The paper introduces GIST, a method for tabular foundation models that leverages intermediate generator states to inform supervision. By averaging query losses over conditional label distributions and using entropy‑calibrated encodings, GIST provides input‑consistent targets that preserve the population objective and avoid increased gradient covariance. Experiments show that GIST improves mean rank on TabArena and TALENT across three architectures, winning 62.1% of dataset‑level comparisons against hard‑label baselines.
By Zhangyang Peng, Jingze Shi, Tongyi Zhang, Yuyu Luo
The paper introduces Evidential Multimodal Transformers with Dempster-Shafer Fusion (EMT-DS) for emotion recognition, which models predictions as Dirichlet distributions to separate epistemic and aleatoric uncertainty. EMT-DS fuses modality-specific belief masses using Dempster's Rule, explicitly quantifying conflict and producing more diffuse predictions when evidence is contradictory. Across four benchmarks, EMT-DS outperforms baselines, reduces calibration error, remains robust to noise, and its conflict metric can flag misclassifications for selective prediction.
By Muhammad Usama, Yeoreum Jung
The paper introduces Self‑Sparse Speculative Rollouts (S$^3$R), a method that uses sparse attention to propose tokens and then verifies them with full attention, thereby reducing KV‑cache traffic while preserving the dense‑policy distribution for reinforcement learning rollouts. S$^3$R eliminates the need for a separate drafter by sharing policy parameters between drafting and verification, and a KV‑aware scheduler limits cache thrashing to maintain parallelism. Experiments on long‑context mathematical reasoning with the Qwen3 model show that S$^3$R achieves 1.35–1.39× higher end‑to‑end training throughput with comparable task performance.
By Yuhang He, Hongyu Ge, Zihan Lin, Siyi Liu, Yongqi Zhang
Eco‑RNA is an 80‑million‑parameter RNA foundation model that introduces adjustable inference depth via learned iterative refinement. The shared, input‑conditioned Transformer core repeatedly refines sequence representations, with the number of iterations chosen after training to control effective depth. On the RNAGym fitness benchmark, increasing iterations from one to four improves Spearman correlation from 0.242 to 0.286, and at four iterations Eco‑RNA outperforms the 40‑billion‑parameter Evo 2 model across all aggregate fitness metrics, while also excelling in non‑coding RNA classification, secondary‑structure prediction, and chemical probing profile prediction.
By Pengan Chen, Liang Hong, Xinyuan Liu, Han Huang, Ziqian Lin, Jiayang Chen, Yimin Fan, Jiyue Jiang, Yu Li
SBNO (Schr"odinger Bridge Neural Operator) is an equation‑free framework that learns conditional operators for both forward and inverse PDE problems using paired data and a diffusion Schr"odinger bridge matching approach. It preserves empirical input‑solution pairing through a mask‑aware endpoint coupling and supports both task directions within a single model. At inference, SBNO generates deterministic reconstructions via a probability‑flow ODE, achieving comparable or superior accuracy on five PDE benchmarks with only 5% observations, while being at least 9.7× faster and using 4.3× less GPU memory than guided‑diffusion baselines.
By Jaehyeon Park, Inhyeok Jeong, Noseong Park
PANDA (Prior-guided Attentional Dual-path Architecture) is a plug‑in for two‑stage message‑passing Hypergraph Neural Networks that blends learned receiver‑normalised attention with structural incidence priors while also providing a non‑competitive auxiliary path for sender contributions. It generalises a complex‑valued convolution operator used in earlier work and, when applied to four different HNNs across spatial and spectral methods, improves performance by an average of 5.18 percentage points on eight real‑world benchmark datasets.
By Rishabh Jain, Federico Belotti, Stefano Coniglio, Pietro Lio, Stefano Fiorini
TabMF is a framework that repurposes pretrained tabular foundation models (TFMs) as encoders for supervised datasets, adding a dataset-level readout and fine‑tuning via contrastive post‑training on synthetic views. The authors instantiate TabMF on TabICLv2 and TabPFN‑2.5 and evaluate the resulting representations on six tasks involving unseen real datasets. The approach achieves the best aggregate performance among both learned and classical dataset representations, excelling in dataset retrieval, compact benchmark construction, and distribution‑shift severity assessment.
