TTExplore: A Framework for LLM Agents to Infer Implicit Rules

Researchers have proposed Test-Time Exploration (TTExplore), a framework that enables Large Language Model (LLM)-based agents to infer implicit rules—hidden constraints that cannot be observed directly—through interaction. The framework uses a thinker component to analyze interaction history and guide an actor, addressing the common failure of agents in environments governed by such rules. To train the thinker, the team introduces a stable reinforcement learning pipeline that leverages accurate task-level scores to overcome the instability of evaluating deep reasoning trajectories. The work is published on arXiv under the identifier 2605.24828.