The paper introduces VoxPrivacy, a benchmark for assessing interactional privacy in Speech Language Models (SLMs). It evaluates models on a 32‑hour bilingual dataset across three difficulty tiers, revealing that most open‑source SLMs perform near random on conditional privacy decisions and even strong closed‑source systems struggle with proactive privacy inference. The authors also validate these findings on a real‑speech subset and show that fine‑tuning on a 4,000‑hour training set can improve privacy‑preserving capabilities while maintaining robustness.
By Yuxiang Wang, Hongyu Liu, Dekun Chen, Xueyao Zhang, Zhizheng Wu
The paper introduces FedRCD, a family of federated causal discovery algorithms for linear non‑Gaussian acyclic models (LiNGAM). FedRCD overcomes limitations of the existing FedISHC method, which fails under near‑symmetric noise, by leveraging higher‑order cumulant tensors that aggregate across independent client data in a single communication round. Three variants of FedRCD trade off communication rounds against algebraic noise, with two exact federated counterparts of centralised HC and HC‑LiNGAM, and a single‑round variant that supports exact unlearning at any granularity.
By Pablo Torrijos, Fabio Stella, Jos\'e A. G\'amez, Jos\'e M. Puerta
The paper introduces XCal-FL, a federated learning algorithm that dynamically calibrates differential privacy noise using three signals—prediction logit variations, counterfactual margins, and saliency concentration—to improve both predictive accuracy and explanation fidelity. Experiments on medical imaging datasets demonstrate that XCal-FL outperforms static-noise and state‑of‑the‑art adaptive DP methods, achieving over 10% better accuracy and up to fivefold higher explanation fidelity while using privacy budgets more efficiently. The study highlights that explanation fidelity behaves non‑linearly with privacy loss, indicating that explainability is a separate dimension of the privacy trade‑off.
By Michael Khavkin, Kichang Lee, Jaeho Jin, JeongGil Ko, Eran Toch
The paper introduces FlexibleFusion, a method for infrared-visible object detection that adapts to both complete and missing-modality scenarios. It employs a Modality-Aware Experts Collaboration mechanism to selectively fuse cross-modal or intra-modal pathways, and a Residual Self-Paced Entropic Optimal Transport module to align heterogeneous feature distributions without heavy optimization. Experiments demonstrate consistent performance across various modality configurations.
By Yue Zhao, Hua Yu, Yukun Zhao, Yuzhi Zhang, Maoguo Gong, Xin Mei, Zhuping Hu, Yanchi Li, A. K. Qin
The paper introduces IndicReStruct, a benchmark dataset comprising two variants—GSM8K-Reordered and GSM8K-Voice—derived from GSM8K to test multilingual LLMs on Hindi and Malayalam. It evaluates six state‑of‑the‑art LLMs under constrained constituent reordering and active‑passive voice transformation, finding consistent and significant drops in mathematical reasoning performance. Qualitative error analysis and residual‑stream activation patching reveal that failures often stem from disrupted entity‑quantity alignment, with intermediate transformer layers playing a key role in restoring reasoning.
By Karthika Nhayakkat, Rajat Verma, Maharaj Brahma, Vetcha Gnana Mahesh, Maunendra Sankar Desarkar, Ganesh Ramakrishnan, Rohit Saluja
The paper introduces OmniRSCLIP, an end‑to‑end contrastive learning framework that extends the CLIP architecture to handle heterogeneous remote sensing sensors such as SAR, multi‑spectral imaging, and hyperspectral imaging. It achieves this by employing Spectral‑Spatial Basis Decomposition to adapt arbitrary‑channel inputs without losing pretrained visual knowledge, and a spectral‑context‑aware mask‑based contrastive learning scheme to improve fine‑grained image‑text alignment. The authors also build OmniRS5M, a large‑scale image‑text corpus covering multiple sensor modalities, and demonstrate that OmniRSCLIP maintains strong RGB performance while effectively supporting these diverse remote sensing data types.
