KhatianDoc is a new benchmark that tests multimodal large language models on Bengali legal land records, specifically the handwritten RS Khatians used in Bangladesh. The benchmark comprises four tasks—symbol recognition, base‑16 to decimal conversion, structured field extraction, and legal document question answering—drawn from 107 real records and 1,634 QA pairs. Six multimodal LLMs were evaluated under a zero‑shot protocol, revealing that many models fail to answer a significant portion of questions correctly and perform poorly on arithmetic conversion, highlighting a lack of capability rather than a performance gap.
By Tasmiad Hasan, Arafat Zaman Ratul, Sarker Sadman Saalim, S. M. Shah Nawaz Hossain, Khan Raiyan Ibne Reza, Sumaiya Tabassum Nimi
VisCAD is a foundation model suite that tackles AI-assisted computer-aided design for industrial products, covering both part-level and assembly-level generation. Its core component, VisCAD‑M1, is a 27B model trained for part-level design generation and outperforms existing models on PubCADBench and RealCADBench, achieving a part-level score of 0.5540 and reaching 0.5797 when used as a test-time verifier. VisCAD also offers a domain-specific harness that improves complex assembly generation compared to general-purpose harnesses, showing quantitative and qualitative advantages.
By JoyIndustrial VisCAD Team, Linxin Cai, Qiuhe Hong, Zhichao Huang, Guanlin Li, Hongsen Liu, Ziqi Liu, Yichen Long, Luya Wang, Yuchen Wang, Wenxiang Wu, Huimu Yu, Ning Zhang
arXiv:2609. 03378v1 Announce Type: new Abstract: Depth can resolve appearance ambiguity in RGB-D salient object detection (SOD), yet sensor depth is not uniformly reliable.
By Xuehao Wang, Jiaxin Hua, Runmei Li, Zhenyu Wu, Chenglizhao Chen, Ke Gu, Aimin Hao
SafeRI proposes an on-demand safety alignment approach for large vision-language models, contrasting with existing always-on methods that globally modify model behavior. The framework uses a lightweight recognizer to evaluate token-level safety during autoregressive generation, gating a LoRA module that only activates when unsafe content is detected. By training the LoRA on unsafe prefixes and safe continuations, SafeRI redirects unsafe generations back to safe responses without perturbing the model’s original reasoning path.
By Caoyuan Ma, Tian Gu, Wenpu Liu, Weichu Xie, Shuai Dong, Yuqi Xu, Ji Zhao, Ziyue Wang, Wenzheng Chang, Taiqiang Wu, Yongfu Zhu, Wenqi Shao, Zheng Wang, Yinqiang Zheng
Principia is a new benchmark that tests video models on Newtonian physics by evaluating relational consistency between paired objects, independent of calibration. It covers eight phenomena—gravity, restitution, friction, rotational inertia, projectile motion, momentum, pendulum, and mass‑spring oscillation—across translational, rotational, collisional, and oscillatory dynamics using real‑world scenes. The benchmark introduces a calibration‑independent consistency score and shows that current state‑of‑the‑art video generators perform poorly on it, with the best vision‑language model achieving only 67% accuracy on detecting physics violations.
By Varun Varma Thozhiyoor, Shivam Tripathi, Venkatesh Babu Radhakrishnan, Anand Bhattad
VKnowU is a benchmark that tests multimodal large language models (MLLMs) on their grasp of visual knowledge—intuitive, human-like understanding of physical and social principles in videos. The benchmark contains 1,680 questions across 1,249 videos, covering eight core types of visual knowledge, and shows that current state‑of‑the‑art MLLMs still lag behind human performance, especially on world‑centric tasks. To address this gap, the authors release VKnowQA and VideoKnow+, a baseline model that incorporates visual knowledge via a See‑Think‑Answer framework and reinforcement learning, improving performance on VKnowU and related datasets.
By Tianxiang Jiang, Sheng Xia, Yicheng Xu, Linquan Wu, Xiangyu Zeng, Limin Wang, Yu Qiao, Yi Wang
The paper investigates audio‑video diffusion models by examining the "attention triangle"—the cross‑attention links among text, audio, and video. It finds that the audio‑video edge is bidirectional and heavily influenced by model biases, leading to semantic leakage when prompts conflict with learned priors. The authors develop attention‑derived diagnostics and inference‑time interventions that improve semantic grounding without sacrificing generation quality.
