SlimeTree-RLM: Failure-Aware Routing and Controlled Recursive Inference

This paper presents a failure-aware routing and control framework integrating Recursive Language Models (RLMs) into the SlimeTree memory architecture. We introduce Three-Mode Inference (Delta/Mu/RLM), Failure-Aware Routing via continuous Unresolved Pressure scoring, and Automatic Slot Granularity Adaptation through Split/Merge/Freeze pressure dynamics. Key contributions: If-free routing using argmax over continuous mode scores Regret-based inverse learning signals (no model weight updates) Automatic semantic unit convergence via pressure dynamics Self-extinguishing recursive exploration All learning occurs via memory state transitions; no model weights are trained or modified. This positions SlimeTree as the structural substrate required to make recursive inference reliable. PoC Code: https://www.slimetree.ai/pre-print/zenodo-18238339-slimetree-rlm/code/ Patent Reference: JP 2025-183827