Data-efficient Meta-models for Evaluation of Context-based Questions and Answers in LLMs

TL;DR

This paper explores data-efficient methods for evaluating hallucinations in large language models, demonstrating that high performance can be achieved with significantly fewer annotated samples using optimized classification and dimensionality reduction techniques.

Contribution

It introduces a novel approach combining efficient classifiers and dimensionality reduction to reduce data annotation needs for hallucination detection in LLMs.

Findings

Achieves comparable performance with only 250 training samples

Reduces reliance on extensive annotated datasets

Demonstrates effectiveness on standardized QA benchmarks

Abstract

Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) systems are increasingly deployed in industry applications, yet their reliability remains hampered by challenges in detecting hallucinations. While supervised state-of-the-art (SOTA) methods that leverage LLM hidden states -- such as activation tracing and representation analysis -- show promise, their dependence on extensively annotated datasets limits scalability in real-world applications. This paper addresses the critical bottleneck of data annotation by investigating the feasibility of reducing training data requirements for two SOTA hallucination detection frameworks: Lookback Lens, which analyzes attention head dynamics, and probing-based approaches, which decode internal model representations. We propose a methodology combining efficient classification algorithms with dimensionality reduction techniques to minimize sample size demands while maintaining competitive performance. Evaluations on standardized question-answering RAG benchmarks show that our approach achieves performance comparable to strong proprietary LLM-based baselines with only 250 training samples. These results highlight the potential of lightweight, data-efficient paradigms for industrial deployment, particularly in annotation-constrained scenarios.