Science’s Last Exam tests agents on real research

// 3h agoBENCHMARK RESULT

Science’s Last Exam tests agents on real research

AItonomy Foundation’s benchmark evaluates agents on 113 research tasks across nine scientific disciplines, covering hypothesis formation, protocol planning, scientific computing, software operation, experiment observation, and end-to-end research. The initial leaderboard’s top configuration scored 37.5/100, while every system scored zero on the end-to-end track.

// ANALYSIS

This is a more meaningful test of scientific agency than question-answering benchmarks, but the zero end-to-end result is also a reminder that autonomous research remains far from solved.

–Tasks use real scientific software, research data, recordings, and withheld expert references
–The benchmark separates useful diagnostic skills instead of hiding weaknesses behind one aggregate score
–End-to-end research currently has only one task and is explicitly slated for expansion
–Wet-lab execution is still under construction and is not scored
–Developers should treat the 37.5 score as evidence of partial workflow competence, not independent scientific discovery

// TAGS

sciences-last-exambenchmarkevaluationagentresearchtool-use

DISCOVERED

3h ago

2026-11-05

PUBLISHED

5h ago

2026-11-05

RELEVANCE

9/ 10

AUTHOR

omarsar0

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

INFRA23m ago

Vercel for Platforms powers Cursor app deployment

Vercel for Platforms gives AI coding tools such as Cursor a programmatic path to create, deploy, and operate customer apps with isolated projects, custom domains, and managed infrastructure. Its claim-and-transfer flow can bundle third-party resources such as databases, reducing deployment friction for generated applications.

UPDATE39m ago

Token Pass Credits Auto-Apply Across AI Gateway

Vercel now automatically applies Token Pass credits to included-model requests through AI Gateway using standard API keys, OIDC tokens, or fx login. After those credits run out, standard-key and OIDC requests fall back to team credits, while Token Pass keys stop until renewal.

LAUNCH48m ago

Inworld Adds 24 Skills for Coding Agents

Inworld is introducing a 24-skill plugin that lets coding assistants explore, test, and integrate its voice, speech, realtime-agent, and inference APIs. The plugin pairs live documentation with executable MCP tools and CLI workflows for faster implementation.