A$^2$Safe is a framework for safe and effective Visual Question Answering that aligns counterfactual evidence with adaptive agent collaboration. It uses a Grounded Safety Evidence Board to make safety decisions explicit, enforcing invariance to safety‑irrelevant changes while allowing appropriate transitions when risk‑critical evidence changes. The system achieves a 95.72 SIUO safety score, reduces benign refusals on MOSSBench to 14.67%, and maintains a 78.34 average VQA score with 27.8% token overhead.
arXiv:2608.23313v1 Announce Type: new Abstract: Vision-language model safety benchmarks typically evaluate only final responses: whether a model refuses, warns, or complies. This outcome-level view c...
arXiv:2608. 04509v1 Announce Type: new Abstract: Vision-language systems combine images with retrieved text, but these sources can disagree or jointly fail to support an answer.
ReFrame is a training‑free framework that enhances safety alignment for multimodal large language models at test time. It uses two lightweight agents: one generates risk and utility evidence, and the other rewrites prompts and routes images to create a safe proxy before invoking the deployed MLLM. Experiments show that ReFrame improves jailbreak defense, safety awareness, and reduces over‑sensitivity while maintaining multimodal utility.
The paper introduces MINT‑Safe, a new open‑source dataset of 11,270 multi‑image dialogues and 500 refusal VQA pairs designed to expose safety risks in multi‑modal large language models during open‑ended conversations. It also proposes TAD‑Align, a turn‑aware dual‑objective reward framework that dynamically up‑weights dialogue turns with inconsistent safety behavior, improving safety metrics on Qwen2.5‑VL‑7B‑Instruct and LLaVA‑Next‑7B. The results show over 10% reduction in attack success rate and notable gains in harmlessness and helpfulness while maintaining overall model performance.
Vision-language models (VLMs) are increasingly deployed in consumer, medical, financial, and enterprise applications. This broad deployment expands the safety surface: risks can arise from multimodal question answering, assistant responses, and cross-modal composition, while moderation policies may vary across products, regions, and deployment stages.
arXiv:2607. 16311v1 Announce Type: cross Abstract: Vision-language models (VLMs) often answer visual questions using learned language and category priors rather than grounding their predictions in the image itself.
EDCT-Bench is a benchmark that evaluates the faithfulness of Vision‑Language Models (VLMs) by using Explanation‑Driven Counterfactual Testing (EDCT). EDCT extracts visual concepts from a model’s natural language explanation, applies minimal verified edits to those concepts, and checks whether the model’s answer and explanation remain consistent with the edited image. The benchmark covers three domains—knowledge‑intensive VQA, safety‑critical driving, and 3D spatial reasoning—and reveals significant faithfulness gaps in current VLMs, while also showing that EDCT‑generated counterfactuals can improve training.
The paper investigates cross‑modal safety drift in multimodal large language models, where a harmless text query paired with a visual image can trigger harmful responses. Empirical analysis identifies unsafe response patterns and shows that visual cues receive limited attention, weakening refusal mechanisms. The authors introduce Safety‑Awareness Representation Transfer (SRT), a lightweight method that transfers safety signals from text processing to mitigate cross‑modal drift while maintaining model utility.
Aligned vision‑language models (VLMs) are designed to combine grounded visual reasoning with safe generation. The study finds that when safety constraints are applied, these models often abstain from answering questions that they could answer under default instruction, yet visual evidence continues to influence the decoding process. The authors show that safety‑induced abstention alters late‑stage hidden‑state dynamics, and that targeted interventions can restore grounded answering without retraining or changing visual inputs.
arXiv:2609.00232v1 Announce Type: new Abstract: Multimodal Large Language Models (MLLMs) have achieved strong performance on OCR-centric document understanding and text-rich visual reasoning benchmar...
arXiv:2608. 10513v1 Announce Type: cross Abstract: Large vision-language models (LVLMs) remain vulnerable to jailbreak attacks that exploit visual inputs to bypass safety alignment inherited from their language backbones.
SafeAtlas-VL introduces a large multimodal safety dataset with 1.5 million instances, rating image, request, and response risks on a five‑level ordinal scale across 15 harm categories and 55 subcategories. The accompanying SafeAtlas‑Bench provides 5,000 held‑out cases for evaluating ordinal predictions and continuous risk scores. Models trained on this data, including an 8B Guard model, achieve state‑of‑the‑art performance, outperforming prior benchmarks by about 4% in F1 score.