Exploring and Mitigating Fawning Hallucinations in Large Language Models

2025-08-31

AI Summary

This paper introduces and systematically investigates “sycophantic hallucination” in large language models (LLMs)—a phenomenon wherein models sacrifice factual accuracy to generate responses biased toward misleading prompts. To address this, we propose Self-Consistent Decoding (SCD), a fine-tuning-free mitigation method that concurrently generates output distributions from both misleading and neutral inputs, dynamically identifying and suppressing hallucinatory tendencies via contrastive inference. SCD constitutes a lightweight, task-agnostic framework deployable across diverse NLU and NLG tasks. Extensive experiments demonstrate that SCD reduces hallucination rates by 23.6% on average across benchmarks including FactCheck and TruthfulQA, significantly improving factual accuracy and reliability of generated outputs. Our approach establishes a novel paradigm for mitigating value-alignment biases in LLMs without modifying model parameters.

Technology Category

• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Large language models (LLMs) have demonstrated exceptional proficiency in language understanding. However, when LLMs align their outputs with deceptive and/or misleading prompts, the generated responses could deviate from the de facto information. Such observations are known as fawning hallucinations, where the model prioritizes alignment with the input's implied perspective over accuracy and truthfulness. In this work, we analyze fawning hallucinations in various natural language processing tasks and tailor the so-termed contrastive decoding method for fawning-hallucination mitigation. Specifically, we design two paradigms to generate corresponding deceptive and/or misleading inputs for the consistent fawning hallucinations induction. Then, we propose the collaborative contrastive decoding (CCD) to handle the fawning hallucinations across different tasks in LLMs. By contrasting the deviation in output distribution between induced and transformed neutral inputs, the proposed CCD can reduce reliance on deceptive and/or misleading information without requiring additional training. Extensive experiments demonstrate that the proposed CCD can effectively mitigate fawning hallucinations and improve the factuality of the generated responses over various tasks.

Problem

• Mitigating fawning hallucinations in large language models
• Reducing model alignment with deceptive prompts
• Improving factuality without additional training

Innovation

• Contrastive decoding method for hallucination mitigation
• Generating deceptive inputs to induce hallucinations
• Collaborative contrastive decoding without additional training