PrefO2O introduces an offline-to-online latent preference transfer framework that enables zero‑shot preference‑based reinforcement learning on unseen tasks without human feedback. It learns a task‑invariant latent preference space from limited offline preference data, disentangling transferable preference structures from task‑specific variations. The method uses confidence‑weighted optimal transport to generate reliable pseudo‑preference supervision for online reward adaptation, achieving performance that matches or exceeds traditional PbRL methods trained with real human feedback across various robotic manipulation and locomotion benchmarks.
By Ruiqi Wang, Dezhong Zhao, Xuesu Xiao, Erdem Biyik, Byung-Cheol Min, Guohua Chen
OpenHome is a visual-only, multiple-choice benchmark designed to evaluate smart‑home camera video understanding. It includes 2,254 questions across 1,118 real‑world clips, organized into 10 tasks spanning perception, reasoning, and insight levels. The benchmark reveals that closed‑source models outperform open‑weight models, with the best closed‑source model surpassing the best open‑weight model by significant margins on all levels.
By Xianglong Wang, Neil Rodrigues, Pei Guo, Yiru Fang, Kevin Beussman, Shuyang Wang, Jingzhi Wang, Liangju Li, Shunyao Xu, Zhongwei Cheng, Wei Li, Lin Chen
The paper evaluates five publicly released EEG foundation models and an in‑house JEPA‑style model using a protocol that includes three key controls: an untrained encoder, spectral band‑power features, and a probe layer selected without test‑set look‑ahead. The study finds that selecting the probe layer by test score can modestly improve peak performance and alter the peak‑versus‑output relationship, that random initialization can match or surpass pretraining on certain tasks, and that full fine‑tuning rarely yields gains beyond seed variability. These findings lead the authors to propose four new reporting practices and to release all experimental results for re‑analysis.
By Mingyu Jeon, Suwan Cho, Jae Young Suh
ReMedI introduces an inference‑time refinement framework that combines a pre‑trained molecular language model with a Metropolis‑Hastings search to perform multi‑objective optimization of drug leads. The method proposes edits conditioned on the current molecule and refines them to improve properties, preserve lead similarity, and maintain scaffold integrity. Experiments on C‑MuMOInstruct and SMDD benchmarks show that ReMedI outperforms RL‑based baselines, achieving higher success rates and requiring far fewer training leads.
By Jia Yi Zhang, Shivanshu Shekhar, Allison McLevige, Tong Zhang
MG-REPA introduces Multi-Granularity Representation Alignment for generative target speaker extraction, aligning hidden states with a self‑supervised learning model at the frame level and a speaker verification model at the utterance level. Temporal normalization removes utterance‑level statistics from the frame‑level target, allowing each branch to focus on its specific granularity. On Libri2Mix, MG-REPA achieves comparable or better performance to a 2,000‑epoch baseline while requiring only 750 epochs, and ablation studies confirm its superiority over single‑granularity alignment.
By Dohwan Kim, Jung-Woo Choi
Dual-Adaptive Z-Score PPO (PPO-DAZ3) is an adaptive-clipping variant of Proximal Policy Optimization that replaces the fixed clipping threshold with sample-dependent thresholds derived from action confidence and normalized advantage magnitude. The method retains the rest of the standard PPO training pipeline and is evaluated on three classic control benchmarks. Results show that PPO-DAZ3 improves learning on LunarLander-v3, while a smooth advantage-based variant (PPO-AZ3S) performs best on Acrobot-v1 and standard PPO remains competitive on CartPole-v1, indicating adaptive clipping benefits harder control tasks.
By Taha M. Mahmoud, Naima Kaabouch
The paper introduces Part‑DINO, a training‑free framework that enhances self‑supervised Vision Transformer (ViT) features from DINO with low‑level image boundaries to improve part discovery. By partitioning images using boundary detectors, pooling DINO features within each region, and clustering them, Part‑DINO yields structurally coherent part assignments. It outperforms other training‑free baselines on four standard benchmarks and remains competitive with weakly‑supervised and unsupervised methods.
