Seeing is Believing: Rich-Context Hallucination Detection for MLLMs via Backward Visual Grounding

2025-11-15

AI Summary

To address the pervasive hallucination problem in multimodal large language models (MLLMs) on cross-modal tasks, this paper proposes VBackChecker—the first reference-free, pixel-level hallucination detection framework. Its core innovation is the “backward visual grounding” mechanism: leveraging vision-language grounding models to perform fine-grained visual backtracking, jointly integrating reasoning and referring segmentation capabilities to verify pixel-level consistency between generated text and the original image across objects, attributes, and relations. To support this, we design an instruction-tuning data pipeline enabling context-aware grounding mask generation and hard negative sample mining, and introduce R²-HalBench—the first hallucination evaluation benchmark targeting real-world complex scenarios. Experiments demonstrate that VBackChecker achieves state-of-the-art performance on R²-HalBench, matching GPT-4o’s accuracy while improving pixel-level localization precision by over 10%.

Technology Category

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

Application Category

• 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
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Multimodal Large Language Models (MLLMs) have unlocked powerful cross-modal capabilities, but still significantly suffer from hallucinations. As such, accurate detection of hallucinations in MLLMs is imperative for ensuring their reliability in practical applications. To this end, guided by the principle of "Seeing is Believing", we introduce VBackChecker, a novel reference-free hallucination detection framework that verifies the consistency of MLLMgenerated responses with visual inputs, by leveraging a pixellevel Grounding LLM equipped with reasoning and referring segmentation capabilities. This reference-free framework not only effectively handles rich-context scenarios, but also offers interpretability. To facilitate this, an innovative pipeline is accordingly designed for generating instruction-tuning data (R-Instruct), featuring rich-context descriptions, grounding masks, and hard negative samples. We further establish R^2 -HalBench, a new hallucination benchmark for MLLMs, which, unlike previous benchmarks, encompasses real-world, rich-context descriptions from 18 MLLMs with high-quality annotations, spanning diverse object-, attribute, and relationship-level details. VBackChecker outperforms prior complex frameworks and achieves state-of-the-art performance on R^2 -HalBench, even rivaling GPT-4o's capabilities in hallucination detection. It also surpasses prior methods in the pixel-level grounding task, achieving over a 10% improvement. All codes, data, and models are available at https://github.com/PinxueGuo/VBackChecker.

Problem

• Detecting hallucinations in Multimodal Large Language Models responses
• Verifying consistency between generated text and visual inputs
• Handling rich-context scenarios without requiring reference data

Innovation

• Uses pixel-level Grounding LLM for visual verification
• Generates rich-context instruction data with hard negatives
• Creates benchmark with real-world multimodal hallucination annotations