Benchmark / dataset 2026
TSBench: physics-grounded reaction-mechanism evaluation
TSBench is a benchmark for evaluating whether LLM agents can construct three-dimensional transition-state guesses consistent with chemical reaction mechanisms.
Workflow level n/a Source: Preprint (not peer-reviewed)
Research applicability
Offers chemical-mechanism researchers an evaluation design beyond factual chemistry questions, with a reaction-path consistency failure mode that matters for proposed transition states.
Getting started
Untested editorial guidance: read https://arxiv.org/abs/2609.08503 and the full methods before adaptation; start with a bounded supervised evaluation and inspect artifacts, permissions and failure modes. This review is not an exercised recipe.
Maintenance
Source retrieved 2026-10-17; software maintenance and availability not tested.
License
Article, model, data and software rights require separate inspection; no open-licence claim.
Workflow & human oversight
Resource-level workflow classification unspecified; an instantiated workflow depends on the research task, implementation and oversight. Researchers retain responsibility for validation and safe use.
Workflow level is not an evidence score; higher is not better evidence.
Approach
Agents use structure-editing tools and an automated quantum-chemical pipeline supplies pass/fail judgments. The authors report diagnosis-driven revision experiments and failures where plausible saddle points connect the wrong reactants and products.
Limitations
Abstract-only review of arXiv v1. Quantum-chemical settings, benchmark implementation, model configurations and reaction set were not inspected. No downloadable dataset or executable release is confirmed. Automated checking inside the authors' benchmark is not editorial machine-verification or independent reproduction; transfer to synthesis planning is not demonstrated.
Reproducibility
Primary abstract and metadata read only; no installation, code execution, experiment or independent outcome reproduction.
Adoption requirements
Untested editorial guidance: inspect the full benchmark protocol and artifact availability, establish quantum-chemistry expertise and compute requirements, and validate transition-state/reaction-path checks before comparing agents. Check model/data licensing separately.
Comments
No comments yet.
Add a comment
Comments are held for moderation before appearing. Critical or corrective comments are welcome — especially evidence that an entry overstates what has been verified.