Reproduce “Data-efficient Meta-models for Evaluation of Context-based Questions and Answers 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
Julia Belikova, Konstantin Polev, Rauf Parchiev, Dmitry Simakov
The authors evaluate a detector for context-inconsistent answers in question answering. It extracts model signals while processing context and answer, then trains a small classifier. PCA, lookback features, and TabPFN reach ROC-AUC scores of 0.6972, 0.8165, and 0.7679 on EManual, ExpertQA, and RAGTruth. This nearly matches GPT-4o RAGAS on ExpertQA but trails it elsewhere; table values do not specify training-set size.
Key insights
- TabPFNv2 proved to be the top-performing meta-classifier for hallucination detection in low-data settings, demonstrating the effectiveness of tabular foundation models in NLP classification tasks.
- Principal Component Analysis (PCA) consistently enhanced the performance of lookback features and effectively managed high dimensionality.
- Relatively small, open-source LLMs (like Llama-3.1-8B and Qwen2.5-7B) can serve as effective extractors for internal activations, sometimes outperforming specialized proprietary evaluation frameworks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.