The paper introduces SIGMA, an end‑to‑end framework that couples density estimation with principled uncertainty quantification for generative models. By using Gaussian mixture models and treating uncertainties as nuisance parameters in a new goodness‑of‑fit test, SIGMA converts a standard point‑null test into a composite test that accounts for statistical error of the fit. The authors demonstrate that incorporating uncertainties allows the same model to be trusted on test sets up to an order of magnitude larger than with conventional methods, using both a toy two‑dimensional example and a collider physics benchmark.
By Giada Badaracco, Sean Benevedes, Christina Reissel, Philip Harris, Gaia Grosso
Graph Representation Learning via Adaptive Expectile Aggregation introduces ExpectileGCN, a graph neural network framework that learns an expectile level to adaptively aggregate neighborhood information beyond simple mean pooling. The method smoothly interpolates between minimum- and maximum-type behaviors, with mean aggregation as a special case, and employs a differentiable sigmoid‑based approximation for end‑to‑end training. The authors provide theoretical stability and generalization guarantees, analyze classification risk under heterophilic settings, and demonstrate improved performance and robustness on both homophilic and heterophilic benchmark graphs.
By Tianyi Zhu, Wei Lan
ResCom introduces a low‑rank correction that leverages activation rounding residuals to compensate for layer‑output errors in low‑bit weight–activation quantization of large language models. The method selects residual directions based on their impact on output reconstruction and computes the correction in closed form, avoiding gradient training or weight updates. Experiments on Llama‑3‑8B with 4‑bit weights and activations show significant gains: perplexity drops from 7.708 to 7.147, MMLU improves by 2.89 percentage points, and five‑shot GSM8K accuracy rises from 23.96 % to 36.16 %.
By Ning Yang, Bohan Cui, Zhuowen Liu, Yuting Wang
AgentBaton explores how capable and efficient language models can collaborate on long‑horizon agentic tasks by handoffs, where one model takes over from another. The study finds that handoffs from a capable to an efficient model at appropriate stages preserve or improve accuracy while significantly cutting cost, whereas the reverse handoff is less effective. AgentBaton, a lightweight detector, learns the optimal handoff point and achieves an average API cost reduction of 65.9% across 11 model pairs while largely maintaining task accuracy, even gaining up to 8.8 percentage points over capable‑only execution.
By Junlong Li, Jindong Jiang, Yonggan Fu, Jiarui Yao, Hanrong Ye, Wonmin Byeon, Peter Belcak, Hao Zhang, Pavlo Molchanov
The study shows that in memory‑agent benchmarks, simply retrieving the correct fact does not guarantee accurate answers. On the FactConsolidation task, increasing the evidence budget from 64 to 512 tokens caused accuracy to drop, even though the newest statement was present in every retrieved set. Experiments with Qwen models revealed that replacing an unrelated statement with an obsolete competing statement drastically reduced accuracy, indicating that outdated information can be more harmful than irrelevant evidence.
By Oleeviya Babu Poikarayil, Tony Basil Babu, Abdulrahman Nahhas, Het Darshan Mehta, Jawid Ahmad Baktash, Ernesto William De Luca, Klaus Turowski
The paper introduces Foreclassing, a new machine learning problem where the class label of a time series is defined by its future observations. It provides a mathematical formulation, theoretical results on how future information influences Bayes risk and classification error, and proposes ForeClassNet, a neural network that jointly forecasts and classifies while accounting for uncertainty. Experiments across weather, energy, and finance datasets show that Foreclassing can improve classification accuracy, probabilistic predictions, and downstream decision-making compared to conventional methods.
By Daniel Andrew Coulson, Martin T. Wells
PREEN (Preset Regression via an Exemplar-based Editor Network) performs color style transfer by predicting a 29‑parameter preset for a deterministic image editor, rather than generating a raster image. It uses a style‑transfer teacher to create pseudo‑presets, then trains a CLIP‑conditioned regressor refined with LoRA and a CIELAB loss. On the PST50 benchmark, PREEN outperforms baselines in pixel‑ and perceptual‑fidelity, runs in sub‑second on‑device latency, and offers a single inference‑time strength knob that smoothly transitions from default to vivid presets.
By Beomjin Ahn
ScIntBench is a new benchmark for evaluating text‑conditioned 3D indoor scene generation, focusing on interaction readiness rather than just static physical plausibility. It runs prescribed manipulation actions—such as pushing, opening, and probing—in a physics simulation and records metrics like task success, capability, progress, post‑interaction integrity, and dynamic plausibility, supplemented by multi‑view semantic and static physical checks. Across 12 state‑of‑the‑art systems and 10 room types, the study finds that high static plausibility does not guarantee interaction readiness, with the best system succeeding in only 36.92% of tasks and many systems unable to support any interaction at all.
By Wendi Ying, Daochang Liu, Leonid Sigal, Ajmal Saeed Mian
The paper introduces Policy-Compatible Reinforcement Learning (PICO), a method that addresses the learning cliff in reinforcement learning with verifiable rewards (RLVR) by performing distribution-constrained joint decoding. PICO guides token-level decoding with expert data while preserving the model’s token probabilities, and uses a dynamic top‑k mechanism to keep sampling within a finite divergence budget. The resulting trajectories are verified and policy‑compatible, and are jointly optimized with on‑policy rollouts using a mixed‑behavior objective that mitigates distribution mismatch via importance sampling, leading to consistent performance gains over standard RLVR and prior off‑policy baselines on diverse reasoning datasets and agentic benchmarks.
By Zehao Wang, Hanzhu Chen, Runquan Gui, Zhe Wang, Lin Yang
Quaternion Approximate Networks (QUAN) is a deep learning framework that uses quaternion algebra to achieve rotation equivariant image classification and object detection. QUAN approximates quaternion convolution via Hamilton product decomposition with real-valued operations, enabling efficient CUDA implementation and preserving geometric properties. The method introduces Independent Quaternion Batch Normalization and extends quaternion operations to spatial attention, achieving higher accuracy, faster convergence, and improved parameter efficiency on CIFAR, ImageNet, COCO, and DOTA benchmarks compared to conventional CNNs and other quaternion models.
By Bryce Grant, Peng Wang
SetGRPO is a new method for calibrating grader‑induced policy optimization in rubric‑based reinforcement learning. It introduces three modules—Joint Reward Set Calibration, Geometry‑Guided Safe Recovery, and Selective Risk‑Gated Optimization—to construct uncertainty sets for training coefficients, map them into safe update geometry, and control direction‑reversal risk. Experiments on ten open‑ended benchmarks show that SetGRPO outperforms other training methods on both Qwen3.5 and Llama3.2 models.
By Jiale Huang, Zixu Li, Zhiwei Chen, Zhiheng Fu, Jinhe Bi, Yupeng Hu
The study evaluates how well large language models (LLMs) understand geographic lexical variation in Spanish across 21 countries, using an expert-curated database of 934 terms. Ten LLMs were tested with yes-no and multiple-choice questions, generating 10,164 and 16,637 items respectively, and performance was measured with F1 and adjusted Jaccard scores. Results show notable differences by model, country, and question format, with Spain scoring highest and Chile lowest, and suggest that digital resource availability and lexical distance from Spain influence performance.
By Yoshifumi Kawasaki
The paper introduces a spatial-frequency learning framework for enhancing Scanning Transmission Electron Microscopy (STEM) images. It employs a spatial-frequency fusion network that separately enhances features in the spatial and frequency domains before fusing them, leveraging periodic atomic arrangements in the frequency domain and local atomic details in the spatial domain. Additionally, the authors present a STEM noise calibration method to synthesize realistic images, creating a diverse dataset that improves performance on synthetic and real advanced material images, including better atomic localization accuracy.
By Hesong Li, Ying Fu, Ziqi Wu, Ruiwen Shao, Tao Zhang
3D-SceneQ is a new 3D large language model that tackles two key limitations of existing 3D scene understanding pipelines: the inclusion of irrelevant object tokens that add noise, and the exclusive use of point cloud data that lacks rich semantic cues. It introduces a Query‑Guided Adaptive Pruning (QGAP) module that selectively keeps task‑relevant objects based on user instructions, and a Multi‑modal Object‑level Feature Enhancement (MOFE) module that fuses 2D feature embeddings into the 3D representation to enrich semantic information. Experiments show that 3D‑SceneQ surpasses state‑of‑the‑art methods on several benchmarks, achieving significant gains in metrics such as CIDEr and EM@1.