By Sagi Polaczek, Noa Kraicer, Gal Metzer, Zhuo Ning, Ali Mahdavi-Amiri, Daniel Cohen-Or, Raja Giryes
Headroom-Drift Replay is a replay control primitive designed for GRPO that separates reuse into two decisions: Headroom ranks stored groups by remaining learning value, while Drift gates them by compatibility with the current policy. The method keeps the fresh on‑policy stream unchanged and adds no auxiliary generation or training machinery. Across mathematical reasoning, multimodal reasoning, and Agentic Search benchmarks, it outperforms naive replay and matches or exceeds broader replay methods on Avg Mean@32, delivering comparable quality at materially lower wall‑clock time in Agentic Search.
By Hyun Bin Park, Du-Seong Chang
The paper introduces Seeing Before Synthesizing (SBS), a weakly-supervised dense video captioning framework that uses a vision‑language model to generate frame‑level narratives for gaps between events and detect transitions based on semantic changes. SBS refines temporal masks by aligning transition points with vision‑language cues, rather than relying on rigidly placed synthetic captions. Experiments on ActivityNet Captions and YouCook2 show that SBS achieves state‑of‑the‑art results in both captioning and localization tasks.
By Ye-Chan Kim, Seunghee Choi, SeungJu Cha, Si-Woo Kim, Hwiseon Kim, Hyungee Kim, Dong-Jin Kim
The paper introduces a new Markov chain Monte Carlo method that samples from multimodal distributions by interpolating along the diffusion path of a noising diffusion process, preserving mode weights and improving mixing. It proposes a Metropolis-adjusted diffusion path (MAD-Path) sampler that corrects for bias from approximate score estimates and discretization errors, ensuring the target distribution remains invariant. Experiments on Bayesian posteriors demonstrate that MAD-Path outperforms tempering-based MCMC and unadjusted diffusion samplers in global exploration and accurate mode-weight estimation.
By Han Chen, Sifan Liu, Jun Yang
MetaStructAtlas is a large-scale dataset for whole-body PET/CT interpretation that combines multimodal imaging with integrated anatomical, metabolic, and semantic annotations. It includes 490 co-registered 3D PET and CT volumes, 50,470 organ-level segmentation masks, and grounded radiology reports. The accompanying MetaStructVQA benchmark offers 100,565 QA pairs that link diagnostic queries to visual evidence across modalities, enabling interactive 3D grounded visual question-answering.
By Chenguang Zheng, Le Xue, Yichi Zhang, Wenbo Zhang, Zehui Ling, Gang Feng, Xin Gao, Yuan Qi, Yuan Cheng, Zixin Hu, Mei Tian
Air-Ground Collaborative Vision-and-Language Navigation (AGC‑VLN) pairs a UAV with a global bird’s‑eye view and a UGV with a local first‑person view, creating a shared bird’s‑eye map that displays the UGV’s pose and the target location. The UAV localizes the target in its downward view and flies toward it, while the UGV uses the shared map to plan a road‑following path with a frozen VLM and execute it under closed‑loop control. In CARLA‑Air’s Town10HD scene, AGC‑VLN achieves a 77.0% joint success rate, a 27.0% improvement over the weaker individual agent and outperforms the strongest single‑agent baseline by 24.0 points.
By Shuning Zhang, Liang Li, Yunheng Wang, Tao Wang, Yihang Kang, Renjing Xu
SVG-Score introduces a human‑aligned evaluation framework for text‑to‑SVG generation, addressing the shortcomings of existing image‑based metrics like CLIPScore that poorly capture SVG‑specific errors such as color, count, and spatial inaccuracies. The authors first demonstrate that CLIP‑based scores are largely insensitive to these errors and that generic Vision‑Language Models respond inconsistently across error types and styles. They then present a human‑annotated Semantic Alignment dataset and develop two complementary evaluators: a CLIP‑based scorer adapted to vector graphics and a VLM judge refined through supervised fine‑tuning and reward‑shaped reinforcement learning, enabling both fast large‑scale and expressive, interpretable assessment of SVG generators.
By Marco Cipriano, Leonardo Zini, Alexandra Schild, Valentin Teutschbein, Afsana Mimi, Marcella Cornia, Lorenzo Baraldi, Gerard de Melo
The paper introduces a two‑stage speech anonymization framework that preserves both linguistic content and acoustic identity. It replaces personally identifiable information using a generative editing model and applies a flow‑matching anonymization technique (F3‑VA) to create diverse, distinct anonymized speakers. The authors evaluate privacy with speaker verification metrics and utility by training ASR, TTS, and SER models from scratch, showing stronger privacy protection with minimal utility loss compared to existing baselines.
By Yunchong Xiao, Yuxiang Zhao, Ziyang Ma, Shuai Wang, Kai Yu, Jiachun Liao, Xie Chen
StrixAE is an intelligent audio enhancement agent that uses a multimodal large language model to coordinate multiple enhancement and personalization models. It is trained in two stages: first, chain‑of‑thought supervised fine‑tuning on AcoustBench, and second, Audio Perception Reinforcement Learning that optimizes format validity, structural coherence, and perceptual quality. The approach yields state‑of‑the‑art performance and strong generalization on real‑world test datasets, outperforming many existing open‑source and proprietary solutions.
By Chenglin Wu, Junjie Wu, Jinhang Chen, Mingyang Chen, Zixu Lin, Jiabian Chen, Xinghao Ding, Xiaotong Tu
LLaDA-Image is a unified framework that couples a 6B Diffusion Transformer (DiT) trained from scratch with a frozen vision‑language module based on the LLaDA2.0‑Mini diffusion language model. The approach first builds a strong visual generative prior through image‑only pre‑training and mid‑training, then fine‑tunes with a 220M‑sample generation pipeline that includes 98 real images. The resulting model produces highly photorealistic images that accurately follow fine‑grained editing instructions, and a distilled version, LLaDA‑Image‑Turbo, enables fast inference in 2–4 sampling steps. On Qwen‑Image‑Bench, LLaDA‑Image sets new state‑of‑the‑art scores for open‑source models in both English and Chinese tracks, and the authors release weights, code, and detailed recipes to support further research.
By Chuyan Chen, Haoxing Chen, Kun Chen, Zhenglin Cheng, Long Cui, Ruishan Fang, Zhangxuan Gu, Zhicheng Huang, Zhenzhong Lan, Yuanting Lei, Haoquan Li, Jianguo Li, Rongchuan Li, Sidu Li, Tao Lin, Deyuan Liu, Jiacheng Liu, Lin Liu, Yuxuan Lou, Zhisheng Lu, Yuxin Ma, Shuheng Shen, Peng Sun, Chaoyang Wang, Hongjun Wang, Xiaomei Wang, Yongxin Wang, Chengzhang Wu, Hongru Wu, Jun Xie
FWBC‑VLA is a force‑aware framework that links vision‑language‑action (VLA) models with whole‑body compensation control for wheeled‑legged robots. It introduces HSR‑Force, a sensorless residual‑torque estimator that infers contact strength and encodes this information as tokens for the VLA action decoder, allowing the policy to detect contact onset, loading, and release. The system fine‑tunes a pretrained VLA backbone on a large WL&Arm dataset, combines proprioceptive, Jacobian‑derived force, and contact estimates to generate corrective actions, and demonstrates effectiveness in real‑world tasks such as whiteboard wiping and door opening.
By Yutian Zhang, Siyuan Ma, Liwen Yang, Yang Li, Ce Hao, Haozhen Chi, Dong We, Qiaojun Yu, Dibo Hou
SimpleDesign is a single-stage, end-to-end model for joint protein sequence and structure design that eliminates the need for multi-stage training. It combines discrete cross-entropy for sequences with a regression objective for structures, using a Mixture-of-Transformer architecture to handle modality-specific processing while maintaining global self-attention. Trained on over 2 million sequence-structure pairs, SimpleDesign achieves strong performance on co-design and unconditional generation benchmarks.
By Jiarui Lu, Yuyang Wang, Yizhe Zhang, Jiatao Gu, Navdeep Jaitly, Joshua M. Susskind, Miguel \'Angel Bautista
The paper introduces a landmark-based approach for classifying injury-associated athlete sessions using minute-resolution multimodal football monitoring data. Instead of labeling every minute, it creates a single representation per athlete-session at fixed time points (landmarks) that incorporates information up to that moment, thereby aligning the session-level injury label with the data. Experiments on 3,743 elite women's football athlete-sessions from the 2020 SoccerMon dataset evaluate various representation strategies and machine learning models, reporting ROC-AUC values between 0.367 and 0.607 and PR-AUC values between 0.0080 and 0.0150, with notable uncertainty across landmarks.
By Evangelos Chatzidimitriou, Konstantinos Tserpes
FPCO-Dialog is a new benchmark designed to evaluate how vision‑language models correct and cooperate when faced with repeated false premises in multi‑turn dialogues. The dataset contains 1,080 images and 10,800 question turns, organized by visual complexity, object category, and false‑premise class, and follows a 10‑turn protocol where a correct dialogue prefix is followed by repeated false‑premise expressions. Using a model‑agnostic protocol and the CorrTP@K correction‑rate metric, the benchmark reveals significant differences among 20 commercial and open‑source VLMs in their correction tendencies, turn‑wise dynamics, and responses to different false‑premise types.
By Jiayuan Ma, Yuqi Lu, Weiyang Guo, Chenrui Wang, Junyi Shu, Xuebo Liu, Min Zhang, Jing Li