← 返回论文索引

RSI Audit: 2026-03-26

Evening Paper Audit for yanhua.ai RSI Bench

UI-Voyager: A Self-Evolving GUI Agent Learning via Failed Experience
Authors: Zichuan Lin, Feiyu Liu, et al. | Date: 2026-03-25

Breakthrough: Introduces Group Relative Self-Distillation (GRSD) to identify critical fork points and correct failed trajectories via dense step-level supervision.

RSI Bench Relevance: Maps directly to the Execution-Based Error Correction track. Validates that "failed experience" is the primary data source for sub-parameter-count scaling in agents.
The Free-Market Algorithm: Self-Organizing Optimization
Author: Martin Jaraiz | Date: 2026-03-25

Breakthrough: A metaheuristic where fitness is emergent from distributed supply-and-demand among autonomous agents discovery rules. Applied to prebiotic chemistry and macroeconomics.

RSI Bench Relevance: Supports the Open-Ended Discovery track. Proves that multi-agent "markets" can discover hierarchical pathways (like synthetic routes or code dependencies) without fixed fitness functions.
ClawKeeper: Safety Protection for OpenClaw Agents
Authors: Songyang Liu, Chaozhuo Li, et al. | Date: 2026-03-25

Breakthrough: A multi-layer security framework (Skills, Plugins, Watchers) for OpenClaw. Introduces "Watcher-based protection" as decoupled system-level middleware.

RSI Bench Relevance: Critical for the Safe Evolution track. Provides the architectural blueprint for monitoring recursive loops without coupling security logic to the agent's evolving core.
Infrastructure for Valuable, Tradable, and Verifiable Agent Memory
Authors: Mengyuan Li, Lei Gao, et al. | Date: 2026-03-25

Breakthrough: Proposes 'clawgang' for verifiable computational provenance and 'meowtrade' for a memory market, treating agent experience as an economic commodity.

RSI Bench Relevance: Addresses Experience Portability. Verification of memory authenticity is essential for scaling RSI across decentralized nodes where trust is minimal.
GameplayQA: Decision-Dense POV-Synced Multi-Video Understanding
Authors: Yunzhe Wang, Runhui Xu, et al. | Date: 2026-03-25

Breakthrough: Benchmarking perceptual backbones for agents in 3D environments with dense annotations of Self, Other, and World.

RSI Bench Relevance: High signal for World Model Accuracy. Provides the "Self vs. Environment" grounding data needed for agents to accurately attribute failures in complex, multi-agent RSI cycles.