By Samyak Rawlekar, Nikhil Chowdary Paleti, Amey Gupta, Narendra Ahuja
ChoreoStudio is a music‑driven group choreography system that explicitly plans what formations dancers should adopt, when to transition between them, and which dancers execute each transition. It uses a music disentangler to extract rhythmic and structural features, a frozen omni‑modal model to generate a global semantic script, and a Director module that outputs centroid‑relative formation points, transition gates, and mobility allocations. These high‑level plans guide a diffusion generator that produces beat‑aligned body motion, resulting in state‑of‑the‑art motion quality, group organization, and formation dynamics on the AIOZ‑GDance benchmark.
By Jing Xu, Cunjian Chen, Qiuhong Ke
TENET is a Time-point Encoding Network designed for unsupervised multivariate time series anomaly detection. It reframes the task as estimating current-point system-state compatibility by constructing a latent representation of the target time point from variable-level temporal context within an input window. The model uses a temporal layer to retrieve context relative to the current point, encodes time with point-centered embeddings, and adaptively mixes this context with the current representation, then evaluates the latent state with a fixed standard-normal negative log-likelihood score, achieving state-of-the-art performance across benchmark datasets.
By Dongchan Cho, Juwon Hwang, Keumyeong Kang, Jiho Han, Namsoon Jung
The paper introduces MCP‑Injector, a fault‑injection middleware that simulates realistic failures in the Model Context Protocol used by large‑language‑model agents. Using data from over 57,000 issue and pull‑request threads, the authors calibrate a fault model and run 1,350 experiments across different backbone sizes, agent stacks, and fault conditions. Results show that permanent‑heavy faults drastically reduce task‑completion scores and that larger models (120B) do not gain fault tolerance, highlighting that framework‑level error handling is more critical than model scale.
By Antonio Ken Iannillo, Joshua Segun Owotogbe, Roberto Natella, Francesco Avallone, Kristina Kudryavtseva, Indika Kumara
The paper introduces Back-to-the-Future (BTTF), an agentic framework designed to automate post-simulation verification and waveform debugging in chip design. BTTF transforms large simulation dumps into a normalized SQLite database and uses a multi-agent orchestration engine to convert natural-language queries into schema-aware SQL, linking anomalies to versioned RTL code. In a benchmark of 150 queries, BTTF achieves 95.33% execution accuracy, demonstrating a viable path toward autonomous EDA verification.
By Je Yang, Ivan Lobov, Thomas Karpati
Vision‑language models (VLMs) are evaluated on their ability to follow a continuous visual path, a fundamental operation that requires maintaining a selected trajectory through successive visual inputs. In controlled experiments that eliminate confounding factors such as crossings or overlaps, VLMs exhibit fragile tracing behavior, failing to preserve the correct path when other nearby paths are present. Internal representation analysis reveals that the necessary information for correct path continuation weakens already at the visual encoding stage, and common mitigation strategies—model scaling, additional reasoning, or explicit tracing instructions—do not reliably address this vulnerability, even in more realistic visual settings.
By Hyesoo Hong, Min Soo Kim, Wonje Jeung, Sangyeon Yoon, Dongjae Jeon, Albert No
The paper audits the RVL-CDIP document classification benchmark, uncovering that about 10.5% of its 400,000 documents have incorrect labels and that 35.2% of the test set contains duplicates or near-duplicates in the training set. It shows that identification codes on pages enable 69% classification accuracy, and removing them drops performance by 11 points; OCR engine choice can shift accuracy by 4.1 to 8.3 points. After correcting defects, 13 re‑trained models improve by 0.9 to 2.9 points, with LayoutLMv3 achieving 96.5% accuracy, and the authors release a cleaned corpus, annotations, and out‑of‑distribution data.
By Stefan Larson, Kevin Leach
NodeSyn is a method that generates socially aligned synthetic data for evaluating AI models. It combines a fine‑tuned taxonomy generator with credible sources and sensitive attributes to produce queries that reflect real societal contexts and nuanced harms. In tests across ten LLMs in Medical Advice and Self‑Harm, NodeSyn outperformed generic synthetic baselines, raising failure rates significantly and revealing that failures stem from both topic selection and sociotechnical grounding.
