gain. Notably, the 8B model reaches performance comparable to, and sometimes surpassing, much larger frontier models such as GPT‑5.5 and Claude Opus‑4.7.
Code is available at https://github.com/bilemond/ChemDiffAgent
Review