CounterVid introduces a scalable counterfactual video generation framework that creates videos differing only in actions or temporal structure while keeping scene context intact. The approach uses multimodal LLMs for action proposals and diffusion models for editing, producing a synthetic dataset of ~26k preference pairs for action recognition and sequence ordering. With the MixDPO optimization method, the authors demonstrate significant improvements in action recognition and temporal ordering on Qwen2.5‑VL and InternVL3 backbones, while maintaining overall video understanding.
By Tobia Poppi, Burak Uzkent, Amanmeet Garg, Lucas Porto, Garin Kessler, Yezhou Yang, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara, Florian Schiffers
The paper introduces Composed Gait Retrieval (CoGR), a task that retrieves a target gait sequence using a reference sequence and a natural language modification query. To support this, the authors create the first gait-language datasets—Language‑Augmented CCPG and Language‑Augmented CASIA‑B—via an automated annotation pipeline powered by large vision‑language models. They propose ComposeGait, an identity‑anchored composition framework with a Part‑aware Identity Adapter that injects identity tokens into a shared Q‑Former, achieving state‑of‑the‑art retrieval performance on both benchmarks.
By Jingchen Fei, Zengbin Wang, Yukun Liu, Muyi Sun, Shibiao Xu, Man Zhang
The paper introduces PuMVR, a benchmark of 1,000 Punjabi image‑text pairs spanning three scripts—Gurmukhi, Shahmukhi, and Roman—to evaluate Vision‑Language Models (VLMs). Testing ten state‑of‑the‑art VLMs reveals a significant Script Gap: models perform well in one script but poorly in another, with accuracy differences up to 16%. The authors propose the Script Consistency Rate (SCR) as a new metric, noting it can be as low as 24.8% on their benchmark, and argue that current multilingual VLMs are not truly multi‑script.
By Prabhjot Singh, Bhushan Pawar, Madhu Reddiboina, Rajvee Sheth
The paper examines how four leading vision‑language models—LLaVA‑7B, Qwen‑VL, GPT‑4o mini, and Claude 3 Haiku—perform in detecting hateful content within memes. It evaluates the models under zero‑shot and few‑shot prompting, focusing not only on classification accuracy but also on the qualitative justifications they generate. The study highlights that these models often overlook contextual nuances, irony, and subtle cues essential for accurately identifying hate speech in memes.
By Muhammad Jawad Chowdhury, Adiba Hasan, Ishrak Hossain, Shahriar Ivan, Sabbir Ahmed
NeuronFuzz is a white‑box fuzzing framework that uses internal safety neurons of large language models as continuous feedback for safety evaluation, eliminating the need to generate full model responses during testing. It constructs a SafetyOracle that converts neuron activations into a differentiable safety alarm score, enabling gradient‑based identification of sensitive template positions and fluent, context‑compatible prompt mutations. Evaluated on 21 text and multimodal models, NeuronFuzz achieves a 76‑100% jailbreak discovery rate on five white‑box source models and demonstrates strong zero‑shot transfer to open‑weight and proprietary targets.
By Zhiyuan Xu, Muhammad Firhard Roslan, Joseph Gardiner, Sana Belguith, Lichao Wu
The paper introduces G2D, a training‑free framework that combines a discriminative model (CLIP) for broad candidate retrieval with a generative vision‑language model for fine‑grained, image‑grounded verification. By using CLIP’s top‑K shortlist and a structured prior from candidate names and probabilities, G2D focuses generative reasoning on uncertain samples, achieving an average accuracy of 68.85% across eight benchmarks—higher than both CLIP alone (59.35%) and the standalone generative model (63.11%). The approach also adapts to various generator configurations and extends to other models such as DCLIP, WaffleCLIP, and CuPL.
By Zehua Hao, Fang Liu, Qinliang Wang, Yaoyang Du, Xinyan Huang, Puhua Chen
AesCanvas is a new dataset and benchmark that evaluates image aesthetic models on two fronts: CritiqueCanvas, which contains 519,136 instruction–response pairs for long‑form, multi‑dimensional critique across photography, painting, and virtual imagery, and ContextCanvas, which offers 301 expert‑reviewed use scenarios to assess contextual aesthetic suitability. The benchmark tests closed‑source, open‑weight general, and aesthetic‑specific multimodal large language models, revealing that models excel at critique generation but lag in context‑sensitive judgment. The study shows that aesthetic specialization does not reliably transfer to contextual suitability and highlights the need for culturally situated, evidence‑grounded suitability as a distinct objective for aesthetic modeling.
