Mitigating multimodal hallucinations through visual attention tracing and origin-point regeneration.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Natural Language Processing

Background:

• Multimodal large language models (MLLMs) excel at vision-language tasks but suffer from hallucinations, where generated text contradicts visual input.
• Existing methods for hallucination mitigation are often inefficient, requiring extensive retraining or applying broad penalties without pinpointing error origins.

Purpose of the Study:

• To introduce Hallucination Backtracking (HB), a training-free decoding framework for detecting and mitigating hallucinations in MLLMs.
• To develop a method for precisely localizing the source of hallucinations by analyzing visual attention dynamics.

Main Methods:

• HB monitors visual attention during generation, identifying pivotal tokens where the model's focus shifts from visual to textual information.
• A novel Visual Attention Score (VAS) quantifies this attentional drift, enabling origin-point detection.
• Upon detecting an anomaly, HB backtracks to the divergence point and triggers regeneration with stricter visual grounding.

Main Results:

• HB achieved high accuracy in localizing hallucination origins (41.8% exact match, 84.1% before-first).
• Evaluations across LLaVA-1.5, InstructBLIP, MiniGPT-4, and Shikra showed HB surpassing state-of-the-art methods.
• On LLaVA-1.5, HB improved the POPE F1 score to 91.4% and reduced the CHAIR metric to 40.2%.

Conclusions:

• Hallucination Backtracking (HB) offers a precise and efficient approach to mitigate errors in MLLMs by targeting specific divergence points.
• The method demonstrates significant improvements in hallucination detection and correction across various MLLM architectures.
• While effective, a residual false negative rate highlights ongoing challenges in fully addressing inference-driven hallucinations.