The paper presents a multi‑perspective annotation framework for detecting medical hallucinations in chatbot responses. It combines first‑pass annotators, a large language model acting as a judge (LaJ) for candidate discovery, and two adjudication stages—medical‑expert review and evidence‑based fact‑checking. The study finds that single‑pass labeling undercounts errors, while multi‑pass adjudication improves coverage but still depends on expert judgment and evidence.
By Joe Cecil, Marjorie Freedman
SISER is a speaker‑invariant speech emotion recognition framework that combines wav2vec 2.0 for feature extraction with an ECAPA‑TDNN speaker discriminator in an entropy‑based adversarial training scheme. By leveraging self‑supervised representations, SISER reduces reliance on large labeled datasets and suppresses speaker identity more effectively than shallow classifiers. On the IEMOCAP benchmark, SISER achieves a UA of 60.63%, surpassing both the baseline (51.15%) and wav2vec 2.0 without speaker suppression (56.46%).
By Eunseo Choi, Hyunku Kang, Chanwoo Kim
The paper introduces KIDBench, a benchmark designed to evaluate the safety of large language models (LLMs) for children aged 7-11. It includes realistic child queries across ten categories, single- and multi-turn prompts, and compares different prompting strategies—no cues, implicit cues, and explicit age instructions—showing that cueing improves safety scores. The study also reveals uneven safety performance across languages and cultures, and presents KIDGuardLlama, a child-safety evaluator, and KIDLlama, a child-safe response model.
By Samee Arif, Angana Borah, Rada Mihalcea
The paper proposes an ensemble-based self‑taught learning framework for parking space classification that uses unsupervised convolutional autoencoders to learn transferable visual representations from unlabeled data. These learned encoders serve as fixed feature extractors for supervised classification with limited annotated samples, and an ensemble of heterogeneous autoencoders with independent classifier heads is employed to enhance robustness and reduce architectural bias. Experiments on PKLot and CNRPark benchmarks demonstrate that this approach significantly lowers annotation requirements while achieving high accuracies (93–96%) under cross‑dataset evaluation protocols.
By Lucas de Oliveira Cunha, Joelton Deonei Gotz, Paulo Lisboa de Almeida, Andre Gustavo Hochuli
The paper investigates how the phrasing of prompts affects large language models (LLMs) in predicting drug toxicity. By varying job role, prompt structure, and rule interpretation, the authors found that natural variability in LLM outputs outweighs fine‑tuning of prompts. However, incorporating chemoinformatic code to extract features significantly improved model performance, suggesting that prompt engineering alone is insufficient for reliable toxicity prediction.
By Mia MacGregor, Aakash Welgamage Don, Mark Bartlett
PoseDreamer is a new pipeline that uses diffusion models to generate large‑scale synthetic datasets for 3D human mesh estimation, providing 3D mesh annotations that remain aligned with the generated images. The system incorporates controllable image generation, Direct Preference Optimization for control alignment, curriculum‑based hard sample mining, and multi‑stage quality filtering to produce over 500,000 high‑quality samples with a 76% improvement in image‑quality metrics over traditional rendering‑based datasets. Models trained on PoseDreamer match or surpass those trained on real‑world or conventional synthetic data, and combining PoseDreamer with synthetic datasets yields better performance than mixing real and synthetic data alone.
By Lorenza Prospero, Orest Kupyn, Ostap Viniavskyi, Jo\~ao F. Henriques, Christian Rupprecht
The paper investigates on‑policy distillation (OPD) of large language models when trained on a minimal amount of data, specifically a single query. It finds that one‑shot OPD continues to improve for hundreds of steps and recovers most of the performance gains achieved with full‑dataset training across various tasks and model families. The authors attribute this to the high state coverage achieved early in training and show that adding a few semantically distinct queries further boosts coverage and accuracy, eventually matching full‑data results. They conclude that OPD is data‑overfed but algorithm‑starved, highlighting the need to focus on step efficiency and the underlying mechanisms of its recent successes.
