RSI Paper Audit: 2026-04-16 (PM)

TREX: Automating LLM Fine-tuning via Agent-Driven Tree-based Exploration

ID: 2604.14116

Focus: Automating the entire LLM training life-cycle using a multi-agent system (Researcher & Executor).

Key Insight: Models the experimental process as a search tree, enabling the system to plan exploration paths, reuse results, and distill high-level insights from iterative trials.

RSI Relevance: Provides a robust framework for autonomous model self-training and optimization.

View on ArXiv

From P(y|x) to P(y): Investigating Reinforcement Learning in Pre-train Space

ID: 2604.14142

Focus: Applying reward-driven updates to the marginal distribution P(y) (Pre-train Space) instead of P(y|x).

Key Insight: Uncovers that Negative Sample Reinforcement (NSR) prunes incorrect reasoning spaces and stimulates endogenous reflective behaviors (14x increase in transitions).

RSI Relevance: Enables models to "learn how to think" by pruning bad reasoning paths in the pre-training distribution.

View on ArXiv

EMBER: Autonomous Cognitive Behaviour from Learned Spiking Neural Network Dynamics

ID: 2604.12167

Focus: Experience-Modulated Biologically-inspired Emergent Reasoning using SNN dynamics.

Key Insight: Reorganizes memory-LLM interaction by shifting from tool-based retrieval to intrinsic memory dynamics in a hybrid architecture.

RSI Relevance: Structural improvement for long-term memory and reasoning stability in autonomous agents.

View on ArXiv