Diving into Mitigating Hallucinations from a Vision Perspective for Large Vision-Language Models

Weihang Wang, Xinhao Li, Ziyue Wang, Yan Pang, Jielei Zhang, Peiyi Li, Qiang Zhang, Longwen GaoEMNLP 2025

TL;DR

This paper tackles the critical issue of "hallucinations" in Large Vision-Language Models (LVLMs), where models generate descriptions of objects or attributes not present in an image. They introduce VHBench-10, a fine-grained benchmark for evaluating these hallucinations, and propose VisionWeaver, a novel context-aware routing network that dynamically aggregates features from multiple specialized visual encoders to significantly reduce hallucinations and improve overall performance.

What Problem Does This Paper Address?

Large Vision-Language Models (LVLMs) are powerful tools for understanding and generating content from visual inputs. However, a significant problem is their tendency to "hallucinate" – to describe objects or attributes that are not actually present in the image. This undermines their reliability and real-world applicability. Existing benchmarks often focus on coarse-grained detection of object hallucinations, failing to capture the nuances of different types of visual errors. The authors hypothesize that the choice of visual encoder within an LVLM significantly impacts its hallucination behavior due to the distinct inductive biases learned during their diverse training paradigms. They aim to systematically analyze these effects and develop a method to mitigate these fine-grained hallucinations.

Methods

The core of the paper's contribution lies in two main areas:

  • VHBench-10 Benchmark: The authors introduce VHBench-10, a comprehensive benchmark designed to evaluate LVLMs across ten fine-grained hallucination categories. These categories are systematically grouped into four core visual competencies: detection, segmentation, localization, and classification. This benchmark allows for a more granular diagnosis of specific visual perceptual failures that lead to hallucinations, moving beyond simple object existence checks.
  • VisionWeaver: To address the limitations of single visual encoders and naive feature fusion, the authors propose VisionWeaver. This is a context-aware routing network that intelligently integrates multiple specialized visual encoders ("experts"). It uses a context-aware routing module that leverages global image features (specifically, the CLIP [CLS] token) to generate soft weights. These weights dynamically guide the selection and aggregation of features from the most appropriate experts for a given visual input. This allows the model to harness the complementary strengths of different encoders, leading to more accurate and grounded visual understanding.

Datasets

  • LLaVA-ReCap-118K dataset: Used for selecting images and their corresponding factual captions for constructing the VHBench-10 benchmark.
  • VHBench-10: A newly introduced benchmark comprising approximately 10,000 samples, each with an image, a factual caption, and a hallucinated caption designed to probe specific hallucination types.
  • POPE, AutoHallusion, MME, MMStar, MMBench, OCRBench, MathVista: Standard benchmarks used for evaluating the performance of LVLMs and the proposed VisionWeaver method.
  • LLaVA-Pretrain and LLaVA-Finetune datasets: Used for training the VisionWeaver model.

Key Insights and Ideas

  • Visual Encoder Matters: The choice of visual encoder significantly influences an LVLM's hallucination patterns. Different encoders have specialized strengths (e.g., CLIP for global perception, DINOv2 for fine-grained details, Vary for text recognition) that make them better or worse at specific types of visual tasks, and thus, specific types of hallucinations.
  • Fine-Grained Analysis is Crucial: Simply detecting the presence or absence of objects is insufficient. Understanding why hallucinations occur (e.g., failures in color perception, spatial reasoning, or text recognition) is key to developing effective mitigation strategies.
  • No Single Encoder is Optimal: No single visual encoder is universally best for all images and queries. Leveraging the diverse strengths of multiple encoders is a promising direction.
  • Naive Fusion is Suboptimal: Simple feature fusion techniques like addition or concatenation of features from multiple encoders often yield suboptimal performance compared to more intelligent aggregation methods.
  • Context-Aware Routing Works: Dynamically selecting and aggregating features from specialized visual experts based on the image context is an effective way to improve hallucination mitigation and overall model performance.

Key Results

  • VHBench-10 Effectiveness: Evaluations on VHBench-10 confirm that different visual encoders exhibit unique hallucination characteristics.
  • VisionWeaver Performance: VisionWeaver consistently achieved the lowest error rates across all hallucination categories on VHBench-10, significantly reducing hallucinations.
  • Benchmark Improvements: VisionWeaver demonstrated superior performance on established hallucination benchmarks like POPE and AutoHallusion, as well as general LVLM benchmarks like MMBench and OCRBench. For instance, on POPE, VisionWeaver achieved 89.5% accuracy, outperforming other methods.
  • Computational Efficiency: VisionWeaver introduces negligible additional latency, making it a practical and lightweight module for real-world applications.
  • Expert Strengths: The paper highlights specific strengths of various visual encoders: DINOv2 excels at fine-grained details, Vary at text-related tasks, SAM at object segmentation, ConvNext for general feature extraction, and EVA02 for general visual representations.

Limitations and Future Work

  • Benchmark Limitations: VHBench-10, while comprehensive, still has about 20% of hard-to-categorize realistic samples. The benchmark's reliance on GPT-4 for hallucination generation may introduce slight errors. Experiments were conducted on smaller-scale models due to computational resource limitations.
  • Future Work:
    • Exploring the possibility of distilling the collective knowledge of multiple experts into a single, more efficient encoder.
  • Deeper analysis of the routing mechanism to understand which visual cues trigger specific experts.
  • Determining the optimal quantity and composition of the expert pool for VisionWeaver.
  • Investigating how VisionWeaver can synergize with other mitigation strategies like improved training data or decoding-time algorithms.

Summary updated

BibTeX
@inproceedings{wang2025diving,
  author = {Weihang Wang and Xinhao Li and Ziyue Wang and Yan Pang and Jielei Zhang and Peiyi Li and Qiang Zhang and Longwen Gao},
  title = {{Diving into Mitigating Hallucinations from a Vision Perspective for Large Vision-Language Models}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure