The paper presents a study on securing large language model (LLM) inference services against adversarial use, such as jailbreaking, denial‑of‑service, and distillation attacks. It introduces a structural causal model to generate a realistic, labeled dataset of user sessions, including coordinated multi‑account campaigns and varying label observability. Using this dataset, the authors train a gradient‑boosted detector that achieves near‑perfect binary classification on oracle labels but shows lower performance on operational labels, and they demonstrate a simple thresholded engine that improves attack‑type attribution accuracy.
By Keifer Lee
ThunderEP is a new communication design for expert parallelism on PCIe-connected consumer GPUs that eliminates relay hops, uses DMA engines to avoid GPU compute contention, and reduces CPU polling overhead. Integrated into vLLM, it outperforms NCCL on RTX 4090 and RTX 5090 GPUs, delivering average speedups of 2.00× for dispatch, 1.53× for combine, and up to 1.66× end‑to‑end over existing MoE inference frameworks.
By Jaehwan Lee, Sangmin Lee, Chaewon Kim, Junsik Shin, Jaejin Lee
The article outlines a Ph.D. research agenda aimed at creating provably efficient and practical algorithms for data‑driven sequential decision‑making under uncertainty. It focuses on reinforcement learning and multi‑armed bandits, targeting applications such as recommendation systems, computer networks, video analytics, and large language models. The work seeks to overcome limitations of existing methods—such as reliance on idealized models, lack of robustness to adversarial perturbations, and poor instance‑dependent performance—by developing algorithms that are more efficient, robust, instance‑adaptive, and generalizable to new environments.
By Zhiyong Wang
The paper introduces SEAR, a four-task benchmark that evaluates audio language models (ALMs) for audio deepfake detection by requiring acoustic evidence identification, quantification, deepfake detection, and forensic rationale generation. It also proposes BAEA, an acoustic evidence agent that equips a frozen ALM with controlled acoustic tools under fixed or adaptive evidence-acquisition policies. Experiments with six ALMs show a gap between plausible rationales and verifiable acoustic evidence reasoning, and demonstrate that BAEA‑Fixed improves final verdicts and forensic rationales while revealing that misleading evidence harms detection and grounding performance.
By Rong Wan, Suliu Qin, Jiaxi Li, Wei Xie, Wenwu Wang, Xiaolong Han, Lu Yin, Xilu Wang
The paper introduces Model-to-Data (M2D) distillation, a method that transfers properties learned by a complex graph neural network (GNN) teacher into the graph data itself. By jointly learning augmented node features and graph structure, M2D encodes the teacher’s behavior, allowing simpler GNNs to recover high predictive performance, fairness, and robustness. The resulting graph can be reused with various downstream models, enabling them to approximate sophisticated teachers such as fairness-aware GNNs, Graph Attention Networks, and Graph Transformers.
By Debolina Halder Lina, Arlei Silva
GPEC is a lightweight pre‑LLM Gaussian Process Embedding Correction that refines visual representations for cardiac ultrasound caption generation. By inserting a residual correction layer between the visual projection and the language model, it aligns projected embeddings with annotation‑guided targets derived from structured video annotations. Experiments show that GPEC improves caption similarity and content alignment while adding less than 0.05 s of inference overhead, all without fine‑tuning the pretrained multimodal backbone.
By Arefeh Rezaei
The paper introduces RateAR, a benchmark of AR images and videos that captures perceptual factors such as object placement, scale, and shadow consistency. Eleven commercial vision‑language models (VLMs) are evaluated on this benchmark, with VLM‑based quality predictions showing strong correlation with human judgments (Spearman’s ρ up to 0.8695). Using contextual prompting, the authors build an automated AR content adjustment system that, in a 21‑participant study, improved placement and size coherence of virtual content for over 90% of users.
