RSI-Anything Open-Sourced: Unlocking Domain-General Recursive Self-Improvement for AI Agents with World Models

Background and the Problem

Recursive Self-Improvement (RSI) offers the theoretical promise of AI systems autonomously scaling toward superintelligence through continuous interaction. However, practical RSI algorithms fail due to two bottlenecks:

1. Scarce and Costly Ground-Truth Feedback: Interacting with real execution environments (running test suites, executing wet-lab experiments, or performing GUI operations) is slow and financially prohibitive.
2. Co-drift Collapse of World Models: Simulating environments using predictive world models frequently breaks down as evolving policies wander out-of-distribution, producing hallucinations and leading to catastrophic reward hacking.
3. Domain-Specific Brittleness: Prior self-improvement techniques relied on rigid game rules (e.g., AlphaZero) and failed to generalize to open domains like code repair and physical sciences.

Architecture and How It Works

Researchers from HKUST, Shanghai AI Lab, Peking University, and CUHK introduce RSI-Anything (arXiv:2610.13912, code: github.com/Gen-Verse/RSI-Anything):

1. Six-Stage P-P-S-E-L-R Paradigm: Implements a disciplined lifecycle: Propose (policy generates diverse actions) → Predict (JEPA world model anticipates outcomes in latent space) → Select (filters top candidates via predictive confidence) → Evaluate (anchors against scarce ground-truth execution) → Learn (co-adapts policy and world model) → Retain (enforces compatibility gates).
2. Conditional World-Model Adaptation: Dynamically updates the predictive world model using freshly acquired interaction traces, preventing representation drift across policy iterations.
3. Unified Latent Architecture: Built on JEPA-Anything, predicting high-level representations rather than low-level pixel/token tokens to minimize compute overhead.

Benchmarks and Measured Results

Evaluated across code repair, GUI navigation, laboratory planning, and reaction optimization:

1. Code Repair SOTA (31.2%): Raises automated code repair accuracy from a 26.3% static verifier baseline to 29.2% via predictive selection, reaching 31.2% after world-model adaptation (+4.9 percentage points).
2. 39.4% Reduction in GUI Action Footprint: On OSWorld and GUI benchmarks, cuts average interaction steps by 39.4% while maintaining superior task completion rates.
3. Laboratory Planning Doubled (21.80% to 46.01%): Over five recursive policy iterations, planning task performance surges from 21.80% to 46.01% (a >111% relative increase).
4. Chemical Reaction Yield (96.15% vs 82.16%): Surpasses Bayesian optimization under equivalent exploration budgets, attaining a 96.15% mean best yield.

Getting Started for Developers

RSI-Anything is fully open-sourced on GitHub (Gen-Verse/RSI-Anything). AI teams building autonomous agents can deploy JEPA-based world models as lookahead verifiers before invoking expensive real-world executions, integrate strict Retain gating to avoid policy regression, and harness domain-general self-evolution workflows across coding, UI automation, and scientific discovery.