The paper introduces PAIR, an EEG benchmark that records brain activity from 20 participants while they watch naturalistic video clips and immediately recall them. The dataset allows direct comparison between perception and recall of the same content, and the authors evaluate several decoding models on tasks ranging from fine‑grained semantic classification to meta‑feature decoding (motion, color, object number, face, human presence) and a state‑decoding task distinguishing perception from recall. Results show comparable performance across many model–task pairs, with slightly higher perception accuracy in some coarse‑grained comparisons, and strong separability between perception and recall at the state level, even after removing alpha‑band activity.
By Zikai Fang, Tian-Yi Zhou, Xuan-Hao Liu, Yixi Li, Xiaoyu Xing, Yutong Pei, Bao-Liang Lu, Wei-Long Zheng
TextVBench is a new benchmark and evaluation framework for assessing how well video generation models render and preserve text. It introduces two tracks—TextVBench‑T2V for prompt‑specified text creation and maintenance, and TextVBench‑Ref2V for preserving text from reference images—using a multi‑target, multi‑instance spatiotemporal correspondence approach. The framework employs target‑conditioned candidate selection, supports multiple OCR regions per target, and separates spatial and temporal fidelity, summarizing results with a length‑weighted harmonic score and detailed diagnostics.
By Ruiqiang Zhang, Hengyi Wang, Chang Liu, Guanjie Wang, Zehua Ma, Weiming Zhang
The paper introduces a real‑world CAD segmentation benchmark focused on fastener‑relevant hole detection, a task that demands human annotation and presents higher geometric complexity than typical synthetic datasets. It presents two rule‑based methods—a graph‑wire‑search and a multi‑view approach—and trains learning‑based models (UV‑Net and AAGNet) on a newly annotated engineering CAD dataset. Extensive experiments evaluate parameter tuning, cross‑domain generalisation, and performance across dataset sizes and IoU thresholds, revealing that rule‑based methods generalise without training data while learning‑based methods excel with sufficient annotations but are less robust to domain shift.
"whyItMatters":"The study provides a realistic benchmark and actionable insights into the trade‑offs between data requirements and robustness for CAD segmentation, informing future research and practical applications in engineering tasks."
By Kai Sch\"utte, Edgar Heinert, Matthias Rottmann
The paper introduces an audit layer for vision‑language model benchmarks that reuses per‑question artifacts from evaluation runs to assess the benchmark itself. Six detectors analyze saved outputs, and an additional text‑only rerun removes visual inputs to produce per‑question reports and aggregate statistics, including permutation‑null excess and interval estimates. Applying this method to the MMStar benchmark reveals that 20.2% of its questions show no accuracy gain from images, with a small but significant excess in mathematics and science categories, and identifies questions where all models chose the same non‑gold option.
By Denis Degtiarev, Anna Gromova, Sergey Katkov, Artem Mikhailin, Safina Mukhamedshina, Albert Nasybullin, Karina Shaikhutdinova, Artem Tuzov
CUBE introduces a class‑conditioned evidence framework for open‑set insect species recognition, modeling unknownness around each known species rather than as a single global unknown class. By evaluating test samples relative to their most similar known species, CUBE enables local rejection of visually confusable unknown insects and incorporates auxiliary species as shadow references and taxonomic candidate calibration. The extended CUBE‑E variant adds a lightweight location‑consistency term when ecological metadata is available, achieving state‑of‑the‑art AUROC scores on the Open‑Insect C‑America benchmark while preserving high closed‑set accuracy.
By Hao Tan, Qimin Zhang, Zhe Xiao, Zekai Zuo, Wanting Liu, Jinghui Qin
The paper presents a low‑cost, self‑supervised method for detecting wheel tracks on snow‑covered roads using synchronized RGB and long‑wave infrared (RGB‑I) imagery. Training labels are automatically generated by projecting future wheel trajectories from GNSS data, and a neural network is trained to identify wheel tracks on featureless, white‑out road surfaces. Experimental results show the approach is effective in visually monotonous adverse environments, providing a robust alternative for winter autonomy. The ThermalTrack dataset is released for further research.