By Elias Rotondo (Duke University), Lin Duan (Duke University), Yanming Xiu (Duke University), Sangjun Eom (Duke University), Conrad Li (Duke University), Maria Gorlatova (Duke University)
The paper introduces Align Then Reason (ATR), a multilingual lip‑sync judge that aligns frame‑level lip representations with phonetic units of a candidate text line and then uses a language model to evaluate both content and timing. ATR achieves significant improvements over existing baselines on a seven‑language benchmark, with mean AUC gains of up to 59.4% for 2B reasoners and similar gains across other LLM families. The method also transfers well to unseen languages and outperforms lip‑reading baselines on real dubbing tasks such as dub‑line reranking and script‑to‑clip assignment.
By Rui Liu, Bhavin Jawade, Haoqi Li, Shivam Mehta, Karan Saxena, Yinghong Lan, Cameron R. Wolfe
Dyna3 is a training‑free framework that extends the depth foundation model DA3 to perform 4D dynamic scene reconstruction without fine‑tuning. By leveraging DA3’s cross‑view features and a best‑match search, it distinguishes static surfaces from moving objects, and uses vision‑language models to generate semantic prompts for SAM 3 to achieve precise instance‑level segmentation. Experiments on four datasets show Dyna3 outperforms correspondence‑trained methods, improving dynamic object segmentation by +5.5 pp, speeding pose estimation 13×, and reducing memory usage 4–8×.
By Xinhao Xiang, Weiyang Li, Zhijie Zheng, Abhijeet Rastogi, Jiawei Zhang
The paper introduces MOSAIC, a method for personalized federated vision‑language models that addresses domain heterogeneity by focusing on class‑specific cross‑domain residuals. It constructs a decision‑aware harmfulness score to identify residuals that hurt image‑text decision margins, then uses a low‑rank residual adapter with shared class factors and private domain factors, an image‑conditioned gate, and harmful‑pair‑aware reweighting to refine local updates. Experiments on Office31, OfficeHome, and DomainNet100 show consistent improvements in macro‑client top‑1 accuracy across various domain‑shift scenarios.
By Wentao Yue, Qingyu Mao, Tianyou Lai, Ahmed M. Abdelmoniem, Qilei Li
The paper introduces a Multi-layer Fusing Transformer that uses cross‑attention to merge visual and textual features across multiple layers, enabling extraction of information from low to high levels. Experiments and ablation studies on the ViVQA dataset demonstrate that this architecture outperforms competitive baselines for Vietnamese Visual Question Answering. The work addresses a notable gap in VQA research for Vietnamese, a language with limited prior resources.
By Cong Phu Nguyen, Huy Tien Nguyen, Tung Le
MapLightning introduces a new approach to online vectorized HD map construction that replaces dense bird’s‑eye‑view grids with a compact set of 1D learnable map tokens. Using a transformer‑based mapper that applies full self‑attention to combined image and map tokens, the method achieves efficient inference, lower memory usage, and improved global context without relying on camera projection parameters. Experiments on nuScenes and Argoverse 2 show state‑of‑the‑art accuracy, with a lightweight variant outperforming MapTRv2 by over 10 mAP and running 1.73× faster at 40+ FPS while using 53 % less memory.
By Shen Zheng, Anurag Ghosh, Mani Ramanagopal, Srinivasa Narasimhan
The paper introduces Anti-Persona, an image-level defense that prevents unauthorized identity binding and recognition in personalized vision–language models. By aggregating visual features from multiple reference images into an identity prototype and applying subtle perturbations, the method disrupts prototype alignment in the vision‑encoder space while maintaining visual fidelity. Experiments on two personalized LVLMs show protection rates up to 95.0% and robustness across prompt variations and encoder mismatches.
By Abhishek Basu, Fahad Shamshad, Karthik Nandakumar
The paper introduces Adaptive Tversky Policy Optimization (ATPO), a reinforcement learning framework for multi‑label video safety detection. ATPO uses an Adaptive Tversky Reward (ATR) that dynamically adjusts false‑positive and false‑negative penalties, allowing controllable precision‑recall trade‑offs. Experiments on SafeWatch‑Bench and XD‑Violence demonstrate significant performance gains, raising the Jaccard Index from 40.66 to 75.44 on SafeWatch‑Bench‑Real.