By Yongqi Shan, Lu Zhang, Jiazuo Yu, Yunzhi Zhuge, Huchuan Lu
MedJev is an open System One model that answers medical questions in a single forward pass without generating text, enabling fast, local clinical decision support. Across ten medical benchmarks, MedJev achieves competitive accuracies—72.5 % on MedQA, 75.4 % on USMLE, and 80.1 % on NEJM—while maintaining well‑calibrated confidence estimates. On patient‑level tasks, it attains an AUROC of 0.926 for in‑hospital mortality, surpassing XGBoost, and introduces a lightweight two‑parameter recalibration method that reduces miscalibration from 0.286 to 0.007 using only 50–200 labeled examples.
By Simon A. Lee, Jiahang Sha, Moonseong Jeong, Yunbei Pan, Jeffrey N. Chiang
PDRVLA is a Vision‑Language‑Action model that incorporates a lightweight Physical Dynamics Module to extract physical tokens from vision‑language features, enabling action generation without online video synthesis. The module predicts future RGB, depth, and optical‑flow features, guided by auxiliary supervision from a frozen pretrained video world model. Trained solely on publicly available data augmented with human‑to‑robot synthesis, subtask annotation, and embodied chain‑of‑thought (ECoT) annotation, PDRVLA achieves high success rates across four simulation benchmarks, with ablations demonstrating the benefits of world‑model supervision and various conditioning strategies.
By Hanqing Wang, Zhaotian Cai, Ruichuan An, Bohan Zeng, Tao Zhang, Yilin Chen, Xu Jiang, Kaiquan Chen, Chongyang Wu, Haoran Li, Fan Yang, Mingan Lin, Wentao Zhang, Zenan Zhou, Youzheng Wu
The paper introduces a model‑based boosting method for fitting structured additive distributional copula regression models with bivariate responses, specifically targeting joint modeling of insurance claim frequency and severity. It builds on generalized additive models for location, scale, and shape (GAMLSS), allowing multiple distribution parameters to depend on covariates and thus providing interpretable additive predictors. The authors evaluate the framework on synthetic and real data, showing it can recover covariate effects, capture dependence structure, identify informative variables, and match benchmark predictive performance while offering transparent interpretation.
By Banghee So, Na Zhang
The paper investigates whether discrete visual tokens can preserve semantic information as effectively as continuous representations. Contrary to the common belief that quantization error inevitably causes semantic loss, the authors show that larger quantization errors can still yield better multimodal understanding, and that the effective bit budget of discrete indices is the critical factor. They propose Factorized Vector Quantization, which uses parallel codebooks to encode complementary views of visual features, achieving performance that matches continuous representations and surpasses prior discrete models on a broad set of metrics.
By Jialv Zou, Xiaotong Li, Zhengrong Yue, Xu Huang, Yaxin Luo, Wenhao Yan, Haodong Li, Wenyu Liu, Xinggang Wang
PIVOT (Plan-Inspect-eVOlve Trajectories) is a self‑supervised framework that refines large‑language‑model agent plans by treating trajectories as optimizable objects. It operates in four stages—PLAN, INSPECT, EVOLVE, and VERIFY—using structured losses and textual gradients to iteratively improve trajectories and ensure compliance with task constraints. Experiments on DeepPlanning and GAIA show that PIVOT achieves up to 94% relative improvement in constraint satisfaction with human‑in‑the‑loop feedback and still delivers significant gains autonomously, while using 3x to 5x fewer tokens than competing methods.
By Tuo Zhang, Alin-Ionut Popa, Yan Xu, Rui Song, Dimitrios Dimitriadis
The paper "HateSense: Hierarchical Sentiment Analysis Framework for Hate Speech Detection: Implementing Binary and Multiclass Classification Strategy" revisits a transformer-based framework that combines hierarchical classification and sentiment features for hate speech detection. Experiments on the Davidson and HateXplain datasets show that a dominant-class-first hierarchy actually lowers macro‑F1, sentiment features have minimal impact, and class‑weighted training yields the largest performance gains. Cross‑dataset transfer and slur reuse reveal that in‑domain metrics overstate real‑world reliability.
By Faria Naznin, Md Touhidur Rahman, Shahran Rahman Alve