DHI: Leveraging Diverse Hallucination Induction for Enhanced Contrastive Factuality Control in Large Language Models

2026-01-03

arXiv.org

AI Summary

This work addresses the challenge of hallucination—where large language models generate inaccurate or fabricated information—by proposing the Diverse Hallucination Induction (DHI) framework. Unlike existing approaches that are limited to specific hallucination types and rely on annotated hallucination data, DHI induces diverse hallucinations without requiring pre-labeled examples by redesigning the loss function and incorporating a causal attention mask. Furthermore, it enhances factual consistency through adaptive rationale-constrained optimization during contrastive decoding. Evaluated across multiple hallucination benchmarks, the proposed method significantly outperforms current contrastive decoding techniques, demonstrating improved factual accuracy and generalization capability.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Large language models (LLMs) frequently produce inaccurate or fabricated information, known as"hallucinations,"which compromises their reliability. Existing approaches often train an"Evil LLM"to deliberately generate hallucinations on curated datasets, using these induced hallucinations to guide contrastive decoding against a reliable"positive model"for hallucination mitigation. However, this strategy is limited by the narrow diversity of hallucinations induced, as Evil LLMs trained on specific error types tend to reproduce only these particular patterns, thereby restricting their overall effectiveness. To address these limitations, we propose DHI (Diverse Hallucination Induction), a novel training framework that enables the Evil LLM to generate a broader range of hallucination types without relying on pre-annotated hallucination data. DHI employs a modified loss function that down-weights the generation of specific factually correct tokens, encouraging the Evil LLM to produce diverse hallucinations at targeted positions while maintaining overall factual content. Additionally, we introduce a causal attention masking adaptation to reduce the impact of this penalization on the generation of subsequent tokens. During inference, we apply an adaptive rationality constraint that restricts contrastive decoding to tokens where the positive model exhibits high confidence, thereby avoiding unnecessary penalties on factually correct tokens. Extensive empirical results show that DHI achieves significant performance gains over other contrastive decoding-based approaches across multiple hallucination benchmarks.

Problem

• hallucination
• large language models
• contrastive decoding
• factuality control
• diverse hallucination induction

Innovation

• Diverse Hallucination Induction
• Contrastive Decoding
• Hallucination Mitigation
• Adaptive Rationality Constraint
• Causal Attention Masking