Propose and Attend: Training-free MLLM Grounding Confidence via Multi-Token Localized Attention

2026-07-07

AI Summary

This work addresses the challenge of hallucination in multimodal large language models (MLLMs) when generating localized predictions—such as bounding boxes or temporal segments—and the inadequacy of raw token confidence scores in reflecting localization quality. To this end, the authors propose Multi-Token Local Attention (MTLA), a training-free post-processing method that constructs a task- and modality-agnostic confidence score by aggregating local attention weights across all output tokens within the predicted region. MTLA is universally applicable across diverse modalities, including images, videos, and audio. Experimental results demonstrate that MTLA substantially improves hallucination detection performance, yielding AUROC gains of 7–38 points across multiple MLLMs. When integrated into a re-ranking pipeline, it elevates the zero-shot detection AP of an open-source 8B model on COCO from 20.4 to 37.0.

Abstract

Multimodal large language models can emit localized predictions, bounding boxes for objects and temporal windows for video and audio events, but they hallucinate these regions prolifically. The model's own token log-probabilities are nearly uninformative: they conflate grounding quality with input ambiguity, and coordinate tokens become near-deterministic once the model commits. We propose Multi-Token Localized Attention (MTLA): a training-free, post-hoc score that measures how strongly a prediction's tokens attend to the region they claim. Prior attention-based detectors, which sum attention over the entire input modality and read a single response token, are weaker special cases; we show that summing only within the claimed region and aggregating across all prediction tokens recovers a stronger grounding signal. The same recipe applies almost trivially to other modalities and tasks: object detection in images and temporal localization in video and audio. Across multiple MLLM families and three modalities, MTLA improves hallucination AUROC by +7 to +38 over the best prior training-free baseline. Used as a confidence score for re-ranking, it nearly doubles the zero-shot COCO detection AP of an open-source 8B generalist (from 20.4 to 37.0), narrowing the gap to supervised detectors without any task-specific training.

Problem

• multimodal large language models
• grounding confidence
• hallucination
• localized predictions
• attention-based detection

Innovation

• Multi-Token Localized Attention
• training-free
• grounding confidence
• multimodal large language models
• hallucination detection