The paper introduces Llama-Mobile, a framework that quantizes vision‑language models for efficient mobile deployment. It uses a quantization pipeline that generates training data from the model itself, eliminating the need for the original training setup, and employs a novel 2.7‑bit‑per‑parameter format optimized for Arm CPUs. Applying this method, the authors compress the Llama 3.2 11B Vision Instruct model to 3.7 GB with 8‑bit activations while maintaining strong performance on visual question answering tasks.
By Luka Ribar, Jeevan Bhoot, Douglas Orr
arXiv:2608.21030v1 Announce Type: cross
Abstract: Video multimodal large language models have advanced significantly, yet fine-grained motion-temporal understanding remains fragile. The core bottlene...
By Chenghua Zhu, Zhaolu Kang, Qifan Shi, Siyan Wu, Kehan Jiang, Lei Wei, Lianyu Hu, Guangyuan Dong, Mingbo Yang, Rui Lu, Guibo Luo
The study examines how the realism of synthetic room impulse response (RIR) datasets influences the training of DeepFilterNet3 for single‑channel speech enhancement. By comparing a DNS4 image‑source‑method RIR set with a higher‑fidelity hybrid wave‑based and geometrical acoustics RIR set, the authors find that the more realistic dataset consistently improves objective speech enhancement metrics and significantly reduces ASR word error rates on unseen measured RIRs. The results suggest that overall realism in synthetic acoustic training data enhances DeepFilterNet3’s generalization to new environments.
By Alessia Milo, Georg G\"otz, Steinar Gu{\dh}j\'onsson, Daniel Gert Nielsen, Jesper Pedersen, Finnur Pind
CogCanvas is a new benchmark for multi-subject reference-based image generation, featuring 1,952 curated reference images of 100 celebrities, 115 objects/fashion items, and 29 real-world backgrounds. It generates 1,361 compositional prompts with 2–5 people, using a pipeline that includes DINOv2 deduplication, aesthetic filtering, and automated graph derivation for interaction and positioning. The benchmark evaluates three tasks—reference-based multi-human-object generation, text-to-image compositional generation, and reference retrieval—under a six-axis protocol, and introduces BG‑Sim and Attr‑VQA metrics to assess background fidelity and attribute binding.
By Long-Bao Nguyen, Quang-Khai Le, Tam V. Nguyen, Minh-Triet Tran, Trung-Nghia Le
The paper introduces the Generative Embedding Benchmark (GEB), which evaluates how much content from an embedding can be recovered by a decoder that only has access to the frozen embedding and a question, without the original image or intermediate features. GEB uses a curated visual‑question‑answering dataset with 1,800 development and 900 test items covering natural images, scene text, and visual documents. Experiments on seven public embedding models show that visual‑only scores range from 28.25 to 33.21, while joint image‑question encoding boosts scores up to 65.56, revealing that generative readout uncovers information bottlenecks not captured by traditional separability‑based benchmarks.
By Yun Li, Biao Yang, Peixi Wu, Yunhao Zhou, Mingzhou Jiang, Wei Yuan, Fan Yang, Wenwu Ou
ClawSentry is an open‑source, framework‑agnostic security supervision gateway designed to protect autonomous large language model (LLM) agents from progressive risks that can arise at four points in the agent control loop: skill admission, invocation‑time intent, execution‑time effect, and post‑action consequence. It introduces a multi‑tier decision engine—deterministic L1, rule‑anchored L2, and read‑only L3—alongside a First‑Use Skill Package Review (FSPR) and an Agent Harness Protocol (AHP) that applies a single policy across multiple agent runtimes without modifying their internals. Evaluation on SkillInject and the SkillsSafety benchmark shows that ClawSentry significantly reduces contextual adversarial skill risk (ASR) while maintaining high task success rates (TSR).
By Kai Wang, Zeming Wei, BiaoJie Zeng, Chang Jin, An Wang, Xiaokun Luan, Zhixiao Lin, Jingjing Qu, Xia Hu, Xingcheng Xu
ForeTime‑VLA is a causal vision‑language‑action policy that distills future‑aware representations from a frozen Fast‑WAM teacher, enabling it to anticipate contact events during conveyor‑belt manipulation. The method compresses current and future video latents into a 64‑dimensional target, uses an eight‑frame history encoder to predict this target along with manipulation phase and time‑to‑transition, and conditions a VLM prefix on future tokens and phase. On a deduplicated conveyor‑belt dataset, ForeTime‑VLA reduces test MAE by 2.63% and L2 by 3.02%, while real‑robot experiments show significantly higher grasp success rates compared to the next‑best reference.
whyItMatters":"The approach demonstrates that distilling future‑token knowledge from a world‑action model can improve dynamic manipulation performance without the computational cost of running the teacher at inference time."
By Siyuan Ma, Yutian Zhang, Boshi Zhang, Qinglian Wu, Jiaqi Zhai, Dong Wei, Xiaojin Huang
Poly-InstructTTS is a text‑to‑speech system that learns expressive speech from open‑ended natural‑language instructions using a 1,000‑hour, 1,000‑plus emotion and style annotated audiovisual corpus. The approach employs a prompt‑free GPT with attribute‑based thinking tokens and a flow‑matching module to inject timbre from reference audio, and includes a speaker fine‑tuning procedure to transfer instruction control while preserving speaker persona. Experiments demonstrate strong instruction adherence and expressiveness, with audio demos and an expanded test set available on the project page.
By Junhui Zhang, Qianhui Xu, Qingxiang Guo, Dawei Yang, Ling Miao, Qiangqiang Wang, Yang Song
arXiv:2608.20749v1 Announce Type: new
Abstract: Identity-preserving video generation aims to synthesize videos that follow natural-language instructions while maintaining the visual identity of a giv...
