Science’s Last Exam
Evaluating AI agents on scientific work.
From understanding experiments to developing hypotheses, SLE tests five scientific capabilities through concrete tasks and inspectable outputs. Explore the task collection, evaluation protocols, and current results.
Research preview · Task counts and results remain provisional.
Manuscript checked · 24 October 2026. Counts remain provisional; six Hypothesis Discovery databases are counted separately.
The benchmark
Five benchmarks for scientific agents
Each sub-benchmark isolates a different capability. Together they cover stages of scientific work; they are not one end-to-end experiment. Select a task family to see its inputs, deliverables, and evaluation.
Research preview · Collection sizes remain provisional. The first four capabilities contribute 198 public tasks; the six Hypothesis Discovery database settings are counted separately. Full task packages are being prepared for release.
Current results
Public evaluation
| Capability | Mean score | Tasks scored | Errors | Coverage |
|---|
Research hypotheses
Hypothesis Discovery
Five models evaluated across four research domains. Inspire Rate measures the fraction of generated hypothesis versions with an actionable connection to later literature; it does not establish a verified discovery.
Inspire Rate
Actionable hypothesis versions / all generated versions
| Rank | Model | Generated | Actionable | Inspire Rate |
|---|
Execution & evaluation
SLE-Agent uses a shared interaction loop with the tools required by each capability. The final answer, code, saved artifact, or research tree is frozen before evaluation.
Read the methodology ↗- 01
Task & environment
Scientific inputs, the target outcome, and a defined set of available tools.
- 02
Actions & artifacts
The agent works in the environment and delivers an inspectable result.
- 03
Independent evaluation
Capability-specific checks assess the scientific requirements.
Public evaluation
Representative & Last-Exam
Two disjoint subsets: a broad sample of scientific tasks and a more demanding frontier. Hypothesis Discovery is evaluated separately on research databases.