ANCHOR: Abductive Network Construction with Hierarchical Orchestration for Reliable Probability Inference in Large Language Models

TL;DR

To address the 'unknown' prediction problem in sparse factor spaces, we propose ANCHOR, a framework that hierarchically clusters factors generated by LLMs and models dependencies with a causal Bayesian network to estimate reliable probabilities.

Problem

Estimating reliable probabilities under incomplete information is a key challenge in large-scale decision-making. Existing LLM-based approaches generate explanatory factors and refine probabilities with a Naive Bayes model, but sparse factor spaces lead to frequent 'unknown' predictions, and expanding factors increases noise and spurious correlations, weakening the conditional independence assumption and degrading reliability.

Approach

ANCHOR is an aggregated Bayesian inference framework over a hierarchical factor space. It first constructs dense factor hierarchies through iterative generation and clustering of factors by LLMs. Then, it maps contexts via hierarchical retrieval and refinement, and augments Naive Bayes with a Causal Bayesian Network to model latent factor dependencies.

Results & Contribution

Experiments show that ANCHOR markedly reduces 'unknown' predictions and produces more reliable probability estimates than direct LLM baselines, achieving state-of-the-art performance while significantly reducing time and token overhead.