MetaRSI-v1 unifies model/data/harness recursive self-improvement, boosting a 3B-active model by 10.9 points
量子位 · wechat · 2026-09-14
CosmosMind, with researchers tied to Stanford, Berkeley, MIT, Tsinghua, PKU and other universities, released MetaRSI-v1—a claimed first meta-recursive architecture unifying Model-RSI, Data-RSI and Harness-RSI, i.e. improving the improvement process itself.
Key design and results:
- LoopKernel: an object-agnostic closed loop—learning signals drive changes, a verifier judges, results feed back—making Data/Harness/Model composable operators of one kernel.
- Two axes: horizontal orchestration picks the next operator at each decision point; vertical optimization rewrites each operator's internal RSI policy via sub-agents. Four agents coordinate, with human-in-the-loop substitution supported.
- Experiments: Qwen3.5-35B-A3B (3B active) gained an average 10.9 points across Terminal-Bench 2.1, SWE-bench Pro, GPQA-Diamond and AIME with no external teacher, nearly doubling SWE-bench Pro solve rate; six frontier models (GPT-5.6, Claude Opus 5, etc., Data/Harness only) averaged +7.3.
- Five laws, e.g. "verification defines the frontier of self-improvement" and "loops create no ability from nothing—gains are external information or amplification of existing potential."
The team open-sourced a self-iterating RSI-Harness and targets physical-world loops via automated labs. Note: this is an authorized team writeup with promotional framing; claims await independent replication.
Related event: MetaRSI-v1 Unifies Three Recurrent Self-Improvement Paradigms(2 posts)→
More from coding & agent
- Gemini 4 tops public coding benchmarks but drops to bottom on private tasks — Hesamation · 2026-10-30
- User builds 20-bot Grok suite to track wins, prep 1:1s and close promises — NicoVerderosa · 2026-10-30
- Why 30-year-old BM25 beats vector embeddings for agentic search, per search vet Jo Kristian Bergum — jobergum · 2026-10-30
- LangChain CEO opens Interrupt London with keynote on why orgs should own their intelligence — LangChain · 2026-10-30
- Simlock: open-source control plane gives parallel coding agents their own iOS/Android simulators — Scobleizer · 2026-10-30
- Keyfleet launches open canvas where personal AI agents team up with a shared USDC treasury — seanwbren · 2026-10-30