LetTokenBurnEurekaBench: Measuring Agentic Ability to Discover New Scientific Insights
A cross-domain benchmark with 26 long-horizon tasks across six scientific domains that evaluates AI agents' ability to conduct iterative experiments, discover mechanisms, and derive novel scientific insights.
$ npm i -g lettokenburn-cli$ ltb bench info eurekabench
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
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Opus 4.8 | 0.1 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GPT-5.6 Sol | 0.1 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Content & Construction: EurekaBench is an expert-verified set of 26 long-horizon tasks spanning neuroscience, computer science, chemistry, astrophysics, geophysics, and plasma physics. It challenges AI agents to start from initial observations challenging existing understanding, conduct iterative experiments autonomously via a provided environment interface, investigate underlying mechanisms, and derive novel scientific insights.
Scoring & Key Findings: The evaluation framework assesses three axes: adherence to scientific constraints (common-sense assumptions), predictive accuracy of discovered mechanisms on test sets, and the ability to yield meaningful scientific insights. Empirical results show that while current AI agents can approach human levels in predictive accuracy, they substantially underperform humans in deriving novel scientific insights.