DocAttriBench (DAB) is a large‑scale benchmark for fine‑grained, element‑level source attribution in Document Visual Question Answering (VQA). It introduces MAPPET, a Mask‑based Perplexity‑Derived Attribution method that uses document layout and language modeling to identify the most informative layout element for each answer. The benchmark contains 237k documents and 296k question‑answer pairs with element‑level grounding, and it evaluates multimodal LLMs on answer accuracy, attribution accuracy, and overall answer quality, revealing that even strong models often fail to localize supporting elements.
By Luca De Grandis (University of Modena and Reggio Emilia, Modena, Italy), Silvia Cappelletti (University of Modena and Reggio Emilia, Modena, Italy), William Raccagni (University of Modena and Reggio Emilia, Modena, Italy, University of Pisa, Pisa, Italy), Marcella Cornia (University of Modena and Reggio Emilia, Modena, Italy), Lorenzo Baraldi (University of Modena and Reggio Emilia, Modena, Italy), Rita Cucchiara (University of Modena and Reggio Emilia, Modena, Italy)
The paper introduces MiX, a micro‑inverted‑scaling format that replaces shared exponents with shared mantissas to avoid microscaling collapse in low‑bit vision‑language models. An adaptive dual‑format inference framework (MiX‑MX) maps this format to a custom accelerator, replacing multipliers with shifters. Experiments show 4.5‑bit MiX matches or outperforms NVFP4 accuracy while improving area efficiency by 25 % and delivering 2.3–4.5× speedup with 1.4–2.9× energy savings over the Focus accelerator.
By Yuan Liao, Jae-sun Seo
Navi-Agent is a zero‑shot Vision‑Language Navigation in Continuous Environments (VLN‑CE) agent that builds a coordinate‑free spatial state from visual observations and motion history. It represents this state as a navigation topology with nodes as visual places and edges as motion transitions, enabling observation‑based self‑localization, task progress verification, and recovery. Experiments on a zero‑shot VLN‑CE benchmark and real‑world robot platforms demonstrate that Navi‑Agent achieves state‑of‑the‑art performance among geometry‑constrained methods while remaining competitive with geometry‑based approaches.
By Wenyuan Xie, Mengyang Hong, Yongzhong Wang, Yanbiao Ji, Yijin Zhou, Shaokai Wu, Shalayiding Sirejiding, Huayi Zhou, Yi-Chao Chen, Ma Ling, Yue Ding, Hongtao Lu
The paper introduces Fast Preemptive Robustification (FPR), a lightweight defense that enhances the robustness of deep neural networks against transferable adversarial attacks. By sharpening Laplacian responses through a single 3×3 channel‑wise convolution, FPR eliminates the need for surrogate models, iterative optimization, or specialized training. Experiments show that FPR lowers untargeted attack success rates by 12.7% and reduces targeted attack success from 10.7% to 4.1%.
By Jiaming Liang, Chi-Man Pun
SalsaAgent is a multimodal embodied language model that generates expressive, full‑body salsa follower motions in response to a human leader and music. The approach treats partner interaction as nonverbal token passing, extending a large language model’s vocabulary to include discrete motion, pairwise relation, and audio tokens. A two‑stage token‑to‑diffusion pipeline, combined with full‑body and pairwise‑relation tokenizers and alignment with automatically derived text descriptions of skeleton dynamics, yields improved motion quality, spatial coordination, and music‑partner synchrony compared to prior baselines.
By Payam Jome Yazdian, Zoe Stanley, Angelica Lim
The paper introduces AttnSign, a vision‑language model that improves sign language translation by steering spatial‑temporal attention. It first supervises attention on sign‑relevant regions such as faces and hands in each frame, then uses an RL‑based motion‑cadence method to focus on keyframes. Experiments on How2Sign and OpenASL show AttnSign consistently outperforms existing methods.
By Meibo Hu, Guohao Sun, Annemarie D. Ross, Sheng Li, Zhiqiang Tao
HunyuanOCR-1.5 is a lightweight, end‑to‑end OCR‑specialized vision‑language model that unifies document parsing, text spotting, information extraction, text‑image translation, and multi‑image document understanding. It builds on the HunyuanOCR‑1.0 architecture, improving efficiency with DFlash‑based OCR decoding for faster inference (6.37× Transformer speedup, 2.14× under vLLM) and enhancing capability through an Agentic Data Flow system that autonomously constructs high‑quality training data for long‑tail OCR tasks. The model achieves top‑tier performance on OmniDocBench v1.6 and sets new milestones in ancient‑script OCR, chart/table parsing, multilingual parsing, and hallucination evaluation, while remaining lightweight for deployment.
