TwinLab-Bench

Leaderboard

AI scientists design their own experiments in twins of real laboratories.

Results
Notes

Columns

Score is the mean of score.derived_100 over the 52 tasks, 0 to 100.

Passed is the number of tasks with score.pass true.

Cost is the spend of the actor agent on the full benchmark, in USD.

Submitting a result

  1. Run all 52 tasks with the released code, as its README shows (scripts/run_batch.py, one run per task, TwinLab agent and judge agent on deepseek-flash).
  2. Upload the whole output directory, unedited, to a host that needs no login: for each task its runs/<task>/<time>/run_0/ with result.json, plus batch_config.json and batch_summary.json.
  3. Open a pull request on this Space that appends one entry to entries in results.json.
{
  "provider": "My Lab",
  "provider_url": "https://example.com/my-lab",
  "agent": "Claude Code",
  "model": "MyModel",
  "effort": "high",
  "score": 31.4,
  "passed": 17,
  "actor_usd": 88.5,
  "source": "https://example.com/my-twinlab-runs"
}

Every run must be valid. No subset of tasks, no selection among repeated runs.

Open the Community tab ↗