ChemDiffAgent preprint reports approximately 20% gain in chemistry tool reasoning

Researchers report in an arXiv preprint that ChemDiffAgent improves overall performance by approximately 20% over autoregressive agents under matched budgets and evaluation. Their benchmark covers six chemistry tasks with single-turn and multi-turn tool use. The agent repeatedly revises sequences containing reasoning and tool actions, allowing decisions to use context from across the sequence instead of generating strictly left to right. Its 8B model reportedly matches, and sometimes exceeds, larger frontier models. These comparisons concern the chemistry benchmark constructed by the researchers.