By Gengluo Li, Xingyu Wan, Shangpin Peng, Weinong Wang, Hao Feng, Yongkun Du, Binghong Wu, Zheng Ruan, Zhiqiong Lu, Liang Wu, Pengyuan Lyu, Huawen Shen, Zibin Lin, Shijing Hu, Jieneng Yang, Hongbing Wen, Guanghua Yu, Hong Liu, Bochao Wang, Can Ma, Han Hu, Chengquan Zhang, Yu Zhou
E-AVI is a new framework for automated video interview assessment that combines verbal, acoustic, and visual data. It extracts timestamped multimodal evidence and uses dimension‑conditioned attention with source‑level embeddings to score candidates. The system also provides a shared evidence pool for natural‑language feedback and follow‑up question answering, outperforming multimodal baselines on RecruitView and a private hospitality dataset.
By Haoshen Wang, Dongbo Che, Zeyi Xie, Yuanjie Du, Shicheng Hua, Xingyu Wang
The paper shows that the wording of prompts in vision‑language models (VLMs) can either improve or worsen robustness to image corruption. Verbose prompts broaden the cross‑modal attention’s frequency filter, making the model less sensitive to corruptions, while semantically complex prompts narrow the filter and increase vulnerability. Experiments on Qwen3‑VL and LLaVA‑OneVision confirm that adding padding or verbose phrasing reduces answer drift by 70–81% on 8B models.
By Farooq Ahmad Wani, Maria Sofia Bucarelli, Mujtaba Hussain Mirza, Oleksandr Pryymak, Aryo Pradipta Gema, Iacopo Masi, Pasquale Minervini, Fabrizio Silvestri
The paper introduces Lens, a training‑free framework that aligns multimodal representations with the semantic perspective required by downstream tasks. Lens uses a task‑specific readout phrase to anchor the perspective and then aggregates token states after the full input, ensuring the extracted representation reflects task‑conditioned evidence integration rather than generic salient content. The method achieves a Precision@1 of 63.9 across 36 MMEB datasets, outperforming the nearest training‑free baseline by 10.2 points.
By Xinran Liu, Shouqian Shi, Yixian Chen, Ruizhi Chen, Xin-Wei Yao, Sheng Zhong
FORGE is a forensic deepfake analysis system that provides region‑grounded natural language explanations for image manipulations. It addresses the inductive bias mismatch of multimodal large language models by adding a Vision‑Only Model trained on dense patch prediction, allowing the language model to interleave tokens with preserved spatial correspondence. Across face‑manipulated and fully synthetic content, FORGE delivers fine‑grained attribute queries and outperforms in‑domain baselines, with region‑specific evaluation and human studies confirming explanation faithfulness.
By Rohit Kundu, Shan Jia, Vishal Mohanty, Athula Balachandran, Amit K. Roy-Chowdhury
VisKG‑LM proposes compiling retrieved knowledge graph subgraphs into static visual memories rather than re‑encoding them during each inference step. The method serializes each subgraph as Relation‑Labeled Paths, renders them as images that preserve the graph’s branching structure, and caches these images for reuse. At inference, a language model processes the question and candidate text first, then consults the cached visual memory only at its final layer, yielding improved performance on CommonsenseQA, OpenBookQA, and MedQA‑USMLE compared to both text‑only baselines and a large vision‑language model.
By Yixin Peng, Er Jin, Shiwei Luo, Diego Collarana, Stefan Decker
The paper introduces the Modality Discrepancy Transformer (MDT), a model designed to detect ambivalence and hesitancy in clinical videos by capturing cross‑modal disagreement across facial, vocal, and linguistic signals. MDT expands a 6‑token representation to 9 tokens that include modality embeddings, absolute‑difference features, and Hadamard‑product discrepancy features, which are processed through Transformer self‑attention with FiLM‑based text conditioning and LoRA fine‑tuning. On the BAH dataset from the 3rd ABAW Challenge, MDT achieves a Macro F1 score of 0.7408 on the labelled test split and 0.7368 on the private leaderboard, surpassing the strongest baseline by over 10 points while training in under 20 minutes on a single GPU.
