BALAR: A Bayesian Agentic Loop for Active Reasoning
BALAR is a task-agnostic loop that enables LLMs to reason actively by maintaining a structured belief about user intent. It uses Bayesian updates and maximizes information gain through intelligent question selection to refine this belief dynamically. This method allows agents to handle complex, multi-turn interactions efficiently without needing specific training for the task at hand, leading to better performance across diverse reasoning benchmarks.