Overview
The HalluScope project addresses the issue of hallucinations in Multimodal Large Language Models (MLLMs). Hallucinations are defined as model outputs inconsistent with visual content, textual context, or commonsense knowledge. The proposed approach focuses on fine-grained hallucination diagnosis, a unified task combining hallucination detection, classification, and interpretable explanation generation, contrasting with existing coarse-grained detection methods.
Research Context
MLLMs have demonstrated performance across various vision-language tasks. However, the occurrence of hallucinations remains a significant challenge. Prior efforts primarily addressed this through coarse-grained detection, which often lacks the detailed diagnostic information necessary for understanding specific hallucination types or for supporting subsequent mitigation strategies.
Approach
To enable fine-grained hallucination diagnosis, the researchers developed several components:
- Automated Data Generation Pipeline: An automated pipeline was created for generating diagnostic data.
- HalluScope-30K Dataset: A large-scale diagnostic dataset, HalluScope-30K, was constructed. This dataset covers eight distinct hallucination sources and five different task categories.
- Multi-granular Joint Reward Function: To facilitate the training of diagnosis models, a multi-granular joint reward function was designed.
- Diagnosis Models: Two specific diagnosis models, HalluScope-4B and HalluScope-8B, were trained based on the HalluScope-30K dataset and the defined reward function.
Findings
- The HalluScope-4B and HalluScope-8B diagnosis models demonstrated state-of-the-art performance against two established benchmarks: the MHALO benchmark and the newly introduced fine-grained hallucination classification benchmark.
- The study indicated that detection and classification tasks exhibit mutual benefit when optimized jointly.
- Through diagnosis-driven feedback experiments, the fine-grained diagnostic explanations generated by the HalluScope models were observed to effectively guide target models in correcting their hallucinations.
- Full diagnosis, incorporating detection, classification, and explanation generation, significantly outperformed all baseline approaches in hallucination correction experiments on both Qwen3-VL-8B-Instruct and LLaVA-1.5-7B models.
Why This Matters
The development of fine-grained hallucination diagnosis offers a more detailed understanding of MLLM inconsistencies compared to coarse detection. The ability to classify hallucination types and generate interpretable explanations provides specific information that can be utilized to guide the correction of these errors in target models. This refined diagnostic capability contributes to improving the reliability and accuracy of MLLM outputs.