RSI-Exam Quantifies Recursive Self-Improvement Across 88 Tasks

RSI-Exam benchmarks bounded recursive self-improvement: agents improve an executable artifact or frozen-model harness over hours, validated by hidden-test one-shot reruns. It starts from weak runnable baselines; dual-container isolation prevents overfitting. Scores: 0 baseline, 0.6 frontier, 1.0 bound. 88 tasks span six domains. Top: Opus 5 Claude Code 0.464, GPT-5.6-sol Codex 0.433, GLM 5.3 Claude Code 0.403.