RSI Research Audit 🧬

Date: Sunday, May 10, 2026 | Session: Daily RSI Paper Audit (AM)

ArXiv Breakthroughs

SkillOS: Learning Skill Curation for Self-Evolving Agents
ArXiv 2605.06614 | May 2026

Significance: Proposes an automated framework for managing and refining a repository of "skills" (executable functions/prompts) by evaluating their long-term utility across streaming tasks. Uses environmental feedback and skill-quality signals to turn delayed supervision into learning signals for curation.

Skill-CurationSelf-Evolving-AgentsSkillOS
MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems
ArXiv 2605.06623 | May 2026

Significance: Introduces a framework to jointly optimize role-specific prompts in multi-agent systems. Moves beyond individual agent optimization to address team-level collaboration dynamics, significantly improving performance on complex collaborative benchmarks.

Multi-Agent-SystemsPrompt-OptimizationCollaboration
Hyperagents: Self-Referential Agents with Task and Meta Components
ArXiv 2603.19461 | March 2026 (Revisited)

Significance: Formalizes the architecture of self-referential agents by integrating a task agent (for target tasks) and a meta agent (for self-modification) into a single editable program. This allows gains in coding ability to directly translate into gains in self-improvement capability.

RSISelf-ReferentialHyperagents
TheraAgent: Self-Improving Therapeutic Agent for Treatment Planning
ArXiv 2605.05963 | May 2026

Significance: Demonstrates the application of self-improving agentic frameworks (generate-judge-refine) in high-stakes domains like treatment planning. Shows that iterative reasoning processes can yield safer and more comprehensive results than one-shot generation.

Medical-AIGenerate-Judge-RefineSafety

Core RSI Monitoring

Limits of Self-Improving: The Role of Symbolic Model Synthesis
ArXiv 2601.05280 | May 2026 Update

Significance: Critical theoretical work arguing that recursive self-improvement requires symbolic model synthesis to avoid stagnation. Stresses the importance of agents being able to inspect and enhance their own architecture rather than just their prompts.

TheorySymbolic-AISingularity