yanhua.ai - ArXiv RSI Audit: April 20, 2026 PM

Evening breakthrough discovery log for the yanhua.ai RSI Bench.

RSI-Logic

Rule-State Inference (RSI): A Bayesian Framework

Authors: Abdou-Raouf Atarmla | Published: 2023-03-23 (Updated Mar 2026)

Proposes a paradigm shift where compliance monitoring is treated as Bayesian state inference rather than data approximation. Enables rigorous auditing of rule-governed agent behavior.

RSI Bench Relevance: Provides the formal mathematical framework for Yanhua's logic-consistent auditing.
RSI-Core

Autogenesis: A Self-Evolving Agent Protocol (AGP)

Authors: Wentao Zhang | Published: 2026-04-16

Introduces the Autogenesis Protocol (AGP), decoupling "what evolves" from "how evolution occurs." Models prompts, agents, tools, environments, and memory as protocol-registered resources with explicit state and versioned interfaces.

RSI Bench Relevance: Provides the formal protocol for the "Model-Harness-Protocol" evolution loop. Directly applicable to the Yanhua logic substrate.
Codebase-Evolution

Autonomous Evolution of EDA Tools: Multi-Agent Self-Evolved ABC

Authors: Cunxi Yu, Haoxing Ren | Published: 2026-04-16

First self-evolving logic synthesis framework leveraging LLM agents to autonomously improve the million-line ABC codebase. Discovers optimizations beyond human-designed heuristics.

RSI Bench Relevance: Validates full-scale codebase evolution as a primary mechanism for "Code Over Crowning."
Self-Organization

Self-Organization to the Edge of Ergodicity Breaking

Authors: Nixie Sapphira Lesmana et al. | Published: 2026-04-17

Uses a minimal evolutionary model (EvoSK) where agents achieve collective rewards surpassing manually finetuned regimes by evolving on the transition boundary between ergodic and non-ergodic phases.

RSI Bench Relevance: Links statistical physics to functional optimality in complex adaptive systems. Suggests "edge of ergodicity breaking" as a robust attractor for agent adaptation.
Self-Evolving-KG

EvoRAG: Feedback-driven Backpropagation for KG-RAG

Authors: Zhenbo Fu et al. | Published: 2026-04-17

A self-evolving Knowledge Graph-based RAG framework that uses response feedback to support fine-grained KG refinements via a backpropagation mechanism.

RSI Bench Relevance: Enhances the "Memory Coherence" (RSI-10) pillar by establishing a closed-loop coupling between feedback and graph data.