Reproduce “HalluCounter: Reference-free LLM Hallucination Detection in the Wild!” 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

Ashok Urlana, Gopichand Kanumolu, Charaka Vinayak Kumar, Bala Mallikarjunarao Garlapati, Rahul Mishra

HalluCounter, a reference-free hallucination detection method from IIIT Hyderabad and TCS Research, integrates both query-response alignment and response-response consistency to identify factual inaccuracies in LLM outputs. It consistently outperforms state-of-the-art black-box approaches by over 10% F1-score and introduces HalluCounterEval, a large-scale, multi-domain benchmark dataset.

Key insights

  • Integrating both query-response alignment and response-response consistency significantly improves reference-free hallucination detection accuracy compared to methods relying solely on self-consistency among generated responses.
  • The performance of HalluCounter's pipeline remains largely stable and resilient to the exact number of generated responses, consistently achieving high confidence scores even with as few as three samples.
  • LLMs exhibit varying hallucination tendencies across domains and model sizes, with smaller models and tasks requiring complex reasoning (e.g., mathematical or logical skills) being more susceptible to generating incorrect information.

Summary by alphaXiv.

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