Reproduce “HalluGuard: Demystifying Data-Driven and Reasoning-Driven Hallucinations in LLMs” 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

Xinyue Zeng, Junhong Lin, Yujun Yan, Feng Guo, Liang Shi, Jun Wu, Dawei Zhou

Researchers from Virginia Tech, MIT, Dartmouth College, and Michigan State University present HALLUGUARD, a theoretical framework and detection method to jointly identify data-driven and reasoning-driven hallucinations in Large Language Models. HALLUGUARD leverages Neural Tangent Kernel properties to quantify hallucination risk, achieving state-of-the-art performance across 10 benchmarks and 9 LLM backbones with minimal computational overhead.

Key insights

  • Hallucination risk can be formally characterized by decomposing it into terms reflecting systematic bias from training-time knowledge flaws and stochastic variance from inference-time dynamics.
  • Reasoning-driven hallucinations exhibit an exponential amplification with increasing generation depth, while data-driven errors are more prominent at lower depths, revealing a phase transition in error propagation.
  • Neural Tangent Kernel properties provide an effective and efficient means to proxy and jointly quantify both data-driven and reasoning-driven hallucination risks by reflecting a model's representational stability and response to perturbations.

Summary by alphaXiv.

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