By Xiangyang Miao, Kelu Yao, Yekai Huang, Xiaogang Xu, Junxiao Xue, Minjun Shen, Chenghui Lv, Shanji Liu, Yaying Chen, Chao Li
The paper presents an end‑to‑end JEPA world model that enhances latent prediction with inverse dynamics and state alignment to improve goal‑conditioned robotic planning. By preventing latent collapse and grounding representations in physical configuration, the model achieves top success rates on tasks such as TwoRoom, PushT, and OGBench‑Cube, outperforming the baseline LeWorldModel. Ablation studies confirm that state alignment consistently boosts planning success over inverse dynamics alone across all four benchmark tasks.
By Muyuan Liu (GENISOM AI, Beijing, China), Yue Huang (GENISOM AI, Beijing, China), Zheng Liang (GENISOM AI, Beijing, China), Xiang Gao (GENISOM AI, Beijing, China)
The paper investigates whether the text of chain‑of‑thought reasoning steps actually reflects their true importance for a model’s final answer. By defining step importance as the advantage in expected reward when a step is included, the authors use Monte Carlo rollouts to estimate ground truth and then test whether large language model judges can identify high‑advantage steps. They find that capable LLMs can beat a prevalence baseline but still fall far short of a noise ceiling, and that fine‑tuning a step‑level critic improves detection for incorrect responses but remains distant from the ceiling for correct ones, indicating that step importance is only partially recoverable from the reasoning trace text.
By Kevin Du, Alexander Hoyle, Laura Ruis, Acyr Locatelli
The paper introduces a federated inference framework that enables multiple commercial large language model (LLM) APIs—such as LLaMA‑3.3‑70B, GPT‑4o‑mini, and Claude‑3‑Haiku—to collaborate on cognitive diagnosis tasks without accessing raw student data or proprietary model internals. Each entity’s predictions are perturbed with Laplace noise to provide epsilon‑local differential privacy, and a residual‑based aggregation scheme mitigates model heterogeneity. Experiments on three educational benchmarks demonstrate strong privacy guarantees with minimal accuracy loss, confirming the framework’s practical usability and cross‑domain generalizability.
By Yagna Manasa Boyapati, Chong Yu, Tianyu Jiang, Justin Zhan
PrivateHub is a contrastive diffusion model designed to generate synthetic multi‑sensor data that protects private user activities while keeping non‑private applications detectable. It operates in two stages: App‑Conditioned Pre‑training, which conditions the model on application embeddings, and App‑Aware Fine‑tuning, which uses contrastive learning to separate private from non‑private data. Experiments on three real‑world datasets demonstrate that PrivateHub reduces private‑application inference accuracy by 40–50% without harming non‑private performance and remains robust even when attackers retrain on the synthetic data.
By Jiechao Gao, Yuandong Pan, Jie Wang, Michael Lepech, Bradford Campbell
The paper introduces a simple, model‑agnostic time‑domain augmentation called Sliding‑Window Reordering with Overlap Averaging. It transforms the joint input‑target sequence into overlapping windows, randomly reorders a fraction of them based on a variance criterion, and reconstructs the sequence by averaging overlaps to generate synthetic samples with controlled variation and minimal temporal distortion. Experiments show strong performance gains across nine long‑term forecasting benchmarks and four short‑term traffic benchmarks, with detailed ablations and diagnostics highlighting the effectiveness of each design choice.
By Jafar Bakhshaliyev, Johannes Burchert, Niels Landwehr, Lars Schmidt-Thieme
The paper introduces View distance, a novel metric that projects high‑dimensional data onto all pairwise two‑dimensional planes and sums the Euclidean distances across these projections. It satisfies metric axioms, couples features, suppresses redundancy, and captures anisotropic geometry. To make it scalable, the authors propose a plane‑selection strategy using iterative Maximum Weight Matching, reducing complexity from ω(n²) to ω(k) and demonstrating competitive performance on twelve datasets.
By Yiqun Zhang, Hou-biao Li
The paper introduces MSign, an optimizer designed to prevent training instability in large language models by restoring the stable rank of weight matrices. It identifies two precursors to gradient explosions—rapid stable rank decline and increased Jacobian alignment—and proves that these jointly cause exponential gradient growth. Experiments on models ranging from 5 M to 3 B parameters show that MSign stops training failures while adding less than 7.0% computational overhead.
By Lianhai Ren, Yucheng Ding, Xiao Liu, Peng Cheng, Yeyun Gong
The paper introduces PACE, a dataset that tests whether personalized assistants can detect hidden conflicts in user requests by integrating implicit, egocentric knowledge from a knowledge base. PACE pairs user requests with persona-specific facts, requiring models to retrieve contextual evidence before deciding if a request is inappropriate. The authors also propose PaceMaker, a multi-agent system that improves conflict detection by coordinating query reformulation, multi-hop graph traversal, and conflict-aware filtering, outperforming existing methods on the PACE benchmark.
By Yoojin Kim, Jihyoung Jang, Hyounghun Kim
The paper introduces a novel zero‑shot anomaly detection framework for multivariate IoT traffic data that combines adversarial learning and contrastive loss within a sequence‑based Variational Autoencoder. It achieves domain‑invariant latent representations and semantically structured embeddings without labeled data, using encoder/decoder adaptor layers to align feature distributions and a destination‑based segmentation strategy to model real‑world communication patterns. The method is evaluated on six diverse datasets across 44 transfer scenarios, showing strong zero‑shot generalization and competitive performance against a contrastive domain‑adaptation baseline in heterogeneous, privacy‑constrained IoT environments.
By Mahshid Rezakhani, Tolunay Seyfi, Fatemeh Afghah
The paper introduces Targeted Active Search (TAS), a black-box attack that extracts forgotten prompts from models that have undergone unlearning. TAS first identifies forgotten entities by querying the model with canonical templates and an entity pool, then reconstructs the prompts by instantiating templates with these entities. Experiments on three unlearning methods, datasets, and large language models demonstrate that TAS can recover 100% of forgotten entities and up to 95% of prompts while using far fewer queries than naive probing.
By Au Ashley Hoi-Ting, Meghdad Kurmanji, William F. Shen, Nicholas D. Lane, Ligang He
ObserverBench is a benchmark framework that evaluates whether internal mechanistic estimators—called observers—are suitable for guiding interventions, control, or safety actions in language models. It separates estimation accuracy from the loss incurred by the chosen action, showing that accurate predictions do not always lead to better decisions. Experiments on GPT‑2‑small, Qwen2.5‑7B, Gemma‑2‑9B‑it, and Qwen3.5‑9B demonstrate that observers trained on action loss can reduce deployment loss, while traditional metrics like AUROC may rank monitors differently from actual performance.
By Vijay Erramilli
EasySteer is a unified framework for high‑performance, extensible large‑language‑model steering built on vLLM. It offers a modular architecture with pluggable interfaces for analysis‑based and learning‑based methods, fine‑grained parameter control, pre‑computed steering vectors for eight application domains, and an interactive demo system. Integrated with vLLM’s optimized inference engine, EasySteer delivers a 10.8–22.3× speedup over existing frameworks and demonstrates effectiveness in overthinking mitigation, hallucination reduction, and other key applications.
By Haolei Xu, Xinyu Mei, Yuchen Yan, Rui Zhou, Wenqi Zhang, Weiming Lu, Yueting Zhuang, Yongliang Shen
The paper introduces a method to detect which web scrapers feed data to large language models (LLMs) by deploying dynamic websites that issue unique canary tokens to each scraper. By querying LLMs for information about these sites, the authors can identify when an LLM consistently outputs the unique tokens, indicating exposure to a specific scraper. Experiments on 22 production LLM systems show the technique reliably uncovers both known and undisclosed scrapers, offering a tool for third parties to monitor and control unwanted web scraping.
By Steven Seiden, Triss Ren, Caroline Zhang, Taein Kim, Enze Liu, Emily Wenger
The paper investigates how different post‑training methods—supervised fine‑tuning, reasoning‑augmented fine‑tuning, and preference optimization (ORPO)—affect the internal computation of refusal behavior in language models. Experiments on Llama‑3.1‑8B, Gemma‑2‑9B, and Qwen3‑8B show that reasoning‑augmented training consistently creates a distinct refusal computation across models, while the architecture influences the internal structure and steerability of refusal. None of the studied methods simultaneously achieve a distributed refusal mechanism, preserve general capability, and allow easy corrective edits, indicating that current post‑training approaches are not a fully reliable defense for safety-critical applications.
By Hoang Cuong Nguyen, Mark Dras, Usman Naseem