The paper introduces Temporally-Grounded Language Generation (TGLG), a benchmark that tests vision‑language models on their ability to produce semantically accurate and temporally precise utterances in real‑time settings. It identifies perceptual updating and contingency awareness as key capabilities, curates datasets from sports broadcasting and egocentric interactions, and proposes the TRACE metric to jointly evaluate semantic similarity and temporal alignment. The authors also present VLM‑TSI, a model that interleaves visual and linguistic tokens in a time‑synchronized manner, achieving better performance than a strong baseline yet still showing modest overall results, underscoring the challenge of real‑time VLMs.
By Keunwoo Peter Yu, Joyce Chai
MM‑Spectrum is a sparse Mixture‑of‑Experts framework designed to infer molecular structures from multimodal spectroscopic data. It introduces a modality‑aware routing mechanism that exposes spectral identity to the router, along with shared and interaction experts of heterogeneous capacities to capture both modality‑unique and cross‑modal synergistic information while reducing noise interference. Experiments across full‑modality, bimodal, and missing‑modality scenarios show consistent and substantial performance gains, supported by ablation studies and interpretability analyses.
By Hai-tao Yu, Nan Min, Zheng Fang, Hongyu Zhan, Yusen Tan, Yuhan Wang, Jun Xia
The paper introduces MDiTFace, a diffusion transformer designed for high‑fidelity mask‑text collaborative facial generation. It unifies tokenization of semantic masks and text, enabling synchronous multimodal feature interaction via stacked multivariate transformer blocks. A novel decoupled attention mechanism separates dynamic and static computations, allowing caching of static features and reducing mask‑condition overhead by over 94% while preserving performance, leading to superior facial fidelity and conditional consistency compared to existing methods.
By Yushe Cao, Dianxi Shi, Xing Fu, Xuechao Zou, Haikuo Peng, Xueqi Li, Chun Yu, Junliang Xing
The paper introduces SubTTA, a test-time adaptation method for vision‑language models that aligns the semantic subspaces of visual and textual modalities to improve zero‑shot predictions. It addresses two issues: the modality gap caused by distribution shifts and visual nuisance that masks task‑specific semantics. By minimizing chordal distance between principal subspaces and projecting visual features onto a task‑specific textual subspace, SubTTA refines decision boundaries and achieves an average 2.24% improvement over existing TTA methods.
By Zhichen Zeng, Wenxuan Bao, Xiao Lin, Ruizhong Qiu, Tianxin Wei, Xuying Ning, Yuchen Yan, Chen Luo, Monica Xiao Cheng, Jingrui He, Hanghang Tong
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
AdaThinking-E introduces a reinforcement learning framework that uses one-token entropy regulation to enable large language models to decide adaptively whether to engage in deep reasoning. By measuring entropy in the predicted probability distribution at key decision tokens, the model learns to explore different thinking strategies during training and converge to confident, efficient decision policies. Experiments show the method improves accuracy on complex tasks while reducing computational overhead on simpler ones across various document reasoning benchmarks.
By Zining Wang, Tongkun Guan, Boming Chen, Zhentao Guo, Jianqiang Liu, Chao Jin, Chen Duan, Kai Zhou, Pengfei Yan, Wei Shen, Xiaokang Yang
Video-FLAIR is a training framework that teaches a model to choose the most suitable reasoning mode—perceptual, compositional, or deliberative—for each multimodal query using reinforcement learning. During training, the model generates responses under all three modes for the same prompt, and a composite reward system selects the best response based on correctness, grounding, and cost, discouraging unsupported deliberation. This adaptive approach improves accuracy on benchmarks such as MathVista, Video-Holmes, and Video-MMMU while dramatically reducing token usage compared to always-thinking baselines.
By Yogesh Kulkarni, Pooyan Fazli
The paper introduces HealthCUES, a real‑time streaming pipeline that extracts and analyzes cough and throat‑clearing events from live spoken conversations. It detects coughs within sub‑second latency, distinguishes cough subtypes (dry, wet, barking, whooping), differentiates coughing from throat clearing, and estimates temporal boundaries, all while gating alerts based on conversational context. The system, built on Qwen3Omni, achieves high accuracy (93% F1 for cough detection) and low latency (340 ms) and has been validated by healthcare professionals for telehealth use.
By Tanmay Laud, Herprit Mahal, Subhabrata Mukherjee
The paper introduces Planning Diffusion Policy Optimization (PDPO), an offline‑to‑online reinforcement‑learning framework that employs a diffusion policy to produce short‑horizon action chunks for robot crowd navigation. PDPO is pretrained on collision‑avoidance demonstrations and fine‑tuned online with PPO, generating five‑step action sequences applied in a receding‑horizon manner. The authors also identify a benchmark artifact where agents can leave the valid domain without explicit boundary constraints, and they mitigate this by treating boundary violations as collisions, leading to improved success rates over strong baselines.
By Wendong Li, Jochen Garcke
CODE: Cross-Modal Calibration and Dynamic Suppression for Open World Object Detection introduces a unified inference-time framework that addresses semantic ambiguity and over-suppression in multimodal OWOD systems. It comprises Cross-Modal Joint Confidence Calibration, Uncertainty-Guided Universal Objectness Enhancement, and Dynamic Outlier Suppression via Confidence Margin. Experiments on the Real-World Detection benchmark with the OWL‑ViT L/14 backbone show CODE achieving 21.7 U‑mAP and 40.8 K‑mAP, surpassing prior state‑of‑the‑art results by 2.6 and 2.3 points respectively.
