HALT-RAG: A Task-Adaptable Framework for Hallucination Detection with Calibrated NLI Ensembles and Abstention
This addresses the critical challenge of ensuring safe deployment of generative language models by providing a flexible, task-adaptable framework for hallucination detection, though it is incremental as it builds on existing NLI models and benchmarks.
The paper tackles the problem of detecting hallucinations in Retrieval-Augmented Generation outputs by introducing HALT-RAG, a post-hoc verification system that achieves strong out-of-fold F1-scores of 0.7756, 0.7391, and 0.9786 on summarization, dialogue, and QA tasks, respectively, using calibrated NLI ensembles and an abstention mechanism.
Detecting content that contradicts or is unsupported by a given source text is a critical challenge for the safe deployment of generative language models. We introduce HALT-RAG, a post-hoc verification system designed to identify hallucinations in the outputs of Retrieval-Augmented Generation (RAG) pipelines. Our flexible and task-adaptable framework uses a universal feature set derived from an ensemble of two frozen, off-the-shelf Natural Language Inference (NLI) models and lightweight lexical signals. These features are used to train a simple, calibrated, and task-adapted meta-classifier. Using a rigorous 5-fold out-of-fold (OOF) training protocol to prevent data leakage and produce unbiased estimates, we evaluate our system on the HaluEval benchmark. By pairing our universal feature set with a lightweight, task-adapted classifier and a precision-constrained decision policy, HALT-RAG achieves strong OOF F1-scores of 0.7756, 0.9786, and 0.7391 on the summarization, QA, and dialogue tasks, respectively. The system's well-calibrated probabilities enable a practical abstention mechanism, providing a reliable tool for balancing model performance with safety requirements.