Reproduce “Teaching with Lies: Curriculum DPO on Synthetic Negatives for Hallucination Detection” 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
Shrey Pandit, Ashwin Vinod, Liu Leqi, Ying Ding
HaluCheck from The University of Texas at Austin enhances hallucination detection in large language models by combining Direct Preference Optimization with a curriculum learning approach that leverages carefully crafted, high-quality hallucinated content as negative training examples. This method achieves significant performance improvements across both in-domain and zero-shot benchmarks for factual accuracy.
Key insights
- Utilizing carefully crafted, high-quality hallucinated content as dispreferred examples ("teaching with lies") significantly enhances a model's ability to detect factual inaccuracies.
- Integrating curriculum learning with Direct Preference Optimization (DPO) allows models to learn progressively from easier to harder hallucination examples, leading to more robust detection capabilities.
- A systematic framework for assessing hallucination difficulty, such as using MiniCheck, is crucial for structuring an effective training curriculum.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.