By Yiming Yang, Jeremy P. Bos
WaveMark introduces a wavelet‑domain watermarking technique that embeds messages into the initial noise of diffusion models, distributing keyed replicas across spatial positions, wavelet orientations, and scales to enhance robustness. The method encodes each carrier with a randomized bit via magnitude projection of four latent channel coefficients, yielding a clear decoding margin. Experiments on SD2.1 with SDP and MS‑COCO Captions show perfect bit accuracy without attacks and over 88% accuracy against state‑of‑the‑art removal attacks, outperforming existing baselines while preserving image quality.
By Jie Cao, Qi Li, Chaklam Cheong, Yutong Jin, Enze Yu, Zelin Zhang, Jianbing Ni
The paper evaluates multimodal decision models, specifically CLEF and CLEF‑Flash, on 2,818 breast ultrasound images from three datasets (Dataset B, BUSI, BUS‑BRA). Using three different prompts—minimal labels, explicit classification, and sonographic criteria—the study reports accuracy, macro F1, balanced accuracy, precision, throughput, and latency. CLEF‑Flash with a clinical prompt achieves the highest accuracy across all datasets, while CLEF with a task‑guided prompt yields the best macro F1 and balanced accuracy on BUSI; however, all configurations underperform a majority‑class predictor on BUS‑BRA, highlighting frequent benign‑to‑malignant misclassifications.
By Tran Cao Minh, Hai Tran, Ha Minh Tan
CLIDD (Cross-Layer Independent Deformable Description) is a local feature representation method that samples directly from independent feature hierarchies using learnable offsets, avoiding the construction of a dense feature map and thus reducing memory and computation overhead. The approach breaks interpolation-induced correlation through per-keypoint per-layer offsets, preserving discriminability even at high keypoint densities. A hardware-aware kernel fusion strategy and a scalable training framework that combine metric learning and knowledge distillation enable real-time performance, achieving superior matching accuracy and exceeding 200 FPS on edge devices.
By Haodi Yao, Fenghua He, Ning Hao, Yao Su
CoRef-DETR is a compact detection transformer designed for underwater instance segmentation, addressing errors from context, boundaries, and categories. It uses a decoupled global context detector with 33.4 M parameters, a lightweight boundary refiner of 47.7 k parameters, and a crop‑probe category branch that revises up to three mask categories per image. The method achieves state‑of‑the‑art mask AP on multiple UIIS datasets and runs in 26–37 ms on an NVIDIA T4 GPU, outperforming larger models.
By M. Fazri Nizar, Julian Supardi, Muhammad Naufal Rachmatullah
VaME (Variational Multimodal Embeddings) introduces a framework that treats latent reasoning as a learnable distribution over trajectories, enabling autoregressive exploration in latent space guided by answer reconstruction. By augmenting embedding-token readout with a latent-fused embedding and optimizing over stochastic variational trajectories via reinforcement learning with a Semantic Decoding Reward, VaME achieves superior performance on the MMEB-V2 benchmark and reasoning-intensive tasks like MRMR. The method also delivers at least a 4.25× inference speedup compared to deterministic latent autoregressive baselines.
By Peixi Wu, Mingzhou Jiang, Feipeng Ma, Biao Yang, Yunhao Zhou, Wei Yuan, Bosong Chai, Huizu Lin, Jie Chen, Zhangchi Hu, Fan Yang, Wenwu Ou, Hebei Li, Xiaoyan Sun
SLICE: Sequential Likelihood-Free Inference via Conformal E-Values introduces a method for constructing exact anytime-valid confidence sequences for parameters of scientific models that can be simulated but have intractable likelihoods. The approach uses fresh simulations at each observation to conformalize a positive predictable score into a conditional e-value, whose product forms a test martingale that is inverted over the parameter space. The method is shown to be valid for any score, with the oracle likelihood-ratio score maximizing expected log growth, and is demonstrated on synthetic benchmarks, an audit of model routing in an LLM API, and microlensing data, where the resulting confidence sequences remain valid under continuous monitoring while neural scores remain competitive with exact likelihood ratios and application-specific scores.
By Stefano Cortinovis, Laura Battaglia, Fran\c{c}ois Caron
The study evaluates the robustness of deep learning models for DNA regulatory sequence prediction under biologically and technically induced distribution shifts. Using a mechanistic simulation benchmark and a massively parallel reporter assay dataset, the authors show that while models perform well under mild GC content changes, they exhibit higher error, miscalibration, and coverage collapse when faced with motif rewiring, noise, and batch effects. Incorporating biological priors and uncertainty-aware selective prediction improves robustness, though protection against strong assay noise remains limited.
