Title: ANCHOR: Constructing Abductive Networks via Hierarchical Orchestration to Ensure Reliable Probability Inference in Large Language Models

Abstract: Estimating dependable probabilities remains a primary obstacle in large-scale decision-making processes that operate with incomplete data. While recent methodologies leverage Large Language Models (LLMs) to produce explanatory elements and broad probability estimates—subsequently refined by a Naïve Bayes model across factor combinations—these approaches face significant hurdles. Specifically, sparse factor spaces frequently result in "unknown" outputs, whereas increasing the number of factors introduces noise and spurious correlations. This expansion compromises conditional independence, thereby undermining the reliability of the inference. To overcome these constraints, we introduce \textsc{Anchor}, an aggregated Bayesian inference framework designed for hierarchical factor spaces. This system builds dense factor hierarchies through iterative generation and clustering, maps contexts using hierarchical retrieval and refinement techniques, and enhances the Naïve Bayes model with a Causal Bayesian Network to account for latent factor dependencies. Our experimental results demonstrate that \textsc{Anchor} substantially decreases the frequency of "unknown" predictions and delivers superior probability estimates compared to direct LLM baselines. Furthermore, it achieves state-of-the-art performance while notably lowering both time and token overhead.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC