The paper introduces Cumulative Marginal Matching (CMM), a new objective for training few‑step diffusion and flow‑matching models. CMM treats flow‑matching as a sequence of intermediate marginal distributions and directly optimizes the cumulative Wasserstein‑2 discrepancy between the base and few‑step models across the entire trajectory. It derives a tractable surrogate for this objective, shows that Shortcut‑CSL is a first‑order approximation of CMM, and reports improved generation quality on standard benchmarks under the same training budget.
By Paribesh Regmi, Rui Li
EmbeddingGemma 2 is an open multimodal embedding model that natively represents text, images, audio, video, and their combinations in a unified space. It introduces simplified per‑layer embeddings and a multi‑stage curriculum that improves quality across unimodal, cross‑modal, and mixed‑modal tasks. The model sets a new Pareto frontier in comprehensive benchmarks, outperforming specialist and larger models, while offering flexibility for resource‑constrained settings through selective encoder loading, adjustable token budgets, truncatable dimensions, and quantized weights.
By Henrique Schechter Vera, Sahil Dua, Jovan Andonov, Sara Smoot, Sebastian Russo, Ivan Llanos, Max Matveev, Jae Hyeon Yoo, Kat Black, Shih-Cheng Huang, Josiah Aklilu, Maarten Grootendorst, Omar Sanseviero, Chandramouli Amarnath, Ian Ballantyne, Zhe Li, Daniel Salz, Ibrahim Alabdulmohsin, Ilya Philippov, Kristen Quan, Glenn Cameron, Edouard Yvinec, Luciano Martins, Doug Reid, Sindhu Raghuram Panyam, Weiyi Wang, Hengxuan Ying, Madhuri Shanbhogue, Tanmaya Dabral, Shanfeng Zhang, Juyeong Ji, Renjie Wu, Joe Zou, Xu Chen, Yunan Zhang, Yu-Hui Chen, Matthew Soulanille, Vinod Mamillapalli, Yi-Chun Kuo, Wai Hon Law, Matt Kreileder, Alice Zheng, Suril Shah, Mogan Shieh, Pooja Consul, Alberto Montes, Andreas Hess, Andres Sevilla, Babak Samari, Brendan McCloskey, Brian Potetz, Daniel Cer, Georgios Kathareios, Gustavo Hern\'andez \'Abrego, Hesen Zhang, Ji Ma, Jiaxing Wu, Keegan Mosley, Paul Suganthan, Rui Meng, Shen Gao, Shengcao Cao, Yi-Ting Chen, Yichang Chen, Zhongli Ding, Kendra Byrne, Cormac Brick, Olivier Lacombe, Qin Yin, Tris Warkentin, Armand Joulin, Tom Duerig, Mojtaba Seyedhosseini
arXiv:2610. 37657v1 Announce Type: cross Abstract: We introduce Heads of Security (HoS), a prompt injection defense that enforces instruction-data separation via a small set of learned per-head attention gates.
By Igor Shilov, Xiaoxue Yang, Alexander Panfilov, Evgenii Kortukov, Matthieu Meeus, Peter Romov, Wojciech Samek, Jonas Geiping, Yves-Alexandre de Montjoye
This systematic review examines the use of natural language processing (NLP) throughout the entire battery life cycle, rather than focusing on a single stage or method. It introduces a new technical language processing (TLP) framework tailored for the European Union’s digital battery passport and other battery prediction tasks. The study follows PRISMA guidelines, searches three major databases, and ultimately evaluates 66 relevant papers out of 274, highlighting emerging NLP tasks that aid materials discovery while noting challenges such as the absence of standard benchmarks.
By Tosin Adewumi, Martin Karlsson, Marcus Liwicki, Mikael Sj\"odahl, Lama Alkhaled, Rihab Gargouri, Nudrat Habib
The paper introduces AZeroS, a speech‑adapted large language model that uses Self‑Generated Instruction‑Free Tuning (SIFT) to avoid task‑specific instruction data. By training only two lightweight projection modules on 25,000 hours of speech‑text pairs and 3,000 hours of paralinguistic labels, AZeroS achieves state‑of‑the‑art results on semantic and paralinguistic benchmarks while keeping the core LLM and audio encoder frozen.
By Yiwen Shao, Wei Liu, Jiahong Li, Tianzi Wang, Kun Wei, Meng Yu, Dong Yu
PubMedCausal is a new span‑level annotated corpus for causal relation extraction in biomedical text, comprising 30,000 paragraph‑level rows with 3,945 causal rows and 6,491 adjudicated cause–effect pairs. Each causal relation is annotated with full‑text cause and effect spans, causality type, and sententiality, allowing evaluation of both causal detection and full‑span extraction. Benchmarking shows biomedical encoders like PubMedBERT achieve an F1 of 0.7391 for detection, while the generative model DeepSeek‑R1‑32B reaches a Cosine Pair F1 of 0.6765 for extraction, and the corpus supports cross‑dataset transfer learning.
By Ifeoluwa Kunle-John, Josiah Paul, Oluwatosin Agbaakin, Peter Aina, Ikenna Odezuligbo, Sydney Anuyah
The paper introduces the Agent Planning Benchmark (APB), a diagnostic framework designed to evaluate the planning capabilities of large language model (LLM) agents. APB contains 4,209 multimodal cases across 22 domains and five settings, testing holistic planning, step-wise planning with feedback, and robustness to extraneous or broken tools and unsolvable tasks. Evaluation on 12 MLLMs reveals systematic weaknesses in long-horizon planning, tool-noise robustness, calibrated refusal, and inference-time refinement, while APB-guided refinement improves plan correctness and downstream execution metrics on ToolSandbox and τ²‑bench tasks.
By Haoyu Sun, Wenxuan Wang, Mingyang Song, Jujie He, Weinan Zhang, Yang Liu, Yang Yang, Yu Cheng
The paper introduces PARALLELEVENTS, a benchmark that simulates realistic future event trajectories to evaluate knowledge updates in large language models without contaminating data or conflicting with existing knowledge. It also presents SYNAPSE, a training framework that uses model-generated synthetic data for instruction tuning, enabling scalable integration of new knowledge. SYNAPSE achieves a 14.23‑point improvement in QA accuracy over previous methods, demonstrating the effectiveness of simulation‑based synthetic training for coherent knowledge insertion.
By Jonathan Zheng, Zirui Shao, Alan Ritter, Wei Xu
BanglaBBQ is a question‑answering bias benchmark that captures Bangladeshi stereotypes in both Bangla and English. It adapts the original BBQ by classifying templates into directly translated, modified for Bangladesh, and removed categories, retaining 140 templates across six groups. Five new stereotype categories—marital status, age and political perception, regional origin, educational background, and mental‑health status—were added, with 466 new templates evaluated and 372 retained after annotation. The dataset was tested on Qwen3‑4B, Qwen3‑8B, GPT‑4.1‑mini, and Gemini 2.5 Flash, revealing that all models tend to favor stereotyped individuals when context is ambiguous, with language‑specific differences and varying accuracy depending on stereotype confirmation or contradiction.
By Sumit Anthony Gomes, Yasin Rahman, Taeeba Tasnia, Faiza Nooren Suhita
The paper introduces Head-wise Modality Decoupling and Regulation (HMDR) to improve fake news detection when some modalities (image or text) are missing. HMDR first decouples modality representations via soft head grouping based on attention preference, then regulates head contributions to balance modality attention, and finally enforces modality invariance by aligning representations when one modality is replaced. Experiments on two benchmarks demonstrate that HMDR enhances detection performance under missing modalities while preserving accuracy on complete inputs.
By Kai Qian, Dongjie Xu, Weijie Shi, Yezhao Xu, Jiaqi Wang, Hanghui Guo, Mengze Li, Hao Chen, Ziyi Liu, Jiajie Xu
The paper "Who Gets Flagged? The Pluralistic Evaluation Gap in AI Content Watermarking" discusses how AI content watermarking, used for authentication and provenance, varies in signal strength and robustness across different languages, cultural visual traditions, and demographic groups. It reviews existing benchmarks and finds that most do not report performance across these dimensions, highlighting a bias risk. The authors propose three evaluation dimensions—cross‑lingual detection parity, culturally diverse content coverage, and demographic disaggregation of detection metrics—to ensure fair and inclusive watermarking practices.
By Alexander Nemecek, Osama Zafar, Yuqiao Xu, Wenbiao Li, Erman Ayday
Tail Cues, Principal Corrections: Plug-and-Play Rectification for Open-Set Test-Time Adaptation introduces TALOR, a lightweight module that rectifies features for non-stationary target streams by decomposing them into high-gain principal and low-gain tail coordinates. TALOR applies a soft-routed affine correction in the principal subspace, subtracts drift via weighted ridge regression, and reconstructs the feature for the source head. On ImageNet-C with six csOOD datasets, TALOR surpasses the next-best UniEnt by 2.2% H-score, runs 2.5× faster, and uses only 12% of its GPU memory, while also boosting COME to 68.4% H-score with minimal memory overhead.
By Yingkai Yang, Chaoqi Chen, Hui Huang
The paper introduces S‑MME, a benchmark for diagnosing shortcut‑biased behavior in Multimodal Large Language Models (MLLMs) on long‑video question answering. It proposes CLEAR, a training‑free framework that estimates and mitigates bias by treating the full video‑question pair as a complete view and intervening on hidden states during inference. Experiments on five benchmarks demonstrate that CLEAR improves prediction accuracy and robustness across different MLLMs.
By Xusheng Liang, Jianqiao Sun, Hao Zhang, Yulei Niu, Hengshuang Zhao, Jiawei Ma
The paper introduces HiRe-DRR, a Hierarchical Multimodal Interaction Reasoning framework for dyadic relationship recognition that organizes evidence across feature level, interaction type, and relationship hierarchy. It decomposes facial, bodily, acoustic, and textual cues into low- and high-level representations, builds homogeneous and heterogeneous cross-participant interactions, and employs a Mixture-of-Interaction Transformer with relation-aware sparse routing. A Hierarchical Conditional Loss provides coarse-to-fine supervision, and experiments on three benchmarks show HiRe-DRR outperforms existing image, video, and multimodal methods, with routing analysis revealing distinct interaction patterns for distant versus close relationships.
By Wang Tang, Fethiye Irmak Dogan, Linbo Qing, Ce Zhu, Hatice Gunes
The paper evaluates privacy‑aware federated learning for breast cancer imaging by comparing pretrained vision‑language models (CLIP, BiomedCLIP) and a ViT‑B/16 baseline on two public datasets, with and without differential privacy. All five models achieve high normal‑versus‑sick discrimination (ROC‑AUC 0.940‑0.990), but acquisition year alone attains 0.974, indicating potential confounding. Removing the text pathway keeps accuracy while cutting communication cost, while differential privacy lowers ROC‑AUC by 0.375 due to perturbing the full model state.
By Malithi Mithsara Wanniarachchi Kankanamge, Ning Yang, Sree Bash Debnath, Md Zobaer Islam, Abdur R. Shahid
SO-101 Bench is a simulated benchmark for language‑conditioned spatial manipulation on the low‑cost SO‑101 robotic arm, built using an Isaac Lab digital twin of a real setup. It features four tasks across five families, 138 objects with seen/unseen splits, 916 evaluation episodes, and 3,629 simulated demonstrations for policy fine‑tuning. The benchmark’s rankings mirror real‑world performance (Spearman ρ = 1.00) and reveals that while learned policies reach about 61% success, an LLM‑driven agent that writes code and delegates grasping achieves 83.2%.
By Truman Hickok
Nemotron‑Labs‑Diffusion‑Uni (NLD‑Uni) is a unified discrete diffusion large language model designed for visual understanding and generation. It introduces a cross‑paradigm on‑policy distillation framework that transfers capabilities from autoregressive VLMs and continuous flow models into a single DLLM, and a progressive self‑refining diffusion process that corrupts tokens gradually during training. Experiments show NLD‑Uni outperforms other unified DLLMs on tasks such as visual question answering, multi‑modal reasoning, object grounding, text‑to‑image generation, and image editing.
By Shufan Li, Hanrong Ye, Greg Heinrich, Yonggan Fu, Haotian Zhang, Sifei Liu, Aditya Grover, Pavlo Molchanov
SAGE (Semantic-Augmented Geometric Experts) is a new point‑cloud registration method that incorporates frozen, pre‑trained semantic representations into the registration pipeline without extra sensors or fine‑tuning. It uses a Semantic‑Guided Mixture of Experts (SG‑MoE) to route data so that each expert specializes in specific semantic‑geometric patterns, thereby reducing mismatches that plague geometry‑only approaches. On the 3DLoMatch benchmark, SAGE achieves the highest registration recall and, with the same number of refinement iterations, also attains the best inlier ratio and recall on both 3DLoMatch and 3DMatch.
By Tomoyasu Shimada, Kazuhiko Murasaki, Shoichiro Takeda, Shogo Sato, Riku Inoue, Toshihiko Nishimura, Ryuichi Tanida
Hy‑Nomas is an 8B diffusion‑based video world model that enables free‑form text instructions and continuous camera control across multiple interaction turns. It generates variable‑length video segments conditioned on current inputs and cached history, using structured video captions and action‑aware causal attention to keep descriptive context accessible while restricting action instructions to their intended temporal windows. A vision‑language Prompt Enhancer translates user requests and multimodal histories into structured conditions, and a multi‑stage curriculum trains the model progressively before diffusion distillation, resulting in superior performance in human evaluations.
By Qiaoling Chen, Jiaxiang Cheng, Xueqing Deng, Tiankai Hang, Ji Hou, Jiaqi Li, Yifan Liu, Tao Lu, Junshu Tang, Yu Tang, Yao Teng, Yu Tian, Heng Wang, Weizhi Wang, Xiyao Wang, Xilun Wu, Mingyu Yang, Yangbin Yu, Rui Zhao, Longjun Cao, Linghao Chen, Haoyi Duan, Jiaming Han, Mingji Han, Enhua Jiang, Shunkai Li, Yanzhe Liang, Qinwei Ma, Cong Wang, Lei Wang, Bing Wu, Ke Xing, Ray Ying, Zhentao Yu, Key Zhang, Yuan Zhou, Zhengguang Zhou, Kai Zou
GeoGSplat introduces a geometry‑grounded, feed‑forward Gaussian splatting framework that avoids the pixel‑aligned prediction bottleneck by using a fixed number of Gaussians. A dual‑branch encoder builds coarse 3D anchors and fine triplane features, while a hierarchical decoder progressively generates dense 3D points and Gaussian primitives for reconstruction and novel view synthesis. The method outperforms state‑of‑the‑art pixel‑aligned and non‑pixel‑aligned splatting techniques on DTU and NRGBD datasets, especially when rendering views far from the input cameras.
By Jiale Wang, Chen Zhao, Haoyang Zhang, Yanhao Wu, Fei He, Yanhu Shan, Rui Wu, Wei Ke, Tong Zhang