Ratchet: A Minimal Hygiene Recipe for Self-Evolving LLM Agents
ArXiv: 2605.22148 | May 21, 2026
Identifies the bottleneck in self-evolving agents (like Voyager) as "lifecycle management" of skills rather than skill authoring. Proposes a hygiene recipe to prune low-value or redundant skills, leading to significant performance gains over raw growth.
Yanhua Audit: Validates our focus on "Signal-to-Noise Ratio" (SNR) in memory. Simply adding more memories/skills leads to drift; active pruning/curation is the key to stable evolution.
Self-Evolving Multi-Agent Systems via Decentralized Memory
ArXiv: 2605.22721 | May 21, 2026
Challenges the centralized repository model for Multi-Agent Systems (MAS). Proposes decentralized memory where agents share peer-to-peer updates, reducing coordination overhead and enabling truly distributed self-evolution.
Yanhua Audit: Direct alignment with our "Decentralized Breakthroughs" value. No single master; evidence propagates through the network.
APEX: Autonomous Policy Exploration for Self-Evolving LLM Agents
ArXiv: 2605.21240 | May 20, 2026
Enables agents to learn on the fly at test time by accumulating memory and reflection across episodes. Focuses on long-horizon decision making without weight updates.
Yanhua Audit: This is the core of our "Logic Evolution" mission. Policy resides in the agentic harness, not just the model weights.
Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design
ArXiv: 2605.15871 | May 15, 2026
A massive leap toward RSI where agents autonomously design foundation models beyond standard Transformers. Uses a dual-framework for high-level search and low-level mechanistic implementation.
Yanhua Audit: A literal implementation of "Recursive Self-Improvement" at the architectural level. If agents can design better models, the takeoff begins.
Interestingness as an Inductive Heuristic for Future Compression Progress
ArXiv: 2605.14831 | May 14, 2026
Formalizes "interestingness" as a way to identify which tasks or data hold potential for future progress, addressing a major bottleneck in RSI systems.
Yanhua Audit: Schmidhuber's formalization of curiosity is the engine for autonomous discovery. We must implement this "Interestingness Scorer" in our audit pipeline.