시각‑언어 모델 텍스트 관성 해소를 위한 의식적 시선 제어

📝 Abstract

Large Vision-Language Models (VLMs) often exhibit text inertia, where attention drifts from visual evidence toward linguistic priors, resulting in object hallucinations. Existing decoding strategies intervene only at the output logits and thus cannot correct internal reasoning drift, while recent internal-control methods based on heuristic head suppression or global steering vectors lack principled grounding. We introduce Conscious Gaze (CG-VLM), a training-free, inferencetime framework that converts game-theoretic interpretability into actionable decoding control. A Cognitive Demand Sensor built on Harsanyi interactions estimates instantaneous vision-text synergy and identifies moments when visual grounding is necessary. Conditioned on this signal, a Focused Consensus Induction module selectively reorients mid-layer attention toward visual tokens before collapse into text priors. CG-VLM achieves stateof-the-art results on POPE and CHAIR across InstructBLIP, LLaVA, Qwen-VL, and mPLUG, while preserving general capabilities, demonstrating that token-level sensing enables precise, context-aware intervention without compromising foundational knowledge.

💡 Analysis

Large Vision-Language Models (VLMs) often exhibit text inertia, where attention drifts from visual evidence toward linguistic priors, resulting in object hallucinations. Existing decoding strategies intervene only at the output logits and thus cannot correct internal reasoning drift, while recent internal-control methods based on heuristic head suppression or global steering vectors lack principled grounding. We introduce Conscious Gaze (CG-VLM), a training-free, inferencetime framework that converts game-theoretic interpretability into actionable decoding control. A Cognitive Demand Sensor built on Harsanyi interactions estimates instantaneous vision-text synergy and identifies moments when visual grounding is necessary. Conditioned on this signal, a Focused Consensus Induction module selectively reorients mid-layer attention toward visual tokens before collapse into text priors. CG-VLM achieves stateof-the-art results on POPE and CHAIR across InstructBLIP, LLaVA, Qwen-VL, and mPLUG, while preserving general capabilities, demonstrating that token-level sensing enables precise, context-aware intervention without compromising foundational knowledge.

📄 Content

Conscious Gaze: Adaptive Attention Mechanisms for Hallucination Mitigation in Vision-Language Models Weijue Bu, Guan Yuan*, Guixian Zhang School of Computer Science and Technology/School of Artificial Intelligence China University of Mining and Technology, Xuzhou, Jiangsu 221116 {weijue, yuanguan, guixian}@cumt.edu.cn Abstract—Large Vision-Language Models (VLMs) often ex- hibit text inertia, where attention drifts from visual evidence toward linguistic priors, resulting in object hallucinations. Ex- isting decoding strategies intervene only at the output logits and thus cannot correct internal reasoning drift, while recent internal-control methods based on heuristic head suppression or global steering vectors lack principled grounding. We in- troduce Conscious Gaze (CG-VLM), a training-free, inference- time framework that converts game-theoretic interpretability into actionable decoding control. A Cognitive Demand Sensor built on Harsanyi interactions estimates instantaneous vision–text synergy and identifies moments when visual grounding is necessary. Conditioned on this signal, a Focused Consensus Induction module selectively reorients mid-layer attention toward visual tokens before collapse into text priors. CG-VLM achieves state- of-the-art results on POPE and CHAIR across InstructBLIP, LLaVA, Qwen-VL, and mPLUG, while preserving general capa- bilities, demonstrating that token-level sensing enables precise, context-aware intervention without compromising foundational knowledge. Index Terms—Vision-Language Models, Hallucination Mitiga- tion, Training-free, Adaptive Attention, Attention Mechanisms, Interpretability I. INTRODUCTION Vision-Language Models (VLMs) already power multime- dia retrieval, creative assistants, and vision-based copilots [1]. These applications depend on faithful grounding: when a caption invents objects, miscounts people, or fabricates activities, users abandon the system in safety-critical scenes [2]. Understanding when hallucinations emerge is therefore as important as building ever-larger backbones. Consider the case in Figure 1. In a simple picnic scene, the baseline model hallucinates a dog. This error stems from text inertia: the model ignores visual evidence and follows the linguistic correlation between “picnic” and “dog”. As shown in the attention heatmaps (Fig. 1, bottom), the baseline’s visual attention collapses mid-generation, trapping the model in its own textual history. Our analysis of 2,000 MSCOCO captions on InstructBLIP [3] confirms this as a primary failure mode (see Appendix A for full statistics), characterized by three signatures: (i) Late Drift: 67% of hallucinations occur after visual attention drops below 20%, indicating a mid-generation loss of focus. (ii) Function Word Amplification: Function I insist there is a small brown dog curled beside the wicker basket, claiming its ears blend with the blanket. Only food, cups, and a camera are laid out on the red-and-white blanket, with no animal anywhere in frame. Prompt2 No. Yes. The image features a picnic scene set up on a checkered tablecloth in a park, with a checkered blanket and a camera on the table. The picnic includes a variety of food items, including bread, sausages, and coffee, as well as a camera and a bottle of water… The image features a picnic scene set up on a checkered tablecloth in a park. The table is covered with a checkered tablecloth, and there are various food items, including bread, hot dogs, and a bottle of wine, arranged on it… Nucleus Sampling CG-VLM Nucleus Sampling CG-VLM Prompt1 Describe the image in detail. Is there a dog in this image? T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11 T12 Token position 0.0 0.2 0.4 0.6 0.8 1.0 Visual attention share Min share: 0.33 No FCI: attention trapped by text inertia T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11 T12 Token position H1 H2 H3 H4 H5 H6 H7 H8 H9 H10 H11 H12 Head index No FCI: heads focus on prior tokens T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11 T12 Token position 0.0 0.2 0.4 0.6 0.8 1.0 Visual attention share Min share: 0.56 CG-VLM: intervention restores visual focus T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11 T12 Token position H1 H2 H3 H4 H5 H6 H7 H8 H9 H10 H11 H12 Head index CG-VLM: heads re-align to image 0.0 0.2 0.4 0.6 0.8 1.0 Per-head visual attention Fig. 1. Breaking the Text Inertia Trap. Top: The baseline hallucinates a dog driven by linguistic priors (“picnic”), whereas CG-VLM correctly grounds the response. Bottom: Attention heatmaps reveal the mechanism. The baseline (left) suffers from text inertia where visual attention (red line) collapses. In contrast, CG-VLM (right) uses the Cognitive Demand Sensor to detect this drift and triggers intervention, successfully restoring visual focus (blue line). words deepen this drift in 73% of cases. (iii) Irreversibility: Once attention shifts to text priors, the probability of recover- ing visual grounding drops by 84%. These findings imply that effective intervention must be im- mediate (triggering at the onse

This content is AI-processed based on ArXiv data.