TrialAtlasBench

TrialAtlasBench is a benchmark for evaluating AI systems on clinical trial design reasoning, grounded in authoritative regulatory feedback from FDA Complete Response Letters (CRLs). It comprises three tasks built over 192 unique clinical trials.

Tasks

Config	Task	Format	Size
task1_deficiency_detection	Trial design deficiency detection	Multi-label multiple choice (10 options)	36
task2_optimization	Actionable trial design optimization	Multi-label multiple choice (10 options)	35
task3_ptrs_prediction	Prediction of technical and regulatory success	Binary classification	146

• Task 1 asks a system to select all applicable FDA-aligned design deficiencies for a trial.
• Task 2 asks a system to select the appropriate protocol-level optimizations for a trial.
• Task 3 asks a system to predict whether a trial will obtain regulatory approval (approved / unapproved).

Data fields

Every task uses the same schema:

• _id (string): unique instance identifier.
• nctid (string): the ClinicalTrials.gov identifier of the trial.
• prompt (string): the task prompt, containing the structured trial design and the question with its answer options (the NCT ID is not included in the prompt).
• answer (string): the ground-truth answer.
  • Tasks 1 and 2: a comma-separated list of option IDs (e.g., "A" or "B, D, F").
  • Task 3: "approved" or "unapproved".

Usage

from datasets import load_dataset

ds = load_dataset("linjc16/TrialAtlasBench", "task1_deficiency_detection", split="test")
print(ds[0])

Evaluation

• Tasks 1 and 2 (multiple choice): for each question, compare the selected option set S with the correct set G; precision = |S ∩ G| / |S|, recall = |S ∩ G| / |G|, and F1 is their harmonic mean. Report the mean over all questions.
• Task 3 (binary classification): report balanced accuracy, F1, and Cohen's kappa to account for class imbalance (106 approved, 40 unapproved).
• Post-release subset: task3_post_release_nctids.txt lists the 56 Task 3 trials whose ClinicalTrials.gov results were first posted after the release of GPT-5 (7 August 2025), used in the paper's pretraining-exposure analysis.

Data sources

All primary sources are publicly accessible: FDA Complete Response Letters and drug records via openFDA, clinical trial records via ClinicalTrials.gov, and biomedical literature via PubMed and PubTator.

Citation

@article{trialatlas,
  title  = {TrialAtlas: Multi-Agent Research Organization for Clinical Trial Design and Optimization},
  author = {Lin, Jiacheng and Wang, Zifeng and Chen, Zheng and Scott, Erick and Yang, Ziwei and Yu, Fanyang and Zhong, Sheng and Sun, Jimeng},
  year   = {2026}
}