Reproduce “A Probabilistic Framework for LLM Hallucination Detection via Belief Tree Propagation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Bairu Hou, Yang Zhang, Jacob Andreas, Shiyu Chang

Researchers at UC Santa Barbara and MIT-IBM Watson AI Lab introduced BTPROP, a probabilistic framework that improves large language model (LLM) hallucination detection by modeling logical relationships between claims as a Hidden Markov Tree. This framework explicitly accounts for the inherent noise and miscalibration in LLM self-confidence scores, leading to performance gains of 3% to 9% in AUROC and AUC-PR on scientific and general knowledge benchmarks over existing methods.

Key insights

  • Explicitly modeling the noise and miscalibration in an LLM's self-assigned confidence scores is crucial for accurate hallucination detection, rather than treating them as perfectly calibrated beliefs.
  • Organizing logically related statements into a structured belief tree enables global consistency checks and allows the framework to leverage inter-statement dependencies for more robust factuality judgments.
  • A Hidden Markov Tree model provides a principled probabilistic framework to integrate both LLM's observed (noisy) confidence and the defined logical relationships between statements to infer the true factual correctness.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.