Constrained Paraphrase Consistency for LLM Hallucination Detection

2026-06-06

AI Summary

This work addresses the pervasive issue of factual hallucinations in large language model (LLM) generations by proposing a constrained optimization training framework. The approach uniquely enforces consistency under semantically equivalent rewrites and preservation of factual labels as training constraints, leveraging a small set of dual variables to enhance the reliability of factuality judgments. By jointly optimizing model parameters and Lagrange multipliers via a gradient descent-ascent algorithm, the method is implemented with DeBERTa and Flan-T5 as backbone architectures. It consistently outperforms strong baselines—including FactCG, MiniCheck, and AlignScore—on standard factuality benchmarks, achieving superior performance without incurring additional inference overhead.

Abstract

Large language models (LLMs) can generate factually inconsistent claims, motivating accurate and scalable hallucination detectors. Prior work largely enlarges training sets via synthesis or new annotations, introducing increasing cost and potential bias while underusing the consistency implied by semantically equivalent paraphrases. We propose Consistency-Constrained Hallucination Detector (CCHD), which formulates training as a constrained optimization problem. The standard cross-entropy on original document-claim pairs is complemented by (i) paraphrase-consistency constraints bounding divergence across paraphrased views, and (ii) label-preservation constraints tying paraphrases to ground truth. We solve the problem by gradient descent-ascent over model parameters and per-view Lagrange multipliers, adding only a few scalar dual variables and no inference-time overhead. With DeBERTa and Flan-T5 backbones, CCHD consistently outperforms strong baselines (FactCG, MiniCheck, and AlignScore) on standard factuality benchmarks, demonstrating its superiority on hallucination detection.

Problem

• LLM hallucination
• factuality detection
• paraphrase consistency
• constrained optimization

Innovation

• constrained optimization
• paraphrase consistency
• hallucination detection
• label preservation
• LLM factuality