← All benchmarks

RSI-Exam 0.1 (Full 88)

Agents · 2026-08-26

AIMING Lab (UNC-Chapel Hill)'s RSI-Exam 0.1 asks whether an agent can improve an executable research artifact over a long horizon. Each of 88 tasks in six domains (AI models and agents, physical sciences and engineering, optimization and control, systems and hardware, life sciences and medicine, and finance, law and business) hands the agent a weak but working method, or a harness that drives a frozen model, to improve for up to 12 hours against a visible split with network access limited to its model provider; the verifier then reruns the submitted artifact once, offline, on a sealed hidden split. Each task's native metric is mapped onto frozen anchors: the inherited starter is 0, a frontier-calibrated reference is 0.6 where the task defines one, and the upper bound is 1, with an exponential tail that approaches but never reaches 1 when no finite bound exists. The score is the equal-weight mean of the hidden-set scores over all 88 active tasks (35 public, 53 private), from one rollout per model and task; a final method no better than the starter scores 0. Each result belongs to a model running in a particular agent harness.

Benchmark source: Rsi Exam

Top models (higher is better)