Science's Last Exam

Science's Last Exam (SLE) is a scientific-agent benchmark from the AItonomy Foundation. Site: https://sciences-last-exam.org/ . Core question: Can AI do research, end to end?

Researchers from Stanford, Oxford, Princeton, and others contribute in a personal capacity; materials note writing/review by PhD students and postdocs from about 34 institutions in five countries.

Highlights:

• 113 public tasks across 9 disciplines;
• 6 capability diagnostics plus 1 end-to-end research assessment (wet-lab execution still under construction/expansion);
• Capabilities include hypothesis discovery, protocol planning, scientific computing, scientific software operation, experiment observation, and more; closed-loop tasks require agents to keep forming hypotheses, choosing experiments, handling counterexamples, and submitting verifiable artifacts;
• Early reports: among ~8 frontier model setups, best composite ~37.5/100 (Claude Opus 5 + Claude Code); 0 on the true end-to-end research track in those reports;
• Public task packages and leaderboard evaluation planned around the 2026-11-06 launch window.

Problem

Most science benchmarks stop at QA, isolated sub-tasks, or paper reproduction with the solution given. SLE targets a full research loop—hypothesis → planning → compute/software → observation → (planned) wet-lab—with expert-authored hidden checks on whether results actually hold.

Structure

• Capability diagnostics: pinpoint which skill fails (literature/hypothesis, protocol, code, GUI software, video observation, etc.);
• Closed-loop end-to-end: whether skills chain into independently verifiable research outcomes;
• Tasks emphasize real research questions and inspectable artifacts, not multiple-choice scores.

Limits

• End-to-end and wet-lab parts are still expanding; some frontier models may be pending on the leaderboard;
• Scores such as 37.5/100 come from public launch/reporting—use the official Leaderboard as source of truth;
• Difficulty and verification depend heavily on domain-expert task design.

Who it's for

• Labs and startups building scientific agents / research automation;
• Evaluators who need end-to-end research benchmarks, not pure QA boards;
• Domain researchers willing to contribute real research tasks.

Getting started

1. Visit https://sciences-last-exam.org/ — Blog, Docs, Tasks;
2. Run the public task packages in agent harnesses such as Claude Code / Codex;
3. Contribute hard problems via Contribute and join the co-author list;
4. Cite: SLE Team, Science's Last Exam, 2026, https://sciences-last-exam.org/ .