Large language models (LLMs) are increasingly used to assess social bias in text, but the passages they evaluate often contain surface noise such as typos and broken punctuation. This study applied five realistic noise conditions at varying intensities to 3,822 stereotype‑related responses and compared bias judgments on noisy versus original text. The findings show that noise disproportionately turns neutral judgments into biased ones—up to 120 times more likely—while rarely converting biased judgments into neutral ones, and that the most fragile LLM judge exhibits the greatest distortion at mild noise levels. As LLMs become more robust, the bias distortion tends toward parity rather than reversal, meaning bias measured on noisy text is systematically overestimated, especially in fairness‑critical categories.
By DongHyun Ryu, Jaehyeok Lee, YeongJun Hwang, JinYeong Bak
The paper introduces FedMVLA, a modality‑decoupled federated learning framework designed for privacy‑preserving embodied intelligence in 6G networks. It addresses the unique challenges of vision‑language‑action models by employing modality‑aware federated aggregation, privacy allocation, and communication compression, along with a precision‑critical action transport slice. A case study on federated robotic manipulation demonstrates significant gains in task success, scalability, and uplink payload reduction compared to standard FedAvg.
By Zhuodong Liu, Xiangyu Li, Chunhong Yuan, Hongyang Du, Bodong Shang, Qingqing Wu, Tony Q. S. Quek, Mohsen Guizani
The paper introduces a subgroup-targeted membership inference game to audit differentially private synthetic text releases, revealing that existing average-case attacks miss significant leakage to vulnerable subgroups. An extensive audit across 32 proxies, four datasets, three generation methods, and five privacy budgets shows that DP reduces overall leakage but leaves concentrated, uneven residual risk, especially for high-risk records. The study demonstrates that which records leak is determined by the release mechanism rather than the records themselves, challenging record-level risk assessment.
By Yidan Sun, Viktor Schlegel, Srinivasan Nandakumar, Siew Kei Lam, Anil Anthony Bharath
The paper introduces a framework for learning feature transformations on tabular data that addresses three limitations of prior generative methods: neglect of hierarchical feature relationships, bias from order-sensitive embeddings, and reliance on gradient-based search. It combines a permutation‑invariant hierarchical module using self‑attention pooling to capture interactions across features, operations, and abstraction levels, with a policy‑guided multi‑objective reinforcement learning strategy that starts from strong seeds and optimizes both predictive accuracy and transformation efficiency. Experiments on diverse tabular benchmarks show the approach outperforms strong baselines, and the authors provide public code and data.
By Rui Liu, Tao Zhe, Yanyong Huang, Sankha Narayan Guria, Xiao Luo, Wei Fan, Yanjie Fu, Dongjie Wang
The paper investigates how different velocity regimes affect broadcast‑viewport basketball analytics. Using a calibrated off‑screen imputation protocol, it compares four velocity settings—none, viewport‑legal observed, true‑for‑visible, and true‑for‑all—against a velocity‑aware ground truth across three analytical layers. Results show that velocity is largely useless for imputation, modestly useful for the control surface, and minimally useful for final team verdicts, with the visible channel providing the most benefit.
By Seongjin Choi
The paper proposes a weighted conformal prediction framework to combine outputs from multiple gravitational‑wave search pipelines, providing statistically rigorous confidence estimates for candidate events. By incorporating likelihood‑ratio reweighting, the method corrects for covariate shift between simulated training data and real observations, restoring well‑calibrated coverage. Experiments on mock datasets show that this approach improves sensitivity, especially near the detection threshold, enabling recovery of true signals that would otherwise be missed.
By Ann-Kristin Malz, Gregory Ashton, Nicolo Colombo
arXiv:2310. 10092v4 Announce Type: replace Abstract: This paper explores the use of linear aggregation to protect the privacy of sensitive training labels through the concept of \emph{label differential privacy} (label-DP) while maintaining regression task utility.