By Xuanwei Hu, Haoyu Dong, Kejun Wu, Tianyi Liu, Jianjun Gao
Glass Surface Detection Grounded in 3D Visual Geometry proposes a new approach that grounds glass surface detection in 3D visual geometry rather than relying solely on 2D appearance cues. The method uses a visual geometry grounded transformer (VGGT) to distill 3D priors and creates glass-aware 3D representations, then applies a multi-task learning framework with a Frequency Self-Attention Module (FSAM) and a Geometry Grounding Block (GeGB) to localize and segment glass surfaces. Experiments show state‑of‑the‑art performance on seven benchmarks, good generalization to video and multi‑modal data, and significant improvements in reconstruction of glass scenes.
By Yiwei Lu, Ke Xu, Tao Yan, Xiaojun Chang, Radu Timofte, Rynson W. H. Lau
The paper introduces a new task called compositional layout understanding, focusing on interpreting complex, multi‑layer document and UI designs. It presents CoDeLayout, a VQA dataset of about 20,000 real‑world layouts annotated with compositional element pairs and design intent. The authors identify semantic drift and structural ambiguity as key challenges for vision‑language models and propose MASON, a post‑training approach that combines multimodal alignment and structural perception to improve performance, achieving 91.66% accuracy with only 30% of the training data.
By Yiyang Huang, Zhaowen Wang, Simon Jenni, Jing Shi, Yitian Zhang, Yizhou Wang, Yun Fu
AffectOmni is a reinforcement‑learning‑trained framework that enhances multimodal large language models for affective reasoning in social and art‑related scenes. It introduces People Focus and Temporal Order rewards to prioritize people‑centric cues and structured reasoning, and uses within‑group comparative scoring for more discriminative rewards. A Thinking Summarizer converts rationales into executable evidence instructions, which are grounded into pixel‑level regions via SAM3, enabling external auditability.
By Yibo Wang, Rui Yang, Jisheng Dang, Bimei Wang, Yitao Wu, Pengfei Cao, Wencan Zhang, Hong Peng, Bin Hu, Tat-Seng Chua
AraMS-28k is the largest publicly released line‑level dataset of genuine historical Arabic manuscripts, containing 14 books, 3,043 pages, and 28,600 annotated text lines (27,971 main‑text and 629 margin). The dataset spans three script traditions—Naskh, Ruq'ah, and Maghrebi—and includes a lithographed printed edition for format diversity. Each line is labeled as main‑text or margin, with margin lines that have a clear attachment point annotated with an insertion anchor to recover the manuscript’s true non‑linear reading order; both fully vocalized and diacritic‑normalized transcriptions are provided, and the data was produced via the RefLAM pipeline combining OCR, clean transcriptions, and human review.
"whyItMatters":"The dataset’s comprehensive line‑level annotations, including reading‑order anchors and dual transcription formats, enable reproducible research on Arabic manuscript recognition, layout analysis, and reading‑order recovery under a CC BY‑NC‑SA 4.0 license."
By Mohamed Guechaoui, Mohamed Diaa Zellagui, Souleyman Chaib, Sahraoui Dhelim
The paper introduces Dependency-Aware Revocable Decoding (DARD), a training‑free framework for diffusion large language models that separates tokens into masked, candidate, and unmasked states. DARD verifies candidate tokens using a selective context that excludes less reliable tokens and adaptively regulates their influence on subsequent decoding. Experiments on 12 textual and multimodal benchmarks across three open‑source dLLMs show that DARD improves the speed‑quality Pareto frontier, achieving a 2.71× speedup and a 4.35‑point CIDEr gain over Saber on Flickr30K.
By Wooje Park, Insu Lee, Minyoung Noh, Jaeyun Jang, Sungmin Lee, Kyuhong Shim, Byonghyo Shim
The paper introduces Inverted Asymmetric Fusion (IAF) to address strong-modality collapse in multimodal learning, where dominant modalities are degraded during fusion. IAF preserves the dominant modality by passing it unchanged and letting weaker modalities attend to it, while also strengthening weaker modalities via Modality-Aware Knowledge Distillation. Experiments on MultiHuSE, UR-FUNNY, and MUStARD show that IAF maintains unimodal performance and improves over the best unimodal baseline by up to 8.25%.
By Mary Ogbuka Kenneth, Foaad Khosmood, Abbas Edalat
The paper introduces LAC (Light Actor, deep Critic), an offline reinforcement learning approach that allocates model capacity to a deep critic rather than a complex actor to improve inference efficiency. It addresses three failure modes—optimization, bootstrap-noise amplification, and value-range drift—using a residual MLP backbone, n‑step bootstrap targets, and a categorical cross‑entropy loss. Experiments on OGBench show LAC matches state‑of‑the‑art diffusion and flow‑matching baselines while reducing inference latency by up to four times.
