MetaRSI-v1: Meta-Recursive Self-Improvement Unifies Model, Data, Harness RSI

MetaRSI-v1, developed by CosmosMind with researchers from Stanford, Berkeley, MIT, Tsinghua, Peking University, and others, presents the first unified meta-recursive self-improving architecture (RSI²) that applies recursive self-improvement to the RSI system itself — moving self-improvement into a "square era." The architecture unifies three RSI dimensions — Model-RSI, Data-RSI, and Harness-RSI — under a single Loop Kernel paradigm, enabling composable, schedulable operators and a meta-level optimization loop.

1. Unified Loop Kernel Paradigm

The Loop Kernel abstracts "how progress happens" into an object-agnostic closed loop: consume feedback to obtain learning signals, propose changes on Data, Harness, or Model, submit to a verifier for adjudication, and feed results back as the next round's signal. This gives self-improvement a unified, reusable skeleton for the first time. Data, Harness, and Model become different instantiations of the same kernel, making them composable and uniformly schedulable.

2. Three Operators: Data-RSI, Harness-RSI, Model-RSI

All three operators share the same Loop Kernel and operate on three writable surfaces:

Data-RSI : Extracts learning signals from the system's own execution trajectories, verifies them, and synthesizes data for downstream consumption, calibrating and amplifying the model's positive and negative capability boundaries.

Harness-RSI : Uses learning feedback signals to generate improvements on five pluggable slots in the Harness — System Prompt, Skill, MCP, Tools, Memory — performing both addition and subtraction.

Model-RSI : Updates the model's own parameters and structure, internalizing repeatedly verified behaviors into the model weights.

3. Horizontal Orchestration and Vertical Optimization

Two orthogonal axes coordinate the operators:

Horizontal orchestration : The RSI² Agent selects the next operator at each decision point based on signal feedback, organically chaining and scheduling it into the RSI Loop Kernel for execution.

Vertical optimization : The RSI² Agent dispatches instructions to a target operator's dedicated RSI² Sub-Agent, which rewrites the operator's internal RSI rules based on learning signals and environmental feedback, thereby optimizing the RSI system itself.

4. Four Agents and Three-Level RSI Nesting

The architecture is coordinated by four agents, each replaceable by human experts for human-in-the-loop operation:

Transition Agent : Bridges the output of an upstream operator in horizontal orchestration to the consumption requirements of the next operator.

RSI² Agent : At each decision point, chooses between horizontal orchestration (specifying the next operator) and vertical optimization (sending strategy-rewrite instructions to a Sub-Agent).

RSI² Sub-Agent : Executes vertical optimization by adjusting the internal RSI strategy of a specific operator.

MetaRSI² Agent : Learns how to perform appropriate horizontal-vertical optimization, optimizing the RSI² Agent's own policy learning.

These four agents form three nested RSI levels: operators improve the target system; dual-axis orchestration improves operator usage; the meta-layer improves the dual-axis orchestration strategy itself.

5. Experiments: Self-Evolution on Small and Frontier Models

Experiments follow two tracks:

Small-model track : Uses the trainable open-source Qwen3.5-35B-A3B (35B total, 3B active parameters). All three operators (Data, Harness, Model) are enabled. Evaluated on Terminal-Bench 2.1, SWE-bench Pro, GPQA-Diamond hard subset, and AIME. MetaRSI-v1 achieves an average improvement of 10.9 points, with SWE-bench Pro solve rate nearly doubling. The meta-layer raises the ceiling for RSI, and cumulative gains from continuous self-evolution are significant.

Frontier-model track : Targets closed-source or massive models — Claude Opus 5, GPT-5.6 Sol, Kimi K3, and three others. Only Data and Harness operators are enabled (Model-RSI requires weight updates). On Terminal-Bench 2.1, the six frontier models improve by an average of 7.3 points, showing that flagship models still have substantial self-improvement headroom.

6. Five Laws of Recursive Self-Improvement

MetaRSI distills two complementary improvement paths: Harness-RSI keeps weights frozen, enabling self-improvement for any model accessible via API; Model-RSI internalizes capabilities through training. Data-RSI is redefined as a capability-boundary probe and amplifier: joint analysis of execution traces and external feedback learning-signals yields verified experience scaled into reusable data, delineating where those experiences hold — defining positive and negative capability boundaries. Data-RSI products feed both Harness-RSI and Model-RSI; their results flow back to Data-RSI to re-calibrate boundaries and drive the next round, with capabilities progressively revealed and accumulated. Harness can subtract: after Model-RSI internalizes a capability, the corresponding Harness knowledge becomes redundant and is deleted under replay verification — capability stays, cost exits.

From this, five laws emerge:

Law 1: Verification determines the self-improvement frontier. Whether a domain can form a self-improvement loop depends on whether its tasks are verifiable and have a suitable verifier.

Law 2: Self-knowledge expires; rediscovering capability boundaries is the rate bottleneck. RSI changes agent capabilities, immediately invalidating its previous system description.

Law 3: Capability is carrier-agnostic, but cost is carrier-dependent. A capability in Harness incurs recurring inference cost; internalized into Model it costs once. Mature loops continuously migrate capabilities to cheaper carriers.

Law 4: Trustworthiness is measured by the unwritable surface. Inside the loop, genuine capability improvement and relaxed success criteria yield identical scores; this bias is undetectable internally and cannot be corrected statistically.

Law 5: Loops do not create capability ex nihilo; all gains are fundamentally external information-entropy input or capability elicitation. Self-improvement only reorganizes, amplifies, and solidifies latent potential. Each improvement must separate what is amplified from what is newly introduced from outside.

7. From Architecture to Ecosystem: A Loop for Every Scientific Domain

MetaRSI-v1 does not require a domain to execute the entire loop at once. Scientific research and engineering are pipelines — hypothesize, design experiment, execute, interpret — each with clear inputs and outputs. Instead of a big bang, start by closing a small loop at any single stage; its output naturally becomes the next stage's input. When enough small loops spin, they connect into a domain-wide research loop. Onboarding a new domain requires only three things: a task family that produces results, a verifier to judge correctness, and an initial scaffold (however rough); the rest is handled by generic mechanisms.

To make this paradigm flow, the team open-sourced RSI-Harness — a self-learning, self-iterating harness that can distill portable Harness Genomes for different scientific and engineering domains. They also launched a Genome open community , aiming for every domain to deposit its own reliable scaffold so new loops can start on the shoulders of predecessors.

Resources: Project homepage: https://github.com/CosmosMind-ai/RSI-Harness; Paper: https://www.cosmosmind.ai/research/metarsi-v1.pdf; HuggingFace: https://huggingface.co/CosmosMind/RSI-Harness