By Jiayi Gao, Changcheng Hua, Jiaqi Tang, Yuxin Peng, Yang Liu
Re$^3$Cap introduces a retrieval‑guided refinement strategy for image captioning that leverages multi‑modal retrieval as a reasoning signal. The method, built on a Caption Refinement Suggester and a Caption Quality Assessor, detects hallucinations and omissions to produce more accurate and detailed captions without extra annotations. Experiments show it surpasses supervised fine‑tuning and improves relation reasoning by 8.64% on the COCO‑LN500 benchmark.
By Haonan Jia, Shichao Dong, Zenghui Sun, Jiawen Zheng, Ziqi Miao, Gege Shi, Qiuyu Zhao, Jinsong Lan, Xiaoyong Zhu, Bo Zheng
arXiv:2608.20868v1 Announce Type: cross
Abstract: Document processing pipelines traditionally cascade optical character recognition (OCR) engines with downstream models for structured information ext...
By A. Said Gurbuz (IBM Research Zurich, ETH Zurich), Ahmed Nassar (IBM Research Zurich), Christoph Auer (IBM Research Zurich), Maksym Lysak (IBM Research Zurich), Lucas Morin (IBM Research Zurich), Matteo Omenetti (IBM Research Zurich), Tim Strohmeyer (IBM Research Zurich), Panagiotis Vagenas (IBM Research Zurich), Nikolaos Livathinos (IBM Research Zurich), Michele Dolfi (IBM Research Zurich), Peter Staar (IBM Research Zurich)
The paper introduces CARD, a diagnostic tool that steers activations along one mental-state axis while measuring predictions on another, to investigate how vision‑language models use internal representations of beliefs, knowledge, and intentions. Using CARD on open‑weight VLMs evaluated on the Relay Chain cooperative grid‑world benchmark, the authors uncover a routing failure: models do not incorporate belief representations into their next action predictions, leaving partner information unused.
By Souptik Kumar Majumdar, Fabian K\"ogel, Andreas Bulling
arXiv:2608.21133v1 Announce Type: new
Abstract: Medical image-text data can expose protected health information (PHI) through both visible image content as well as accompanying text, creating a barri...
By Shiva Shrestha, Zongxing Xie, Chen Zhao, Liran Ma, Zhipeng Cai, Honghui Xu
The paper introduces Linear Discriminant Tree Ensembles (LDT, LDF, LDAB) for multimodal affect and behaviour classification, combining text, audio, and visual streams. It uses tokenization, concept clustering, and tree-based routing to balance accuracy and interpretability, and proposes a modified feature importance metric that reduces negative class bias. The ensembles outperform Multimodal Transformers and Interpretable Multimodal Routing in F1-mod and accuracy, and their feature importance aligns better with human annotations on IEMOCAP and CMU-MOSI datasets.
By Mojtaba Moattari
The paper introduces a synthetic Bengali speech dataset tailored for telecom customer‑care applications, comprising 10,000 audio‑text pairs (≈26.82 hours) with predefined train, validation, and test splits. The data were generated using OmniVoice voice‑cloning, and include both original and normalized transcripts for ASR/STT use. Automatic intelligibility evaluation with a fine‑tuned Whisper model shows an average WER of 2.54% and CER of 0.59%, indicating strong text‑audio consistency, while the authors note limitations of synthetic speech and STT‑based evaluation.
By Kawshik Kumar Paul, Md. Nafiul Alam Fuji
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.
By Wenzheng Jiang, Xuankun Rong, Yuanzhao Zhai, Dawei Feng, Huaimin Wang
arXiv:2608. 20810v1 Announce Type: cross Abstract: Multimodal information systems increasingly route generated visual content back through the same vision-language index that informed its production, so the output must remain retrievable by the queries it was meant to serve.
By Guangyuan Dong, Chuang Liu, Yangchen Zeng, Haoyu Wang, Xiaoyang Yu, Pinlong Zhao, Yuchao Hou, Ziwei Li, Zheng Lin
TLive-Omni is an omni‑modal understanding model designed for e‑commerce live streaming, integrating image, video, audio, and text inputs into a unified representation. It introduces Per‑vGrid for timestamped token organization, a three‑stage supervised training pipeline, and a Faithful‑RFT reinforcement fine‑tuning stage to enhance answer faithfulness and expression quality. The model is supported by a scenario‑oriented capability taxonomy and a compact data production engine that generates training signals for tasks such as speech recognition, product visual grounding, and omni‑modal QA, achieving strong performance on live‑commerce benchmarks and good generalization to general tasks.
By Yibo Hu, Yu Qian, Mao Gu, Yingfan Tao, Yuhao Chen, Yongdong Luo, Zhuoqun Liu, Meiguang Jin, Junfeng Ma
arXiv:2608.20932v1 Announce Type: new
Abstract: Layout-to-image generation enables explicit spatial control through bounding-box layouts, yet bounding boxes specify only instance locations and cannot...
By Wenyang Hong, Yuan Wang, Yanbin Hao, Lanqing Xue, Ke Wang, Xiang Wang, Kuien Liu, Richang Hong
AEGIS is a policy enforcement component designed to prevent resource abuse in the Model Context Protocol (MCP), an open‑source JSON‑RPC protocol that allows large language models to interact with external systems via tools. By leveraging large language models to analyze, categorize, and normalize diverse tool invocations across text, images, video, and location modalities, AEGIS provides administrators with fine‑grained safeguards that can be enforced through the Open Policy Agent and ContextForge AI Gateway. This approach enables detection and mitigation of abusive behaviors—such as excessively large search requests or long video calls—while maintaining the flexibility of MCP‑based agent ecosystems.
By Shriti Priya, Teryl Taylor, Frederico Araujo