AgentPersonaBench
Harbor: pendingA benchmark of 2,460 tasks across 867 traits evaluating latent persona adherence via observable actions in survey, chat, web, and app environments.
- Samples
- 2,460
- Released
- 2026
- Scores
- 7
Autonomous Science Bench
Autonomous Science Bench: Evaluating AI Agents for Autonomous Scientific Discovery
Evaluates AI agents on six budgeted design-build-test-learn campaigns in biology, chemistry, and materials science using discovery and learning gates.
How to run
ltb · autonomous-science-bench
$ npm i -g lettokenburn-cli$ ltb bench info autonomous-science-bench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
8 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 67.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT-5.6-terra | 61.1 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT-5.6 Luna | 27.8 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Claude Sonnet 5.5 | 11.1 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GLM-5.3 Flash | 11.1 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| MiniMax-M3 | 5.6 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Qwen3.8 27B | 5.6 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| Claude Haiku 4.5 | 0.0 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
The benchmark comprises six campaigns grounded in published experiments across biology, chemistry, and materials science. Agents iteratively select experimental designs under a fixed budget via a stateful lab service.
Success requires passing two calibrated gates: a discovery gate evaluating measured designs and a learning gate evaluating the agent's predictive model. In pilot tests, the best configuration succeeded in 67% of trials.