RSI-Exam: New Benchmark Tests AI Agents' Recursive Self-Improvement Across 88 Tasks
yuyinzhou_cs · x · 2026-08-27
As frontier models improve monthly, Recursive Self-Improvement (RSI) is hot — but how to measure it? Researchers introduce RSI-Exam: 88 executable research tasks across 6 domains (AI agents, virtual cells, TPU kernels, chip design, quant finance, distillation, and more), testing whether agents can improve a working method via long-horizon autonomous experimentation and generalize to hidden data.
Protocol: an agent inherits a working method and iterates on visible data for hours; only the final artifact is evaluated once in a fresh container on a hidden set.
Findings: Claude and GPT form a clear first tier (Opus 5 leads the 88-task leaderboard by mean hidden-set score), with a large gap over the rest — yet huge headroom remains. Trajectories show both agents discovering fundamentally better methods and agents spending hours optimizing the wrong idea. Task contributions are open to all fields.
More from Models
- Image inputs now ~10,000 per dollar, but re-referencing an image doubles token cost — smtabatabaie · 2026-10-15
- "Argon doesn't trust me": researcher laments overzealous model refusal — thegautamkamath · 2026-10-15
- User finds Grok shockingly bad at finding X posts, prefers Opus with browser — JasonBotterill · 2026-10-15
- "Claude-Next" Reportedly Returns to Claude CLI and Web via Stealth Opus Routing — koltregaskes · 2026-10-15
- 'Thinking leakage': no-think RL gains in hybrid models just reuse think-mode skills — kaust-generative-ai · 2026-10-15
- Cline: Gemini 4 Argon saw 3x Claude's day-one adoption and 30% fewer tokens — JMateosGarcia · 2026-10-15