Benchmark

TrialDesignBench focuses on clinical trial design tasks that require agents to reason across medical evidence, statistical tradeoffs, patient safety, feasibility, and protocol documentation.

What Agents Receive

Each task should provide the materials a human trial designer would need: disease context, target population, candidate interventions, prior evidence, design constraints, and required deliverables.

What Agents Produce

Agents may be asked to draft protocol components, select design parameters, justify endpoints, identify risks, propose adaptations, or produce machine-checkable artifacts for downstream evaluation.

Evaluation Signal

The benchmark should combine deterministic checks, rubric-based clinical and statistical review, and structured artifact validation. A task should pass only when the full design is coherent, clinically defensible, and aligned with the stated constraints.

Release Status

The public benchmark specification, task examples, evaluation harness, and submission process are under development.