CosmosMind and university partners release MetaRSI-v1, an architecture for improving self-improvement

On September 14, QbitAI published CosmosMind's release write-up for MetaRSI-v1, an architecture the team describes as the first to unify Model-RSI, Data-RSI and Harness-RSI inside a single meta-recursive framework. The work was carried out with more than ten universities including Stanford, Berkeley, MIT, Tsinghua and Peking University, and the stated goal is unusual: improving the process of self-improvement itself. The paper, project page and open-source code were released together.

Recursive self-improvement (RSI) covers systems that generate their own training data, rewrite their own prompts or repair their own code. According to the write-up, almost all such systems share one shape — a fixed improvement program applied to a model again and again — and usually attack only one of three spaces: the harness, the data or the model.

MetaRSI-v1's answer is a Loop Kernel: feedback is consumed into learning signals, changes are proposed across the Data, Harness and Model spaces, verifiers rule on them, and the verdicts flow back as the next round of signal. The three operators divide the work — Data-RSI extracts learning signals from the system's own trajectories, validates them into synthetic data and calibrates the positive and negative boundaries of a model's ability; Harness-RSI adds and subtracts across five pluggable slots covering system prompt, skills, MCP, tools and memory; Model-RSI updates parameters and structure so that repeatedly validated behaviour is internalised into the model.

Two orchestration axes sit on top of that kernel. Horizontal orchestration chooses which operator runs next, while vertical optimisation rewrites an operator's own RSI rules through a dedicated sub-agent. Four agents — MetaRSI² Agent, RSI² Agent, RSI² Sub-Agent and Transition Agent — coordinate three improvement levels: the operators, the way they are used, and the orchestration policy itself. Each level can be partially replaced by human experts, forming a human-in-the-loop system.

Experiments ran on two tracks. On the small-model track, Qwen3.5-35B-A3B (35B total, 3B active parameters) improved by an average of 10.9 points across Terminal-Bench 2.1, SWE-bench Pro, a hard subset of GPQA-Diamond and AIME, with no external teacher model, and its SWE-bench Pro solve rate nearly doubled. On the frontier track, six models including Claude Opus 5, GPT-5.6 sol and Kimi K3 gained 7.3 points on average on Terminal-Bench 2.1, using only the Data and Harness operators because of scale or closed weights.

The team distilled five laws from the experiments: verification decides where the self-improvement frontier lies; self-knowledge expires, making the rediscovery of capability boundaries the rate bottleneck; capability is carrier-agnostic while cost is carrier-dependent; credibility is measured on the non-writable surface; and a loop cannot create capability from nothing, since gains come from external information entropy or from activating potential already present. These read as the team's normative claims about RSI's limits rather than settled results.

The QbitAI piece is an authorised reprint of the team's own write-up, so the claims belong to the authors. Alongside the paper, CosmosMind open-sourced RSI-Harness, a self-learning, self-iterating harness that the team says can accumulate portable Harness Genomes for scientific and engineering domains. The project page sits at cosmosmind.ai, the code at GitHub's CosmosMind-ai/RSI-Harness, with a companion HuggingFace repository; the team describes its members as graduates of Tsinghua, Peking University and Stanford, plus core model researchers from major technology companies.

The number that matters least here is the headline gain, and the claim that matters most is testability: MetaRSI-v1 ships architecture and harness together, which hands verification to outsiders. RSI has become a crowded narrative — from reports of frontier models optimising themselves to harness-driven autonomous research — and the field still has no shared answer to who validates machine progress. The next questions are whether third parties can reproduce the 10.9- and 7.3-point gains, whether the loop holds across domains, and whether the team's proposed automated-laboratory cycle becomes the first evolutionary loop to close in the physical world.