MetaRSI Report Makes Self-Improvement Recursive on Itself, Boosting Flagship Models 7.3 Points

青稞AI · wechat · 2026-09-13

CosmosMind, with researchers from Stanford, Berkeley, MIT, Tsinghua and other institutions, released the MetaRSI-v1 technical report: the first unified meta-recursive self-improving architecture covering Model-RSI, Data-RSI and Harness-RSI.

Key ideas

LoopKernel: a unified closed loop where learning signals propose changes on three writable surfaces (Data, Harness, Model), a verifier adjudicates, and results feed back as next-round signals.
Three operators: Data-RSI extracts learning signals from trajectories and maps capability boundaries; Harness-RSI adds/removes across five pluggable slots (SystemPrompt, Skill, MCP, Tools, Memory); Model-RSI internalizes verified behaviors into parameters.
Dual-axis orchestration: RSI²Agent picks operators horizontally, Sub-Agents rewrite operator policies vertically, and a MetaRSI²Agent optimizes the orchestration itself — three nested levels.

Results

Small-model track: Qwen3.5-35B-A3B (3B active) gained 10.9 points average across Terminal-Bench 2.1, SWE-bench Pro, GPQA-Diamond and AIME; SWE-bench Pro solve rate nearly doubled.
Frontier track: six flagship models (GPT-5.6, Claude Opus 5, etc., Data/Harness only, no external teacher) averaged +7.3 points.

Five laws include: verifiers set the frontier of self-improvement; self-knowledge goes stale so boundary re-discovery is the rate bottleneck; capability is carrier-independent but cost isn't (mature loops migrate capability to cheaper carriers); trustworthiness must be measured by unwritable surfaces; loops never create capability from nothing.

The team open-sourced RSI-Harness and a Genome community, arguing any scientific field needs only a task family, a verifier, and a rough initial scaffold to plug in.

Original post →

MetaRSI Report Makes Self-Improvement Recursive on Itself, Boosting Flagship Models 7.3 Points

More from coding & agent