Skip to content

Scientific and ML agents

ScienceAgentBench Verified

Data-driven scientific coding tasks; original suite has 102 instances.

About ScienceAgentBench Verified

What it measures

Executable task/output verification.

Scope & limitations

Use the verified update released in April 2026. Search-only products need code execution before they fit the full protocol.

Evaluation environment

Scientific code agent with datasets and declared GPU requirements.

Source

Original benchmark

Dataset & task catalog ↗