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

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

Multimodal large language models (MLLMs) frequently hallucinate localized predictions, with 58% to 68% of regions emitted by state-of-the-art models not corresponding to valid objects or events. Token log-probabilities are uninformative for grounding quality. Amazon researchers propose Multi-Token Localized Attention (MTLA), a training-free, post-hoc score that measures how strongly a prediction's tokens attend to their claimed region. MTLA aggregates attention within the proposed region across all prediction tokens, improving hallucination AUROC by +7 to +21 over prior baselines across multiple MLLM families (Qwen3-VL-8B-Instruct, Gemma-4 E4B-it, Audio Flamingo 3) and three modalities (image, video, audio). Used for re-ranking, MTLA nearly doubles the zero-shot COCO detection AP of an 8B generalist from 20.4 to 37.0, narrowing the gap to supervised detectors without task-specific training.

Key takeaway

For AI Scientists and ML Engineers deploying MLLMs for localization tasks, you should integrate Multi-Token Localized Attention (MTLA) as a training-free confidence score. This method significantly reduces hallucinated predictions and boosts accuracy, nearly doubling zero-shot COCO detection AP from 20.4 to 37.0 for an 8B generalist. Consider MTLA to enhance reliability and close the performance gap to supervised systems without additional task-specific training.

Key insights

MLLM hallucination detection improves by localizing attention to predicted regions and aggregating across all output tokens.

Principles

• Grounded predictions attend strongly to their claimed region.
• Hallucinated predictions rely on contextual distractors.
• Aggregating attention across multiple tokens is robust.

Method

MTLA aggregates decoder attention from all prediction tokens (coordinates, label) within the model's own proposed region, then averages across middle transformer layers (L8-21 for Qwen3-VL-8B, Gemma-4).

In practice

• Apply MTLA post-hoc to MLLM localization outputs.
• Use MTLA for re-ranking to suppress hallucinated predictions.
• Extend MTLA to image, video, and audio grounding tasks.

Topics

• Multimodal LLMs
• Object Hallucination
• Localization Confidence
• Attention Mechanisms
• Zero-shot Detection
• Temporal Grounding

Code references

• TalRemez/MTLA
• simedw/coco-gemini

Best for: Research Scientist, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Propose and Attend: Training-free MLLM Grounding Confidence via Multi-Token Localized Attention — MTLA offers a training-free, post-hoc method to assess MLLM localized prediction confidence by analyzing multi-token attention within claimed regions.
• ADAPT: Attention Dynamics Alignment with Preference Tuning for Faithful MLLMs — MLLM hallucination stems from progressive text-to-image cross-attention degradation, which ADAPT mitigates by aligning attention dynamics.
• Mechanistic Insights into Functional Sparsity in Multimodal LLMs via CoRe Heads — Multimodal LLMs exhibit functional sparsity, relying on a sparse subset of "CoRe heads" for efficient, precise cross-modal visual information retrieval.
• Grad Detect: Gradient-Based Hallucination Detection in LLMs — LLM hallucinations can be detected efficiently by analyzing internal layer-wise gradient patterns from a single forward-backward pass.
• MLLMs Know When Before Speaking: Revealing and Recovering Temporal Grounding via Attention Cues — MLLMs internally know event timings during prefill, but often lose this signal during answer generation, a gap recoverable via attention-guided re-inference.
• Detecting Hallucinations in SpeechLLMs at Inference Time Using Attention Maps — Attention-derived metrics enable efficient, inference-time hallucination detection in SpeechLLMs without gold-standard outputs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.