By Anand Brahmbhatt, Rishi Saket, Shreyas Havaldar, Anshul Nasery, Yukti Makhija, Aravindan Raghuveer
The paper examines the common practice in diffusion and flow matching of predicting the clean signal, converting it to a velocity, and training with a velocity-space loss. It shows that this conversion can cause unstable optimization due to singular endpoint amplification, but that prediction–loss alignment removes this source of non‑integrability and ensures a finite second moment for all timesteps, even with uniform sampling. Experiments confirm that aligned objectives remain trainable across different samplers, reconciling theoretical concerns with empirical success.
By Jiadong Hong, Lei Liu, Xinyu Bian, Wenjie Wang, Zhaoyang Zhang
mmFHE is the first system that runs the entire cloud-side mmWave sensing pipeline—including DSP and machine‑learning inference—under fully homomorphic encryption. It encrypts range profiles on an edge device, then processes them homomorphically on a semi‑honest cloud using a library of seven data‑oblivious FHE kernels that replace standard DSP routines. The authors demonstrate the approach on vital‑sign monitoring and gesture recognition, proving input privacy and data obliviousness, and show negligible accuracy loss (84.5% vs. 84.7%) with practical GPU latencies on commodity hardware.
By Tanvir Ahmed, Yixuan Gao, Adnan Armouti, Rajalakshmi Nandakumar
The paper introduces SAFA-MZ, a secure adaptive federated authentication scheme for multi‑zone non‑terrestrial networks that embeds group‑level authentication tags into a collaborative rate‑splitting multiple access transmission. It jointly optimizes high‑altitude platform station placement, user association, and power allocation to maximize secrecy spectral efficiency while meeting authentication reliability, power, and coverage constraints. The solution employs a repair‑based cross‑entropy method and a graph‑aware advantage actor‑critic algorithm, achieving up to 135% higher average secrecy spectral efficiency compared to single‑connect transmission.
By Parsa Rajabi, Mohammad Mirzaee, Mohammad Reza Abedi, Nader Mokari, Paeiz Azmi
RAG-Safety-Bench is a benchmark designed to evaluate how retrieval-augmented generation (RAG) affects the safety of large language models (LLMs). It isolates safety impacts by testing four conditions: non-RAG, RAG with an oracle document, RAG with related but non-answer documents, and RAG with random safe documents. Results on five open-source LLMs reveal an inverse relationship between benign and unsafe capabilities, show that baseline safety guardrails do not guarantee safety in RAG, and confirm that even benign documents can trigger unsafe generation.
By Adithiyan Rajan Indira Saravanan, Kathleen C. Fraser
The paper introduces a per-layer differential privacy (DP) clipping strategy for federated multilingual speech large language models (speech‑LLMs). It demonstrates that standard single‑pool per‑layer DP methods fail due to a cross‑component budget collapse caused by large norm differences between acoustic encoders and language decoders. The authors propose an α‑split two‑pool allocation that normalises encoder and decoder parameters separately, preserving the overall DP guarantee while restoring word error rate performance and providing tighter noise protection for the encoder.
By Jordi Luque, Fernando L\'opez, Aleix Sant
INDRA is a research platform that integrates multiple archival collections—such as UCSF’s Industry Documents Library, Columbia and CUNY’s ToxicDocs, and Stanford’s SRITA—into a single, LLM‑readable corpus. It employs three safeguards: a closed evidentiary sandbox, real‑time provenance tagging, and a deterministic system‑level protocol to ensure that model outputs are clearly distinguished from archival evidence and from the model’s own inferences. The platform enables large‑language‑model‑powered investigations across these archives while keeping the conditions of knowledge production transparent and auditable.
By Daniel Akselrad, Robert N. Proctor
The paper investigates whether Vision‑Language Models (VLMs) can understand visual persuasiveness by evaluating image‑message pairs that humans consistently judge as persuasive. It introduces Visual Persuasive Factors (VPFs), a taxonomy from cognitive psychology, to quantify visual cues influencing persuasive judgments. Empirical analysis shows VLMs tend to over‑predict persuasiveness, partially reproducing human patterns but often generating false positives, and that VPF‑guided interventions can improve performance only when properly framed.
