PruneHal: Reducing Hallucinations in Multi-modal Large Language Models through Adaptive KV Cache Pruning

2025-10-21

AI Summary

Hallucination in multimodal large language models (MLLMs) demands efficient, low-overhead mitigation strategies. This paper proposes a training-free, zero-parameter-update adaptive KV cache pruning method—the first to introduce token-level pruning for hallucination suppression in MLLMs. It dynamically identifies and removes redundant visual tokens via attention-based analysis, thereby enhancing model focus on salient visual cues. The approach is architecture-agnostic, compatible with mainstream MLLMs (e.g., LLaVA, Qwen-VL) and diverse decoding strategies, with negligible inference latency overhead. Evaluated across four representative MLLMs and multiple standard hallucination benchmarks (e.g., HalluBench, MME-Hallu), it achieves consistent reductions in hallucination rates—averaging 12.6%–23.4%—while demonstrating strong robustness. This work establishes a lightweight, plug-and-play paradigm for hallucination mitigation in MLLMs.

Technology Category

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

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

While multi-modal large language models (MLLMs) have made significant progress in recent years, the issue of hallucinations remains a major challenge. To mitigate this phenomenon, existing solutions either introduce additional data for further training or incorporate external or internal information during inference. However, these approaches inevitably introduce extra computational costs. In this paper, we observe that hallucinations in MLLMs are strongly associated with insufficient attention allocated to visual tokens. In particular, the presence of redundant visual tokens disperses the model's attention, preventing it from focusing on the most informative ones. As a result, critical visual cues are often under-attended, which in turn exacerbates the occurrence of hallucinations. Building on this observation, we propose extbf{PruneHal}, a training-free, simple yet effective method that leverages adaptive KV cache pruning to enhance the model's focus on critical visual information, thereby mitigating hallucinations. To the best of our knowledge, we are the first to apply token pruning for hallucination mitigation in MLLMs. Notably, our method don't require additional training and incurs nearly no extra inference cost. Moreover, PruneHal is model-agnostic and can be seamlessly integrated with different decoding strategies, including those specifically designed for hallucination mitigation. We evaluate PruneHal on several widely used hallucination evaluation benchmarks using four mainstream MLLMs, achieving robust and outstanding results that highlight the effectiveness and superiority of our method. Our code will be publicly available.

Problem

• Reducing hallucinations in multi-modal large language models
• Addressing insufficient attention to critical visual tokens
• Mitigating redundant visual tokens dispersing model focus

Innovation

• Adaptive KV cache pruning reduces hallucinations
• Training-free method enhances visual attention focus
• Model-agnostic approach integrates with decoding strategies