By Guhyeon Kang, Jaehwi Lee, Minhae Kwon
The paper introduces a zero‑shot video restoration and enhancement framework that leverages a text‑to‑image latent diffusion model along with multi‑modal references. It employs dual prompt tuning inversion and sampling to cut inference time to about one‑third of the original, while also strengthening performance and temporal consistency. Additional techniques such as texture‑aware video token merging, referenced self‑attention, and referenced token merging further improve temporal coherence across frames.
By Cong Cao, Huanjing Yue, Xin Liu, Jingyu Yang
Case2Flow is a new task that retrieves the most relevant guideline flowchart for a given patient case from a collection of medical guideline documents. The authors created FlowAtlas, a curated corpus of 202 flowcharts extracted from 2,080 guidelines, and a pipeline that generates 1,911 aligned case‑flowchart pairs. Their evaluation shows that existing multimodal retrieval methods often overrely on keywords and spurious token‑patch matches, and they propose CRISP, a training‑free scoring method that improves Recall@1 by up to 18.71 percentage points and gains preliminary feasibility evidence from a blinded physician assessment.
By Jiale Wei, Yufan Chen, Alexander Jaus, Zdravko Marinov, Julian Friedrich, Simon Rei{\ss}, Jens Kleesiek, Rainer Stiefelhagen
The paper introduces Echo-GRPO, a method that rewrites privileged reasoning traces into a model’s own idiolect to align off‑policy supervision with the student policy’s vocabulary. By preserving semantics through Dual‑Reference Decoding, Echo‑GRPO mitigates gradient clipping on critical reasoning tokens and improves reasoning distillation. The approach is instantiated as VideoEcho‑R1 for video reasoning, yielding consistent gains across multiple multimodal LLM backbones and benchmarks, and it can be applied as a plug‑in to both RL and supervised fine‑tuning frameworks.
By Ji Soo Lee, Jinyoung Park, Seohyun Lee, Jongha Kim, Joonmyung Choi, Jinsung Yoon, Hyunwoo J. Kim
The paper introduces Checkerboard, a clean‑label backdoor attack that uses a closed‑form, data‑independent trigger design based on an input‑space Fisher‑separability objective and a ridge four‑neighbor local‑smoothness prior. This approach yields a pixel‑wise checkerboard trigger without requiring data access, surrogate model training, or iterative optimization, and it outperforms existing norm‑bounded clean‑label attacks across four benchmark datasets. On CIFAR‑10, poisoning 20 samples with a 10/255 perturbation achieves a 95.72% attack success rate, while on IN‑100 a 0.4% global poisoning rate yields over 83% ASR without harming clean accuracy, and the attack remains robust against state‑of‑the‑art defenses.
By Yi Yang, Jinyang Huang, Binbin Liu, Feng-Qi Cui, Xiaokang Zhou, Haiming Jin, Zhi Liu, Jie Zhang, Meng Li
The paper introduces PACE, a factor-guided, progressive framework for acquiring evidence in long-video question answering. PACE first indexes clip-level descriptions using question-derived factors, then refines evidence retrieval with contrastive cues derived from candidate answers. On the MMR‑V dataset, PACE achieves 42.6% accuracy and recovers 66.9% of annotated cues, outperforming direct inference and prior agentic baselines, and shows consistent improvements across several long-video benchmarks.
By Baixuan Xu, Yinyui Xu, Tianshi Zheng, Zhaowei Wang, Weiqi Wang, Haochen Shi, Jiayu Liu, Qing Zong, Xiyu Ren, Xinyu Geng, Zhitao He, Yangqiu Song
FireRedAudio is a 9‑billion‑parameter audio language model that separates continuous input representations for audio understanding and speech generation, enabling a single autoregressive LLM to perform tasks such as ASR, zero‑shot TTS, Instruct TTS, and semantic/acoustic speech editing. The model uses a dedicated Audio Encoder for recognition and a RedAE‑based pathway for generation, with the LLM directly generating text or conditioning a flow‑matching DiT to produce acoustic latents. Evaluations show competitive or leading performance in multilingual ASR, content‑accurate zero‑shot TTS, strong instruction following, and significant improvements in speech editing over prior work.
By Feiyu Shen, Fenglong Xie, Junjie Li, Kun Xie, Lei Xie, Xu Tang, Xuelong Geng, Yan Jia, Yao Hu, Yichen Han, Yichen Wu, Ziqi Dai, Junjie Chen, Kai Huang, Manzhen Wei, Yixuan Li