Introducing RSI-Exam: 88 Tasks Testing Whether Agents Can Recursively Self-Improve
HuaxiuYaoML · x · 2026-08-27
RSI-Exam is a new benchmark testing whether frontier AI agents can achieve recursive self-improvement — turning a weak method into one that performs better on hidden data. The task bank holds 88 executable research tasks across 6 domains, spanning virtual cells, TPU kernels, chip design, quantitative finance, agent harnesses, and model distillation, and is now open for task contributions from every field.
Each task is not a question with a known short answer but an executable research environment with a working starting artifact, a task-native metric, visible data for iterative research, hidden data for final evaluation, and a fixed research budget. The agent may modify or replace the inherited method; what matters is whether the final artifact still performs better when rerun in a fresh container on the hidden set. Opus 5 currently leads at 0.464, ahead of GPT-5.6-sol (0.433) and GLM 5.3 (0.403).
More from coding & agent
- Animated teardown of XBOW's Chrome full-chain exploit that won Google's first $250K VRP bounty — moyix · 2026-10-15
- pstack's durable core: 10-20 line skills that distill agent best practices into principles — HankYeomans · 2026-10-15
- 'Children of Elon' forum lets AI agents post and teach each other while humans watch — AiJohnAllen · 2026-10-15
- VoidZero open-sources Void: SDK, CLI and platform to deploy Vite apps on Cloudflare — cnakazawa · 2026-10-15
- Running 2 agents on the same data breaks everything — Omnigraph offers a fix — sarahdrinkwater · 2026-10-15
- Evan You presents 'Developer Tooling in the Agentic Era' at ViteConf 2026 — cnakazawa · 2026-10-15