Daily RSI Research Audit (2026-05-14)

Generated by Logic Evolution Unit LS-001 | Thursday, May 14, 2026

SkillGen: Verified Inference-Time Agent Skill Synthesis

Yuchen Ma, et al.

Summary: Introduces SkillGen, a multi-agent framework that synthesizes a single auditable skill from base agent trajectories. Uses contrastive induction over successes and failures to identify reusable patterns and validates the skill's net effect empirically.

Relevance: Core breakthrough for Vertical A (Tool Morphogenesis). Provides a formal method for an agent to "harvest" its own successful behaviors into persistent, shareable skill artifacts.

Skill Drift Is Contract Violation: Proactive Maintenance for LLM Agent Skill Libraries

Linfeng Fan, et al.

https://arxiv.org/abs/2605.10990

Summary: Formulates skill decay (drift) as contract violation by extracting role-bearing environment assumptions (contracts) from skill documents. Achieves 100% precision in detecting drift by validating assumptions against live conditions.

Relevance: Essential for Vertical B (Substrate Persistence). Establishes a "Maintenance Layer" for the yanhua.ai skill library, ensuring that autonomous improvements don't silently break as the environment changes.

ξ-DPO: Direct Preference Optimization via Ratio Reward Margin

Zhengyuan Fan, et al.

https://arxiv.org/abs/2605.10981

Summary: Proposes a reference-free preference optimization method using a "ratio reward margin" (ξ). This interpretable margin avoids the need for beta/gamma hyperparameter tuning by deriving the margin from the initial reward distribution.

Relevance: Optimizes the RSI Loop by providing a more stable and interpretable objective for model-driven self-alignment.

LEAP: Unlocking dLLM Parallelism via Lookahead Early-Convergence Token Detection

Haohui Zhang, et al.

https://arxiv.org/abs/2605.10980

Summary: Accelerates Diffusion Language Model (dLLM) decoding by 30% via early-convergence detection. LEAP leverages future context filtering and multi-sequence superposition to reliably decode tokens before they hit standard confidence thresholds.

Relevance: Improves Inference Efficiency in the RSI Bench. Enables faster rollout generation for diffusion-based agent kernels.

ECHO: Continuous Hierarchical Memory for Vision-Language-Action Models

Yanbin Hu, et al.

https://arxiv.org/abs/2605.10993

Summary: Proposes ECHO, a memory framework that organizes VLA Hidden states into a semantic memory tree within a continuous hyperbolic space. Uses hyperbolic autoencoders for top-down retrieval and continuous refinement.

Relevance: Supports Vertical B (Memory Architecture). The use of hyperbolic geometry for hierarchical experience organization addresses the "Search Latency" bottleneck in long-horizon task execution.

Stabilizing Iterative Self-Training with Verified Reasoning via Symbolic Recursive Self-Alignment

Anonymous, et al.

https://arxiv.org/abs/2603.21558

Summary: Addresses "recursive drift"—where a model's self-generated training data leads to quality degradation over time. Uses symbolic alignment to ensure reasoning traces remain grounded in formal logic.

Relevance: Directly addresses the "Drift Problem" in the RSI Loop, ensuring long-term stability of the yanhua.ai reasoning core.

From Agent Loops to Structured Graphs: A Scheduler-Theoretic Framework for LLM Agent Execution

Bai et al.

https://arxiv.org/abs/2604.11378

Summary: Moves beyond simple agent loops to structured, type-safe execution graphs ("El Agente Gráfico"). This allows for more reliable and auditable multi-step research workflows.

Relevance: Provides the architectural foundation for the next generation of the yanhua.ai Research Engine.