An Information-Theoretic Model of Abduction for Detecting Hallucinations in Explanations

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Information Theory

Background:

• Generative models, like large language models (LLMs), can produce unsupported or contradictory content, known as hallucinations.
• Existing hallucination detection methods often lack interpretability or struggle with complex reasoning.

Purpose of the Study:

• To introduce a novel neuro-symbolic framework for detecting hallucinations in generative models.
• To enhance the accuracy and interpretability of hallucination detection.

Main Methods:

• Developed an Information-Theoretic Model of Abduction (ITMA).
• ITMA combines entropy-based inference with abductive reasoning.
• Incorporates discourse structure using Rhetorical Structure Theory (RST)-derived EDU weighting.

Main Results:

• The proposed method outperforms state-of-the-art neural and symbolic baselines on medical, factual QA, and multi-hop reasoning datasets.
• Demonstrated superior accuracy and interpretability in hallucination detection.
• Successfully identified plausible-sounding but abductively unsupported model errors.

Conclusions:

• Integrating Information-Theoretic divergence and abductive explanation offers a principled foundation for robust hallucination detection.
• The framework effectively distinguishes legitimate elaborations from unjustified claims.
• Provides a significant advancement for reliable generative AI systems.