MLLM can see? Dynamic Correction Decoding for Hallucination Mitigation

2024-10-15

arXiv.org

AI Summary

Multimodal large language models (MLLMs) often suppress accurate visual information preserved in early layers due to overly dominant linguistic priors, leading to hallucination. We find that such hallucination does not stem from visual perception failure but rather from imbalanced cross-layer knowledge fusion. To address this, we propose DeCo—a model-agnostic, dynamic decoding-time correction method. During autoregressive decoding, DeCo dynamically analyzes per-layer logits to adaptively select and weight visual features from preceding layers, injecting them into the final layer for real-time, vision-guided correction. DeCo integrates seamlessly with standard decoding strategies—including sampling and beam search—without requiring fine-tuning or additional training. Evaluated across multiple standard benchmarks, DeCo reduces hallucination rates by over 40% on average, significantly outperforming existing post-hoc mitigation methods. The implementation is publicly available.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Multimodal Large Language Models (MLLMs) frequently exhibit hallucination phenomena, but the underlying reasons remain poorly understood. In this paper, we present an empirical analysis and find that, although MLLMs incorrectly generate the objects in the final output, they are actually able to recognize visual objects in the preceding layers. We speculate that this may be due to the strong knowledge priors of the language model suppressing the visual information, leading to hallucinations. Motivated by this, we propose a novel dynamic correction decoding method for MLLMs DeCo, which adaptively selects the appropriate preceding layers and proportionally integrates knowledge into the final layer to adjust the output logits. Note that DeCo is model agnostic and can be seamlessly incorporated with various classic decoding strategies and applied to different MLLMs. We evaluate DeCo on widely-used benchmarks, demonstrating that it can reduce hallucination rates by a large margin compared to baselines, highlighting its potential to mitigate hallucinations. Code is available at https://github.com/zjunlp/DeCo.

Problem

• Mitigate hallucinations in MLLMs
• Dynamic correction decoding method
• Integrate knowledge adaptively

Innovation

• Dynamic correction decoding method
• Adaptive layer selection integration
• Model agnostic hallucination mitigation