By Guangyu Yang, Jingbiao Mei, Mingsheng Sun, Jinghong Chen, Yingtong Bu, Pengda Qin, Da Chen, Bill Byrne
FocusGraph is a framework for selecting keyframes in egocentric long‑video question answering. It uses a lightweight Scene‑Graph LLM Selector to identify query‑relevant clips from graph‑based captions, then extracts keyframes with Patch‑wise Sparse‑Flow Retention (PSFR) before feeding them to a multimodal large language model for answer generation. The method achieves state‑of‑the‑art performance on FindingDory and HourVideo while reducing question‑time inference cost.
By Tatiana Zemskova, Solomon Andryushenko, Ilya Obrubov, Viktoriia Khoruzhaia, Ekaterina Eroshenko, Ekaterina Derevyanka, Dmitry Yudin
SlotVLA proposes a slot-attention framework that models both objects and their relations for robotic manipulation, using a relation‑centric decoder and an LLM‑driven action translator. The authors introduce LIBERO+, a fine‑grained benchmark with object‑centric annotations, box and mask labels, and instance‑level temporal tracking to evaluate object‑relation reasoning. Experiments on LIBERO+ show that object‑centric and object‑relation slot representations reduce visual token counts while maintaining competitive generalization, offering a compact and interpretable foundation for multitask visuomotor control.
By Taisei Hanyu, Nhat Chung, Huy Le, Toan Nguyen, Yuki Ikebe, Anthony Gunderman, Duy Nguyen Ho Minh, Khoa Vo, Tung Kieu, Kashu Yamazaki, Chase Rainwater, Anh Nguyen, Ngan Le
ChainLoRA is a replay‑free continual learning framework for large language models that merges task vectors while preserving geometry. It uses chain‑updated training with a one‑sided orthogonality proxy to keep historical state and regularization overhead constant, and applies post‑stream adaptive SVD merging with Procrustes adaptation to separate shared and task‑specific components. Experiments demonstrate state‑of‑the‑art performance on Large and SuperNI benchmarks and competitive results on Standard CL, approaching the scores of replay‑based methods.
By Hang Yin, Haozhe Wang, Yuhua Luo, Zhangqi Pan, Xiaoxing Wang, Junchi Yan
VIDA (Visually-Dependent Ambiguity) is a new dataset comprising 2,500 curated instances designed to test whether multimodal machine translation models can use visual input to resolve ambiguous source spans. The authors also introduce Disambiguation-Centric Metrics, which employ an LLM-as-a-judge classifier to verify correct span-level resolution. Experiments with recent large vision‑language models show that visual disambiguation remains difficult, and that chain‑of‑thought supervised fine‑tuning yields stronger out‑of‑distribution performance, especially for collective‑noun ambiguities.
By Jingheng Pan, Xintong Wang, Longyue Wang, Liang Ding, Weihua Luo, Chris Biemann
IQS-BO introduces an in‑context query selection method for Bayesian optimisation that learns to choose evaluation points via supervised learning on synthetic priors. The approach uses a Prior‑data Fitted Network (PFN) to predict, in a single forward pass, the probability that each candidate maximises the objective, thereby amortising the decision step and eliminating the need for costly surrogate refitting or acquisition maximisation. Experiments show that IQS‑BO matches or surpasses traditional Gaussian‑process‑based BO and other in‑context methods on both synthetic and real‑world benchmarks, and that a mixture prior combining GP samples with challenging function types further improves performance.
By Luca Geminiani, Nadja Klein
The paper presents sample complexity bounds for sampling high‑dimensional categorical distributions with local dependence using discrete diffusions. It introduces a pinning decomposition of the discrete score, enabling a weight‑sharing neural score learner combined with τ‑leaping for efficient end‑to‑end sampling. The authors derive explicit guarantees that depend on vocabulary size, MRF interaction order, and sample size, and demonstrate that weight‑sharing networks outperform fully connected ones on Potts, Ising, and tree‑structured models.
By Shivam Kumar, Nabarun Deb