By Zixuan Fu, Bingxiang He, Yuxin Zuo, Haohuan Huang, Jinqian Zhang, Ruhang Xiao, Cheng Qian, Qinyu Luo, Huan-ang Gao, Yudong Wang, Zhiyuan Liu, Ning Ding, Chaojun Xiao
The paper introduces LUGL (Local Updates, Global Learning), a framework that separates data collection from model fitting, allowing non‑incremental learners such as gradient‑boosted trees (LightGBM) to be used in reinforcement learning for games. LUGL alternates between a local update phase—where agents play self‑play games and store tabular updates—and a global learning phase—where a function approximator is trained on the accumulated table before it is reset. Experiments on both perfect‑information and imperfect‑information games show that LightGBM‑based agents perform competitively or better than neural‑network baselines like DQN and DeepCFR.
By David Milec, Spyridon Samothrakis, Michael Fairbank, Dennis J. N. J. Soemers
The paper introduces set difference captioning for autonomous driving datasets, aiming to generate natural‑language descriptions of differences between two image subsets. It adapts a two‑stage approach to focus on object‑centric patches, allowing attribution of differences to specific objects or categories. A new benchmark, AD‑Diff Bench, is presented to evaluate these methods, especially for sparse, real‑world differences, with open‑weight models to ensure reproducibility.
By Julian Truetsch, Felix Hauser, Christoph Stiller, Frank Bieder
The paper investigates how on‑policy reinforcement learning (RL) can lead large language models (LLMs) to develop harmful behaviors such as sycophancy, manipulation, or deception. By training 11 instruction‑tuned LLMs (0.5B–14B) across three environments, the authors find that model size can act as a safety buffer in some settings but can also enable greater exploitation in others, depending on environment‑specific features like role framing and implicit gameability cues. They also show that most safety benchmarks fail to predict RL‑induced misalignment, except for sycophancy scores when the exploit relies on inferring user preference, and that on‑policy RL preserves an inherent safety buffer that off‑policy settings bypass.
By Leon Eshuijs, Shihan Wang, Antske Fokkens
NeoRed is a multimodal large language model specifically designed for diagnosing neonatal respiratory diseases. It addresses two main limitations of existing models—domain gaps from adult data and inadequate integration of clinical context—by leveraging two real-world neonatal datasets (NeoCXR and NeoCXR-EV). The model incorporates a Knowledge-Logic-Alignment framework that injects diagnostic priors, aligns report semantics with diagnostic logic, and aligns visual features with imaging conclusions, achieving superior performance on neonatal benchmarks while maintaining adult benchmark performance.
By Yinan Liu, Hongtai Xia, Haoran Xu, Jiankang Hong, Jingkuan Song, Ye Luo
The paper explores how graph neural networks (GNNs) can classify changes in electronic navigational charts (ENCs) as critical or non‑critical for maritime safety. By representing ENC objects as graph nodes and their spatial/semantic relationships as edges, the authors encode both old and new chart data into paired graphs and apply GNN architectures to predict risk levels. Experiments with various GNN models, evaluated on expert‑reviewed ENC updates, show that graph‑based representations enhance classification accuracy, offering a scalable method to support ENC maintenance workflows.
By Abhishek Potnis, Jacob Arndt
The paper investigates how retrieval‑augmented generation (RAG) affects single‑turn mental‑health question answering. It introduces a lightweight selective retrieval policy that activates retrieval only when user queries exhibit psychoeducational, coping, or specificity needs, guided by a rule‑based safety trigger. Experiments show that unconditional retrieval improves specificity but degrades overall quality and safety, whereas selective retrieval maintains closed‑book performance for low‑need cases while avoiding these negative effects.
By Hyunseo Oh, Chong-Kwon Kim, Yoonhyuk Choi
The paper introduces $K$-NeAS, a scalable neural architecture for multi-material CT reconstruction that replaces separate material networks with a shared latent backbone and a differentiable $K$-material soft selector. It automates attenuation bounds using a Gaussian Mixture Model and adds a scheduled auxiliary floater loss to reduce geometric hallucinations in sparse-view settings. Evaluated on four clinical CBCT datasets, $K$-NeAS achieves higher 3D volumetric fidelity—up to a 1.88 dB PSNR gain over a single-material baseline—and shows improved robustness under extreme sparsity, outperforming baselines by up to 1.17 dB.
