RSI-top framework unblocks stalled research agents, hits 77.3% medal rate on MLE-bench Lite
dbreunig · x · 2026-10-17
dbreunig shares RSI-top, a framework that helps autonomous research agents explore new directions when progress stalls.
- Results: Up to 15% better than OpenEvolve and ShinkaEvolve on Anthropic's kernel optimization task; 77.3% medal rate on MLE-bench Lite, beating the strongest baseline by 11.4 percentage points.
- Approach: Combines context management (a compact metric dashboard for selectively retrieving useful past trajectories) with three diversity-control mechanisms that push agents toward fresh perspectives and new starting points when stuck.
- New failure mode: "Context Inertia" — persistent models hyper-focus during long-horizon tasks; the fix is deliberately throwing out context.
More from coding & agent
- Building a notes chatbot: why semantic search fails on dates and membership queries — bzarnal · 2026-10-17
- 14B judge no better than 8B: one prompt tweak lifted grading from 3/14 to 10/14 — Lost-Masterpiece2400 · 2026-10-17
- 10 open-source repos for training small models to do your agent's job — victor_explore · 2026-10-17
- Robotics engineer challenges coding agent hype: 'They can't do visuomotor control' — AlexTensor · 2026-10-17
- Notion AI lead: much of what runs on GPUs should run as CPU code — threepointone · 2026-10-17
- Dev uses Google Argon procedural generation scripts to build real 3D game assets — Promptmethus · 2026-10-17