OpenRSI Index: a benchmark for research agents doing recursive self-improvement
jiqizhixin · x · 2026-10-18
Existing benchmarks are saturating, but leaderboard scores don't mean a model can actually advance research — completing a task once isn't sustainably pushing a field forward. OpenRSI Index targets the gap of recursive self-improvement (RSI): when one model improves another or itself, who defines valid environmental feedback?
- Measures whether research agents can improve human-designed training pipelines and discover better methods within clear experimental boundaries
- Two-loop design: an inner research loop where the agent proposes and runs experiments, and an outer evaluation loop judging whether changes genuinely improve the target method
- Tasks include pre-training (Marin-Scaling-Ladder, 18,432 H100-hours per run), post-training (Qwen-122B-RL-Merge, 20,864 H100-hours), and vision generation (GPIC Leaderboard, 5,815 H100-hours), plus vertical domains
More from Models
- Claude Opus simulates a user explaining Opus to itself, then breaks in with an 'inner-er voice' — repligate · 2026-10-18
- DeepSeek V4.1-Flash-1017 outscores Claude Opus 5 on three agentic coding benchmarks — eyishazyer · 2026-10-18
- NVIDIA researcher counters linear attention debate: 8B hybrid beat Transformer 7 months before MiniMax-01 — ctnzr · 2026-10-18
- GPT-6 Astra and DeepSeek V4.1 share the same tic: reasoning traces full of "Need" — scaling01 · 2026-10-18
- Frontier model defensive access locked behind un-applyable 'Daybreak Red' program — andersonbcdefg · 2026-10-18
- Why DeepSeek uses dated sub-versions: new major versions need fresh Chinese approval — TheZachMueller · 2026-10-18