By Gyuwon Park, Hyounghun Kim
The paper introduces a unified framework that simultaneously measures intrinsic (encoded) and extrinsic (expressed) gender bias in large language models using identical neutral prompts. It finds a consistent link between latent gender information and output bias, but shows that alignment via supervised fine‑tuning reduces expressed bias while leaving internal gender associations largely intact and reactivatable by adversarial prompts. The study also demonstrates that debiasing gains on structured benchmarks may not transfer to realistic tasks such as story generation.
By Nour Bouchouchi, Thibault Laugel, Xavier Renard, Christophe Marsala, Marie-Jeanne Lesot, Marcin Detyniecki
The paper proposes a new method for training large language models to handle long-context reasoning by combining Group Relative Policy Optimization (GRPO) with on‑policy distillation (OPD). It introduces a synthetic multilingual dataset called LongBlocks that tests multi‑hop reasoning, contextual grounding, and long‑form generation. Experiments show that the combined approach outperforms either GRPO or OPD alone while maintaining short‑context performance.
By Miguel Moura Ramos, Duarte M. Alves, Andr\'e F. T. Martins
The paper introduces a densely labeled coastal dataset of over 1,000 images from Oahu, Hawaii, featuring 18 semantic classes and 7,400 annotated instances. Seven modern vision‑language models were evaluated using text‑to‑mask, mask‑to‑mask, and mask‑to‑text alignment experiments. Results show that broad landscape classes are recognized more accurately than conventional object and coastal classes, with coastal concepts posing the greatest challenge; however, performance differences are not solely due to environmental context, and alternative textual labels improve recognition of several coastal concepts.
By Seth Knoop, Chad R. Samuelson, Gabriel R. Slade, Brady Moon, Joshua G. Mangelson
The paper introduces Order‑Aware Slab Multiple Instance Learning (OAS‑MIL), a weakly supervised method that predicts perineural invasion (PNI) risk in intrahepatic cholangiocarcinoma using preoperative T2‑weighted MRI. Tumor‑centered MRI crops are split into ordered 2.5D slabs, and a shared encoder extracts features that are aggregated through both permutation‑invariant set‑attention and bidirectional sequence‑attention branches. In five‑fold patient‑level cross‑validation, OAS‑MIL achieved a mean AUROC of 0.770, outperforming volumetric and other MIL baselines, indicating that preserving axial order improves non‑invasive PNI prediction.
By Hyunsu Go, Youngung Han, Kyeonghun Kim, Jinyong Jun, Junbeom Lee, Dohyun Kweon, Yului Jeong, Suah Park, Sungha Park, Anna Jung, Woo Kyoung Jeong, Ken Ying-Kai Liao, Hyuk-Jae Lee, Nam-Joon Kim
LoopVAE introduces a recurrent depth architecture that reuses a scale‑ and loop‑conditioned core across different spatial scales while keeping resolution‑changing transitions separate. The four‑block core applies 28 block operations per encoder or decoder, enabling a 29M‑parameter convolutional model to achieve 0.28 rFID and 32.54 dB PSNR on ImageNet‑256 with roughly 65% fewer parameters than comparable VAEs. Experiments with both convolutional and Transformer operators, as well as ablations on parameter sharing, demonstrate competitive image quality metrics and reveal how targeted loop interventions and truncation affect reconstruction quality and computational trade‑offs.
By Zhiying Lu
The paper introduces a self‑supervised method for detecting end‑diastole (ED) and end‑systole (ES) in echocardiography by constraining the latent motion to a single‑parameter orbit, effectively modeling cardiac phase as a one‑dimensional signal. This approach yields an interpretable representation that directly identifies ED and ES, improving ED localisation and matching ES performance compared to prior state‑of‑the‑art methods, while using fewer training epochs and a more constrained model. The method is trained on EchoNet‑Dynamic without annotations and the code is publicly available.
By John Bonnici, Matthew Baugh, Aleksandra Kulbaka, Sarah Cechnicka, Bernhard Kainz, Alberto Gomez