By Rafiq Kamel, Nicolas Atienza, Aziz Fouch\'e, Nicola Cancedda, Josif Grabocka
The paper critiques the current evaluation practices in generative molecular dynamics (GenMD), arguing that they allow models to appear successful while potentially being incorrect on key claims. It attributes this issue to the high computational cost of obtaining exhaustive ground truth, which forces the community to rely on convenient proxies that become entrenched. The authors propose new observables, reporting standards, and benchmark designs—both computational and experimental—to align evaluation with GenMD’s scientific goals.
By Simon Olsson, Frank No\'e, Grant M. Rotskoff, Kresten Lindorff-Larsen
The paper introduces a blind and verifiable protocol for sampling from quantum generative models, enabling secure inference on a remote quantum server. It builds on Universal Blind Quantum Computing for blindness and uses trap-based techniques for verifiability, also estimating the proportion of corrupted samples. Additionally, the authors extend their earlier work on verifiable delegated variational quantum algorithms to cover the training phase of such models.
By Matteo Inajetovic, Anna Pappa, Petros Wallden
The paper introduces a machine‑learning approach that incorporates the ground‑state electron density as input while enforcing it to be the minimum of a surrogate functional derived from orbital‑free density functional theory. This method outperforms state‑of‑the‑art models on six of seven extensive properties in the QM9 dataset and shows better generalization to larger molecules in the QM40 dataset. Models that use the density without enforcing the surrogate functional constraints perform worse on all extensive properties of both datasets.
By Manuel V. Klockow, Simon Wagner, Roman Remme, Dominic Plein, Fred A. Hamprecht
The paper investigates whether structural diversity among quantum circuit designs can enhance predictive performance through model fusion. Using Combinatorial Fusion Analysis (CFA), which combines rank-score functions and cognitive diversity, the authors fuse 29 diverse classical models and quantum circuits. Experiments on four binary classification benchmarks show that CFA-based hybrid classical‑quantum fusion consistently outperforms the best individual model, achieving accuracy gains up to 6.03% and surpassing related literature on all datasets.
By Jingyan Xu, Yuanhong Wu, Samuel Yen-Chi Chen, Sean Patten, Christina Schweikert, D. Frank Hsu
The paper introduces the Constrained Triple-Head Price Optimization (C3PO) network, a transformer-based model designed for real‑time bi‑level decision making in discrete‑choice settings. C3PO integrates price imitation learning, multi‑task revenue prediction, and in‑context elasticity learning, while a lightweight differentiable constraint layer enforces business pricing rules. Experiments on simulated, synthetic, and real‑world datasets show that C3PO outperforms calibrated MNL baselines by up to 7.9 percentage points in expected revenue and reduces constraint violations by over 50%, with significant improvements in healthcare tender and airline ancillary pricing applications.
By Shivaram Subramanian, Zhengliang Xue, Markus Ettl, Yingdong Lu, Jayant Kalagnanam
The paper evaluates TabPFN, a tabular foundation model, for motor insurance pricing by comparing it to generalized linear models (GLMs) and XGBoost on two motor third‑party liability datasets. Results show that TabPFN does not consistently outperform GLMs: it lags on claim frequency, especially when the in‑context training set is small, but performs competitively on claim severity where data are limited. Additionally, TabPFN’s inference times are substantially longer than those of the baseline methods.
