BALAR : A Bayesian Agentic Loop for Active Reasoning

TL;DR

BALAR introduces a Bayesian agentic loop enabling large language models to actively reason and interact over multiple turns by selecting questions that maximize information gain, improving performance across diverse benchmarks.

Contribution

The paper presents BALAR, a task-agnostic, no-fine-tuning algorithm that structures multi-turn interactions for LLMs using Bayesian reasoning and mutual information maximization.

Findings

BALAR outperforms baselines with 14.6% higher accuracy on AR-Bench-DC.

BALAR achieves 38.5% higher accuracy on AR-Bench-SP.

BALAR improves accuracy by 30.5% on iCraft-MD.

Abstract

Large language models increasingly operate in interactive settings where solving a task requires multiple rounds of information exchange with a user. However, most current systems treat dialogue reactively and lack a principled mechanism to reason about what information is missing and which question should be asked next. We propose BALAR (Bayesian Agentic Loop for Active Reasoning), a task-agnostic outer-loop algorithm that requires no fine-tuning and enables structured multi-turn interaction between an LLM agent and a user. BALAR maintains a structured belief over latent states, selects clarifying questions by maximizing expected mutual information, and dynamically expands its state representation when the current one proves insufficient. We evaluate BALAR on three diverse benchmarks: AR-Bench-DC (detective cases), AR-Bench-SP (thinking puzzles), and iCraft-MD (clinical diagnosis). BALAR significantly outperforms all baselines across all three benchmarks, with $14.6\%$ higher accuracy on AR-Bench-DC, $38.5\%$ on AR-Bench-SP, and $30.5\%$ on iCraft-MD.