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

TL;DR

This paper introduces DHI, a training framework for large language models that enhances hallucination diversity and reduces factual errors by inducing a wider range of hallucinations without needing pre-annotated data, improving reliability.

Contribution

DHI enables the Evil LLM to generate diverse hallucinations through a novel loss function and attention masking, advancing contrastive factuality control without relying on pre-labeled hallucination datasets.

Findings

DHI significantly outperforms existing contrastive decoding methods on multiple benchmarks.

The approach effectively increases hallucination diversity while maintaining factual accuracy.

Empirical results demonstrate improved factuality and robustness in LLM outputs.

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.