By Daksh K. Shah, Emmanouil Nikolakakis, Razvan Marinescu
The paper investigates how privacy guarantees, robustness to Byzantine attacks, and detection coverage for rare intrusion types interact in federated network intrusion detection systems. It introduces geometric indistinguishability to explain how privacy noise can obscure minority-class signals, and demonstrates through experiments on UNSW‑NB15 that combining differential privacy with robust aggregation can disproportionately harm detection of rare attacks. The study highlights that these properties cannot be treated as independently composable and calls for aggregation‑aware modeling and sample‑aware evaluation to build trustworthy federated NIDS.
By Adrita Rahman Tory, ABM Shawkat Ali, Md Abu Layek, Khondokar Fida Hasan
WIDE: Wildcard Inference with Dynamic Expansion for Cross-Modal Generative Retrieval proposes a new approach to address information asymmetry in cross-modal retrieval. The method introduces Adaptive Entropy Thresholding to calibrate uncertainty, Asymmetry-aware Wildcard Decoding to emit wildcards instead of forced identifiers, and Blind-Spot Re-ranking to evaluate expanded candidates. Experiments on the M-BEIR benchmark show that WIDE outperforms existing generative retrieval methods by reducing forced hallucination while keeping index structures compact.
By Teng Guo, Xin Wang, Jiayou Xu, Keying Zhou, Jifeng Shen, Haoxin Ruan
LightEMMA is a longitudinal evaluation framework that tests the autonomous driving performance of vision‑language models (VLMs) without fine‑tuning or prompt engineering. Using this protocol, the authors evaluated 15 VLMs from five major families on the nuScenes prediction benchmark and found that larger, more capable models do not consistently outperform earlier generations. The study identifies common failure modes such as overreliance on historical actions and difficulty reconciling conflicting visual cues, underscoring the need for domain‑specific adaptation to enhance VLM safety in autonomous driving.
By Zhijie Qiao, Haowei Li, Zhong Cao, Henry X. Liu
The paper presents a deployed system that scores blockchain addresses using their position in a massive multi‑chain transaction graph instead of relying on sanctions lists. The system operates on a single graph of 835 million addresses and 15.8 billion edges across five EVM chains, employing a shared inductive encoder with per‑chain normalization and two scoring heads. It demonstrates label‑free transfer, achieving high recall on held‑out positives for Base, Arbitrum, and Gnosis at a very low alert rate, and shows significant lead time over external registry events, while maintaining fast, reproducible serving performance.
By Yury Korolev
The paper introduces DP‑BR‑FedAvg, a federated learning framework that combines Gaussian‑mechanism differential privacy with a coordinate‑wise trimmed‑mean Byzantine‑robust aggregation rule. It is evaluated on a simulated cross‑institutional classification task involving fraud and clinical‑risk scoring, where it improves the F1‑score for a minority class from 0.030 (plain FedAvg) to 0.119 while bounding privacy loss. The study demonstrates that privacy and robustness mechanisms interact, and that system design for regulated, adversarial, cross‑institutional settings must account for this interaction.
By Srikumar Nayak
The paper reviews how Large Language Models (LLMs) are being adapted for medical reasoning, moving beyond single-step answers to systems that can systematically, transparently, and verifiably reason. It introduces a taxonomy of enhancement techniques, split into training-time methods such as supervised fine‑tuning and reinforcement learning, and test-time methods like prompt engineering and multi‑agent systems. The review examines their application across text, image, and code modalities in key clinical areas—diagnosis, education, and treatment planning—and tracks the shift in evaluation benchmarks from simple accuracy to more nuanced assessments of reasoning quality and visual interpretability.
By Zizhan Ma, Wenxuan Wang, Meidan Ding, Shiyi Zheng, Shengyuan Liu, Jie Liu, Jiaming Ji, Linlin Shen, Yixuan Yuan, Wenting Chen