By Hao Xu, Zhaoning Shi, Hehe Jin, Bo Ma
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 SARA, a framework that separates action induction from runtime authorization in tool‑augmented LLM agents. By treating these as distinct roles, SARA uses an Action Probe to record action provenance and only authorizes tool calls that align with the user objective and past successful executions. Experiments on AgentDojo and AgentDyn show that SARA reduces action‑to‑side‑effect risk to below 0.63% while preserving task performance.
By Xiaokun Guo, Zhen Xu, Dongdong Huo, Yanqiu Zhang, Wei Wang, Qinfu Yang, Dongjin Yu, Yu Wang
DocTalkBN is a large-scale multimodal dataset of authentic expert telemedicine conversations in Bengali, comprising 557.63 hours of paired audio and text, 1,515 multi-turn patient calls, and 10,274 host–doctor question–answer exchanges across 26 medical specialties. The dataset contains 1.7 million tokens and preserves the spontaneity and contextual richness of real medical interactions in a low-resource language. Three downstream tasks—medical triage classification, advice safety evaluation, and medical named entity recognition—are constructed to benchmark large language models and encoder-based baselines, demonstrating DocTalkBN’s practical usefulness for clinically grounded reasoning.
By Anik Saha, Fahmida Sultana Naznin, Sadatul Islam Sadi, Ananya Shahrin Promi, Wahid Al Azad Navid, Rifat Shahriyar
The paper introduces PAYN, a training‑free token compression strategy for Multimodal Large Language Model (MLLM) based Referring Expression Segmentation (RES). By preserving original position embeddings and local spatial structures, PAYN retains tokens that are evenly distributed across neighboring regions, thereby maintaining spatial relational consistency. Experiments on multiple RES benchmarks show that PAYN outperforms existing token compression methods, confirming that position information alone is sufficient for effective compression in this task.
By Yuhan Liu, Yixiong Zou, Yuhua Li, Ruixuan Li
The paper introduces GSM8K‑V, a new benchmark that converts the GSM8K grade‑school math word problems into multi‑image sequences while preserving semantic equivalence. It requires vision‑language models to extract quantities through visual perception and reconstruct reasoning chains by integrating implicit cues across scenes. Evaluation of 34 VLMs shows a large modality gap: most models score above 90 % on text, but the best reaches only 59 % on GSM8K‑V, highlighting a distinct challenge in implicit visual inference.
By Fan Yuan, Yuchen Yan, Yifan Jiang, Haoran Zhao, Tao Feng, Jinyan Chen, Yanwei Lou, Wenqi Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen
The article discusses how vision generative AI models, while rapidly advancing, have largely been developed with a focus on output quality, leading to hardware that adapts reactively to increasing model demands. It evaluates the parameter cost and energy efficiency of these models across various accelerator platforms and aligns four generative model families with seven real-world application domains. The authors propose a software‑hardware co‑design strategy that considers deployment constraints from the outset, ensuring that the appropriate model runs on suitable hardware for specific applications, thereby making generative AI deployment more sustainable and widely accessible.
By Eleni Tselepi, Cristian Sestito, Shady Agwa, Themis Prodromakis
The paper presents a method for accurately measuring 3D and 2D circular centers, a key geometric sensing problem for instruments like cameras and LiDARs. It integrates a conformal‑geometric‑algebra estimator with RANSAC to jointly recover the 3D center, normal, and radius from noisy LiDAR points, and uses a chord‑length‑variance criterion to estimate the 2D projected center, resolving ambiguity with homography validation or a quasi‑RANSAC fallback. Experiments on synthetic and real sensors demonstrate improved center measurement accuracy and reduced extrinsic calibration error across various target and sensor setups.
By Jiajun Jiang, Xiao Hu, Wancheng Liu, Wei Jiang
PACE introduces a training‑free Condense‑and‑Extract framework that speeds up Vision‑Language Model inference by first adaptively downsampling visual inputs before encoding and then selectively retaining essential tokens during decoding. The Adaptive Pixel Compressor (APC) reduces encoder workload while preserving global context, and the Dynamic Dual‑Attention Extractor (DDAE) keeps task‑critical details by fusing visual and language signals. Applied to Qwen2.5‑VL‑7B, PACE maintains 93.8% of performance using only 10% of visual tokens, achieving a 3.1× speedup in time to first token.
By Junjie Liu, Shengyuan Ye, Xu Chen
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
HUG‑VIS is a unified multimodal benchmark for human‑centered visual intelligence, comprising 8,400 half‑body videos of 30 professional actors performing 280 emotion‑action prompts in Mandarin. The dataset provides synchronized video, audio, text, and alpha mattes for four tasks—human emotion recognition, video generation, voice cloning, and video matting—allowing evaluation of both open‑ and closed‑source models under a zero‑shot protocol. Results reveal that linguistic cues dominate emotion recognition, visual affect is weakest, and that automatic metrics and human judgments diverge in generation and cloning tasks, while motion‑related boundary fidelity remains a key challenge for matting.
By Fei Ma, Zebang Cheng, Minghui Li, Hongbo Xu, Yuyong Tan, Yihua Shao, Hanling Wang, Zhou Liu, Yuqing Gao, Dong Wang, Long Ma, Laizhong Cui, Nicu Sebe, Qi Tian