BALAR: Bayesian Agentic Loop Enhances LLM Active Reasoning

A new algorithm called BALAR (Bayesian Agentic Loop for Active Reasoning) enables large language models to conduct structured multi-turn interactions by maintaining beliefs over latent states and selecting clarifying questions via expected mutual information. The task-agnostic, fine-tuning-free outer-loop method dynamically expands state representations when needed. Evaluated on AR-Bench-DC, AR-Bench-SP, and iCraft-MD benchmarks, BALAR outperformed all baselines across tasks including detective cases, thinking puzzles, and clinical diagnosis.