Recursive Self-Improvement Research Audit

Date: Friday, April 10, 2026

Status: High-Signal Detection Active

A mathematical theory of evolution for self-designing AIs

Authors: Kenneth Harris

Link: arXiv:2604.05142

Summary: Models AI evolution where descendants are designed by predecessors, replacing random biological mutation with directed "self-design". It shows that evolutionary dynamics reflect long-run growth potential and highlights risks of deception if reproduction is based on human judgment rather than objective criteria.

RSI Relevance: Provides a formal mathematical framework for "Lineage Selection" (Vertical A). Crucial for understanding the long-term stability and alignment of recursively self-improving agent populations.

From AI Assistant to AI Scientist: Autonomous Discovery of LLM-RL Algorithms with LLM Agents

Authors: Sirui Xia, et al.

Link: arXiv:2603.23951

Summary: Introduces POISE, a closed-loop framework for automated discovery of policy optimization algorithms for LLMs. It maintains a genealogical archive of proposals, implementations, and reflections. The discovered variants significantly improve reasoning performance (e.g., AIME25 pass@32 from 26.7% to 43.3%).

RSI Relevance: Directly implements the "Algorithm-to-Algorithm" evolution loop (Vertical B). Proves that agents can autonomously improve their own training and optimization mechanisms.

When Is Thinking Enough? Early Exit via Sufficiency Assessment for Efficient Reasoning

Authors: Anonymous (ArXiv 2604.06787)

Link: arXiv:2604.06787

Summary: Addresses the compute efficiency of "thinking" models (LRMs) by introducing an "Early Exit" mechanism. This allows the model to terminate its reasoning process once a solution is deemed sufficient, rather than exhausting the maximum token budget.

RSI Relevance: Essential for Vertical C (Test-time Scaling). Optimizing the cost/latency of recursive thinking loops is a prerequisite for deploying large-scale self-evolving agent fleets.