By Bruno Deprez, Wouter Verbeke, Tim Verdonck
The Cross-View Attention Fusion Network (CVAF‑Net) is a dual‑view deep learning model that estimates cardiac output from short photoplethysmography (PPG) segments by combining raw PPG signals with a structured prior‑guided feature sequence map through cross‑view attention. Evaluated on simulated, vasoconstriction, and real‑world activity datasets, CVAF‑Net achieved a mean absolute error of 0.19 L/min on simulated data and as low as 1.20 L/min in real settings, outperforming most benchmarks while reducing computational cost by twelvefold compared to a leading Transformer model. Plausibility checks confirmed physiologically consistent correlations with age, heart rate, and systemic vascular resistance.
By Yaowen Zhang, Bo Cui, Libera Fresiello, Peter H. Veltink, Dirk W. Donker, Ying Wang
The paper introduces T‑Mem, a memory framework that uses temporal structure to manage long‑horizon multi‑party conversations. It employs Temporal Evidence Discovery to find locally concentrated evidence, Temporal Memory Structuring to model state transitions and compress low‑information spans, and Evidence Forest Linearization to serialize the evidence for answer generation. Experiments on two benchmarks show that T‑Mem outperforms existing memory frameworks across various settings.
By Junyou Guo, Chongyang Tao, Yan He, Yong Wang, Shuai Ma
MoPrO‑SD is a stance‑specific prompt optimization framework that breaks prompts into semantic modules and refines them sequentially with LLM feedback. It evaluates three refinement strategies—instruction‑only, instruction with reasoning, and instruction with per‑example reasoning—using class‑conditioned errors to improve stance detection. The authors also release VaccineStance, a new human‑annotated benchmark for vaccine stance detection, and demonstrate that MoPrO‑SD outperforms human‑written base prompts across multiple datasets and open‑weight LLMs.
By Meysam Sabbaghan, Nikesh Gyawali, Arman Zareian Jahromi, Sebastian Loera-Lozano, Doina Caragea, Cornelia Caragea
MVGraph is a new benchmark designed to evaluate large vision‑language models on multi‑graph reasoning tasks where evidence is spread across multiple interrelated visual graph views. It contains 8,000 instances covering eight graph‑theoretic tasks derived from partial and temporal graph scenarios, and each instance is provably unsolvable from a single view. The benchmark also offers 29 diagnostic tasks with 64,000 instances to probe four core capabilities—Perception, Alignment, Composition, and Reasoning—providing golden intermediate representations for detailed analysis.
By Yingjie Zhu, Han Cheng, Xuefeng Bai, Youcheng Pan, Yang Xiang, Kehai Chen
APEX-Harm is a benchmark that tests prompt and script injection attacks on long‑horizon agents working in large, shared professional workspaces. The benchmark covers finance, consulting, and legal tasks, each involving 115–328 files and multiple tools such as email, calendar, chat, documents, spreadsheets, PDFs, and code execution. It evaluates six attack categories and measures attack success, exposure, and task performance across 12 frontier models, revealing that many agents can complete tasks while still executing unauthorized actions.
By Ziyi Yang, Adel Bibi, Guohao Li, Bernard Ghanem, Carlos Hinojosa
The paper introduces ARTIS, a framework that separates exploration from commitment by using iterative simulation to test large language model actions before real-world execution. This approach extends inference-time computation to boost action-level reliability and robustness while avoiding environmental risk. The authors also highlight that standard LLM simulators miss rare, high-impact failures, and propose a risk‑aware simulator that focuses on fidelity for failure‑inducing actions, showing significant reliability gains across agentic benchmarks.
By Xingshan Zeng, Lingzhi Wang, Weiwen Liu, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Qun Liu
CaptchaAgent is a structured visual reasoning framework designed to solve modern reasoning CAPTCHAs. It constructs a visual intermediate representation that records entities, attributes, and relations, then refines ambiguous regions and verifies counterfactual consistency before generating actions. The approach yields significant performance gains over baseline browser agents on Open CaptchaWorld and the Oedipus benchmark, highlighting key challenges such as exact counting, geometry, and ordered execution.
By Shangquan Sun, Ruoyu Chen, Zhirui Liu, Xiang Liu, Chenghao Qian, Wenqi Ren, Xiaochun Cao