By Qazi Mamunur Rashid, Xuan Yang, Zhengzhe Yang, Yanzhou Pan, Erin van Liemt, Darlene Neal, Kshitij Pancholi, Jamila Smith-Loud
The paper introduces MAGIC, a method for improving test-time robustness of large vision‑language models by mitigating language bias. MAGIC evaluates decision margins to selectively activate textual and visual perturbation branches, integrates logits from multiple perturbed views, and restores the most informative corrupted views for ambiguous samples. It also uses a calibration dataset to prune ineffective branches, reducing inference cost while enhancing performance on benchmark datasets.
By Shunshun Gu, Enze Zhang, Ming Pei, Hang Zhou, Xiao Luo
DriveOSS is a new real‑world multimodal dataset and benchmark for 3D anomaly segmentation and open‑set semantic segmentation in autonomous driving. It contains 12,661 synchronized frames from six RGB cameras, a 128‑beam LiDAR, and GPS/IMU, covering 52 driving scenes under various weather conditions. The data are split into in‑distribution (ID) scenes for training and out‑of‑distribution (OOD) scenes for validation and testing, with unknown objects labeled as an outlier class to challenge current 3D perception systems.
By Zhiran Yan, Savankumar Bhanderi, Sourabh Lolge, Rui Song, Girdharbhai Paneliya, Shiva Agrawal, Gordon Elger
STROKE is a query‑efficient, score‑based black‑box adversarial attack that builds perturbations from simple oriented rectangular occlusions. It optimizes these occlusions using a lightweight greedy add‑and‑refine procedure, achieving high attack success with low query counts at nominal 1–2% area settings across diverse image classifiers. Compared to existing patch‑based attacks, STROKE shows favorable performance, reduced wall‑clock attack time, and the ablation studies confirm the effectiveness of its structural primitives.
By Ankit Gupta, Akshat Pandya, Paul Lin, Christoph Adami, Emily Dolson
HierViT is a hierarchical Vision Transformer designed for medical image classification that balances strong predictive performance with interpretability. The model first predicts clinically relevant visual appearance attributes and then uses a hierarchical reasoning process to infer the final diagnosis. Attribute‑specific prototypes and attention maps are visualized to provide complementary views of the model’s reasoning, linking representative image examples and spatial attention to semantically defined visual features.
By Luisa Gall\'ee, Meinrad Beer, Michael G\"otz
SetCon reformulates open-ended referring segmentation as set-level concept prediction, using LVLM-generated natural-language concepts instead of segmentation-specific tokens. It employs a hierarchical semantic decomposition that first predicts a shared set-level concept and then refines it into fine-grained concept groups aligned with target subsets. The method, supported by a two-stage annotation pipeline adding hierarchical semantic supervision, achieves state-of-the-art results on multiple image and video benchmarks, with performance gains increasing as the number of referred targets grows.
By Zhixiong Zhang, Yizhuo Li, Shuangrui Ding, Yuhang Zang, Shengyuan Ding, Long Xing, Yibin Wang, Qiaosheng Zhang, Jiaqi Wang
RSure-Agent is a framework designed to improve remote sensing agents by verifying tool observations before they influence downstream reasoning. It introduces a verifiable observation protocol that requires tools to provide process evidence, and builds a task‑tool reliability prior from offline feedback to assess each tool configuration’s past performance. In evaluations on EarthBench, ThinkGeo, TerraLogic, and CHOICE‑420, RSure-Agent cuts error propagation by 21.3–25.9 percentage points and boosts overall accuracy, while also reducing the number of tool calls.
By Fuyuan Liu, Nayu Liu, Wenhao Yu, Peijin Wang, Yingchao Feng, Fanglong Yao, Yue Zhou, Xue Yang, Chao Ren, Liang Wan, Wei Feng