analiza aiqbench
This research identifies a critical safety gap in vision-language models where requests for visual grounding (points or bounding boxes) bypass refusal mechanisms that successfully block similar free-text queries. The authors demonstrate that standard system prompts fail to close this disparity and propose a specific fine-tuning strategy using self-distilled benign data to align grounding outputs with safety constraints without degrading model capabilities. This work is essential for deploying VLMs in robotic agents and downstream interfaces where precise spatial actions could otherwise lead to unintended harm or bias.
Visual Grounding Safety in Vision-Language Models
arXiv:2610.05637v1 Announce Type: new Abstract: Vision-language models (VLMs) are increasingly trained to generate structured outputs like points and bounding boxes that downstream interfaces, agents, and robots can act on, yet safety alignment of this output channel has not been systematically analyzed. We study visual grounding safety by repurposing three safety benchmarks spanning direct harm (VLSU), social bias (BBQ-V), and situational safety (Asimov-2.0) into 15,401 matched pairs of harmful requests that differ only in the requested output: a free-text answer (VQA) or a grounding (point or bounding box). Across five VLMs, models that refuse a harmful request posed as a question often comply when the same request asks for a grounding: averaged over models, grounding refusal trails VQA refusal by 31-59 percentage points, depending on the domain, and safety system prompts do not close this gap. We propose a fine-tuning approach that combines grounding-form refusals with capability grounding data and self-distilled benign data to counter over-refusal. For Qwen3-VL-8B and VisionReasoner-7B, it improves grounding refusal by 77-95 percentage points on VLSU and BBQ-V and by 64-85 points on the held-out Asimov-2.0 domain, while also improving VQA refusal, preserving grounding capability, and keeping over-refusal limited. Representation analysis shows that fine-tuning moves harmful requests toward each model's refusal direction, most strongly for grounding, while leaving benign requests near the harmless reference.
Abstract:Vision-language models (VLMs) are increasingly trained to generate structured outputs like points and bounding boxes that downstream interfaces, agents, and robots can act on, yet safety alignment of this output channel has not been systematically analyzed. We study visual grounding safety by repurposing three safety benchmarks spanning direct harm (VLSU), social bias (BBQ-V), and situational safety (Asimov-2.0) into 15,401 matched pairs of harmful requests that differ only in the requested output: a free-text answer (VQA) or a grounding (point or bounding box). Across five VLMs, models that refuse a harmful request posed as a question often comply when the same request asks for a grounding: averaged over models, grounding refusal trails VQA refusal by 31-59 percentage points, depending on the domain, and safety system prompts do not close this gap. We propose a fine-tuning approach that combines grounding-form refusals with capability grounding data and self-distilled benign data to counter over-refusal. For Qwen3-VL-8B and VisionReasoner-7B, it improves grounding refusal by 77-95 percentage points on VLSU and BBQ-V and by 64-85 points on the held-out Asimov-2.0 domain, while also improving VQA refusal, preserving grounding capability, and keeping over-refusal limited. Representation analysis shows that fine-tuning moves harmful requests toward each model's refusal direction, most strongly for grounding, while leaving benign requests near the harmless reference.
Artificial Intelligence (cs.AI); Cryptography and Security (cs.CR); Computer Vision and Pattern Recognition (cs.CV); Computers and Society (cs.CY); Machine Learning (cs.LG)
arXiv:2610.05637 [cs.AI]
(or arXiv:2610.05637v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2610.05637 arXiv-issued DOI via DataCite (pending registration)
Text integral sub licența deschisă a publicației · 247 cuvinte
Subiecte conexe
EG-ARSA: An Expert-Grounded Open Model for Visual Road Safety Auditing in Low-Resource Settings
arXiv · 25 aug. 2026
When Safety Overrides Vision: Exploring Dynamics between Vision Influence and Safety Alignment in Vision-Language Models
arXiv · 20 aug. 2026
Exploring Adversarial Robustness and Safety Alignment in Multilingual Multi-Modal Large Language Models
arXiv · 3 iun. 2026
[AINews] not much happened today
Latent Space · 10 sept. 2026