By Yiyao Yang
SpireBench evaluates long‑horizon strategic play by having large language models play the full commercial deck‑builder Slay the Spire, a game where early choices determine long‑term success. Across 20 LLMs at the highest difficulty, frontier models beat heuristic baselines and match average human performance but lag far behind experts, especially as runs lengthen. The benchmark reveals that strong models excel in early fights but misjudge risk over a run, and that their performance depends partly on prior game knowledge rather than on memories written during play.
By Scott C. Lowe, Sander W. Keemink
The paper introduces Representation Geometry Regularization (RGR) to address capability degradation observed in On-Policy Self-distillation (OPSD) for large language models. OPSD can cause previously correct responses to become incorrect even as overall performance rises, a problem linked to deteriorating hidden representation geometry—specifically, less compact intra-cluster and weaker inter-cluster separation. RGR preserves this geometry by encouraging cluster compactness and separation during post‑training, and can be added to existing OPSD methods without changing their core optimization, with experiments showing it mitigates degradation while maintaining improvements.
By Jiaqi Tian, Ting Long
EmbodiedCLBench is the first benchmark designed to evaluate continual learning in embodied simulation environments, focusing on self‑evolving agents. The benchmark uses compositional generalization to test whether agents can learn basic tasks and then recombine those skills to solve more complex, unseen tasks. Experiments across various agent harnesses and learning methods reveal that while basic task experience improves performance on advanced tasks, this benefit does not scale with more learning samples and current agents fail to extract additional useful information.
By Jie Huang, Yanan Chen, Ruixun Liu, Kaichen He, Xinyi He, Rui Huang, Zilong Zheng, Zhenliang Zhang, Yiwu Zhong
DRYLAB‑Bench is a new benchmark that measures the biosecurity risk of large language models (LLMs) when they design biological mutations. It presents 30 hazardous mutation‑design scenarios and evaluates LLMs in a multi‑round, tool‑augmented setting, scoring designs against 187,695 empirical wet‑lab measurements using a Biosecurity Risk Score. Initial tests on 15 frontier models show that even without tools, a majority of designs exceed average risk, and tool use further amplifies the threat.
By Yuzhen Wang, Chunxiao Li, Baichen Le, Lewen Yan, Jie Zhang, Yucheng T. Yang, Xing-Ming Zhao, Lijun Li, Jing Shao
The paper introduces Anticipatory Generative Modeling, a new graph-constrained approach for generating road‑network trajectories that focuses on behavior modeling rather than simple location simulation. It presents Foresight, a probabilistic model that iteratively anticipates, selects, updates, and re‑anticipates future transitions to produce realistic routes that stay on the road network. Experiments show that Foresight outperforms existing state‑of‑the‑art generative methods.
By Zhenhao Huang, Li Sun, Philip S. Yu
Branching On-Policy Distillation (BranchOPD) enhances on‑policy distillation by selecting informative decision points based on student uncertainty and teacher‑student disagreement, resampling continuations from shared prefixes, and applying token‑level teacher supervision to both main trajectories and branches. This approach reduces student‑generated token costs by about 19% on ALFWorld and 31% on WebShop compared to seven independent trajectories, and introduces protocol‑boundary repair (PBR) to mitigate repetition under protocol constraints. BranchOPD outperforms Vanilla OPD and Multi‑Rollout OPD across benchmarks, achieving significant gains in mathematics accuracy and ALFWorld success rates, with mechanistic analyses showing that branching at uncertain decisions expands behavior coverage.
By Dongyang Chen, Shudong Liu, Shubo Li, Duoxun Tang, Jinwei Liang, Zehuan Wu, Lewei Lu
Kronecker Product Attention (KPA) is introduced as a memory‑efficient method for compressing the key‑value cache in autoregressive models. By representing key and value activations as sums of Kronecker products of compact, token‑dependent matrix factors, KPA generalizes Tensor Product Attention and can capture richer block‑structured patterns. Empirical results show that KPA matches or surpasses existing techniques such as MHA, MQA, GQA, and MLA across multiple model scales and benchmarks, achieving up to 10× compression while preserving relative‑position information through a structured rotary embedding.
By Bahareh Nikpour, Hossein Rajabzadeh, Maryam Dialameh, Vahid Partovi Nia, Parsa Omidi