By Shiyu Luo, Yu Wang, Jiawen Huang, Zhaoxiang Xiao, Chenxi Huang, Qi Zhang, Bin Liu
The paper introduces a frontend‑backend architecture for full‑duplex speech‑to‑speech models that enables tool calls while preserving natural conversational flow. The frontend emits a delegation token and streams ASR transcripts to a text‑based backend LLM, which performs tool calls and returns results that are re‑injected into the frontend via a lightweight prefill‑and‑repeat mechanism before streaming TTS synthesis. In single‑turn evaluations the system achieves 92‑97% tool‑call recall, 81.2% accuracy in rejecting irrelevant calls, and competitive performance on Full‑Duplex‑Bench‑V3 and EVA‑Bench when paired with a large backend model.
By Ke Hu, Slyne Deng, Chen Chen, Elena Rastorgueva, Edresson Casanova, Punit Kumar, Dharmendra Choudhary, Nikhil Srihari, Ameya Sunil Mahabaleshwarkar, Viet Anh Trinh, Slim Essid, Oluwatobi Olabiyi, Zhehuai Chen
Efficient Unified Multimodal Understanding (EUMU) is the winning solution for the MUMU Track of the 8th LSVOS Challenge, addressing multi‑concept image tagging, open‑vocabulary object detection, and image captioning with a single efficient model. It leverages a shared pretrained multimodal backbone and lightweight heads, while applying task‑aware inference refinement that uses detection cues to improve captioning, caption cues to recover missed detections, and image statistics to refine tagging. With 239.169 M parameters, 23.947 GFLOPs, and 4.5 GB peak memory, EUMU achieves a challenge score of 17.3409 and is publicly available on GitHub.
By Dayoung Kil, Seong-heum Kim
The paper introduces a new task called visual scene negative captioning, which aims to describe what should be present in an image but is actually absent, a capability crucial for safety-critical applications. It proposes the CRCD framework, which uses counterfactual reconstruction and contrastive decoding to overcome affirmation bias, limited mental filling, and representation bias. CRCD employs a dual-branch architecture for amodal completion and functional association, along with multi-condition representation learning, to generate accurate negative captions and sets a high-performance baseline for this emerging task.
By Zhiyun Jiang, Hanyong Wang, Binbin Liang, Yu Xie, Menglong Yang, Wei Li
QCPruner is a training‑free visual token pruning method that conditions both token selection and representation on the query via bilateral utility weighting. It fuses keyword‑matched query anchors with cross‑modal cues to compute a nonnegative facility‑location objective that is monotone and submodular, guaranteeing a (1‑1/e) greedy approximation. Across multiple multimodal large language models, QCPruner consistently outperforms existing pruning methods, achieving over 96% of unpruned performance even with very few tokens retained.
By Shengli He (Guizhou University), Yongchao Liang (Guizhou University), Roumeng He (Shanghai Ocean University), Junjie Zeng (Guizhou University), Jiyuan He (Guizhou University), Can Wu (Guizhou University), Li Zheng (Guizhou University)
The paper introduces Latent-Centroid Steering (LCS), a single-pass classifier-free guidance method for vision‑language autonomous driving models. LCS replaces instance‑level residuals with class‑level latent shifts, projecting conditional representations toward precomputed command‑specific centroids to enhance command adherence. Experiments on Bench2Drive and nuScenes show that LCS cuts inference latency by about 50% while improving driving performance.
By Meibo Hu, Jiamian Wang, Pichao Wang, Zhiqiang Tao
AUDITPLAN introduces a plan-then-answer method for safety alignment in language models, where the model first generates a compact structured safety plan before responding. The plan includes a threat label, intended action, and explicit constraints, allowing machine‑checkable auditing while remaining hidden from end users. Training combines supervised fine‑tuning with reinforcement learning using the FAITHGATE reward, which only rewards correct plans, thereby reducing unsafe shortcuts and improving robustness across Qwen model variants.
By Sai Sri Pushpa Jampani, Kshitij Mishra, Asif Ekbal
The paper introduces Granite 5.0 TurboCTC, a 470 million‑parameter encoder‑only ASR model that balances speed and accuracy. Its architecture combines pyramidal temporal subsampling, block‑diagonal self‑attention, and intermediate prediction conditioning within Conformer blocks. Training relies solely on public data, employs a Muon optimizer, and uses balanced sampling, while inference is accelerated by replacing 1×1 convolutions with linear layers and optimizing attention calculations.
By Brian Kingsbury, George Saon, Masayuki Suzuki, Hong-Kwang J. Kuo, Takashi Fukuda, Samuel Thomas, Vishal Sunder, Avihu Dekel