CoLMIN is an LLM-based framework for cooperative autonomous driving that addresses premature convergence to suboptimal solutions in multi-solution traffic scenarios. It introduces a Multi-Intent Negotiation module that generates multiple candidate driving intentions, an Evaluation-based Shallow Reflection Module that provides feedback to accelerate consensus, and a Deep Reflection Module that mitigates cognitive fixation by reflecting on negotiation histories. Experiments in the CARLA simulation show that CoLMIN outperforms existing methods in challenging interactive driving scenarios.
By Zhe Huang, Zhaoxin Fan, Shuo Wang, Wenjun Wu, Xuan Zhao, Min Liu
The paper introduces Out-of-Distribution Semantic Occupancy Prediction, a task that focuses on detecting unknown objects in 3D voxel space for autonomous driving. It proposes Realistic Anomaly Augmentation to create two new datasets, VAA-KITTI and VAA-KITTI-360, and presents the OccOoD framework, which uses Cross‑Space Semantic Refinement to improve OoD detection while maintaining semantic occupancy accuracy. Experiments show OccOoD achieves an AuROC of 65.50% and an AuPRCr of 31.83% within a 1.2 m radius, demonstrating strong generalization to real‑world urban scenes.
By Yuheng Zhang, Mengfei Duan, Kunyu Peng, Yuhang Wang, Ruiping Liu, Fei Teng, Kai Luo, Zhiyong Li, Kailun Yang
The paper presents Test-Time Adaptation via Cache Personalization (TTA‑CaP), a gradient‑free, cache‑based method that personalizes vision‑language models for facial expression recognition in videos. TTA‑CaP uses three complementary caches—a personalized static cache, a positive target cache, and a negative target cache—controlled by a tri‑gate mechanism to prevent corruption and provide robust subject‑matched evidence. Experiments on BioVid, StressID, and BAH datasets show that TTA‑CaP outperforms state‑of‑the‑art test‑time adaptation methods while keeping computational and memory overhead low.
By Masoumeh Sharafi, Muhammad Osama Zeeshan, Soufiane Belharbi, Alessandro Lameiras Koerich, Marco Pedersoli, Eric Granger
DenseScout is a 1.01M‑parameter dense‑response selector designed for edge platforms that directly optimizes ranked patch‑center prioritization, eliminating the need for detector‑style box regression. It aligns output representation, supervision, and decoding, and is jointly designed with transport‑aware execution and QoS‑oriented evaluation. Experiments on VisDrone and DOTA show that DenseScout achieves stronger low‑budget recall than detector‑derived selectors, and cross‑platform profiling on Jetson Orin NX and RK3588 demonstrates that deployable utility depends on selector quality, memory movement, and heterogeneous runtime realization.
By Zhouzhi Xiong, Zimo Zeng, Yi Chen, Shuqi Xu, Yunfeng Yan, Donglian Qi
ASTRA (Asynchronous Spatio-Temporal Reconstruction via Trajectory Alignment) tackles the challenge of reconstructing dynamic 3D scenes from temporally asynchronous multi‑camera data. By using 2D motion trajectories as texture‑robust supervision, it jointly optimizes temporal offsets and 3D representations, aligning projected 3D point motion with observed 2D paths while masking unreliable constraints. Experiments on Gaussian Splatting backbones show that ASTRA retains high‑frequency spatial detail, improves PSNR by ~1.4 dB, reduces temporal‑offset MAE by 54 %, and nearly quadruples synchronization success even with up to 25‑frame offsets.
By Junyu Zhu, Hao Zhu, Xinzhuo Zhang, Xu Zhang, Hongdong Li, Zhan Ma, Xun Cao
The paper investigates Double Machine Learning (DML) estimators under structure‑agnostic (SA) models, which assume the data‑generating law lies within a neighborhood of fixed machine‑learning estimates. It shows that for two of three studied functionals—the quadratic functional in the Gaussian sequence model and the quadratic density integral functional—the DML estimators are asymptotically inadmissible, being dominated by second‑order empirical higher‑order influence function (HOIF) estimators. For the third functional, the expected conditional covariance, both DML and HOIF estimators remain minimax but neither dominates the other.
By Lin Liu, Rajarshi Mukherjee, James M Robins
The paper introduces Objective-aware Trajectory Credit Assignment (OTCA), a framework that refines reinforcement learning for diffusion-based visual generation. OTCA decomposes credit across denoising steps and allocates multiple reward signals adaptively, addressing the coarse, uniform reward assignment of existing GRPO pipelines. Experiments demonstrate that OTCA consistently enhances image and video generation quality across various metrics.
By Rui Li, Ke Hao, Yuanzhi Liang, Haibin Huang, Chi Zhang, Yun Gu, Xuelong Li
The study reports on three separate executions of a four‑qubit ZZ feature map on the ibm_fez backend, each using 1024 shots per circuit. The authors compare the resulting hardware‑reconstructed Gram matrices to an exact statevector reference, finding off‑diagonal RMSE values of 0.0878, 0.0864, and 0.0427 for the baseline, dynamical decoupling, and gate‑twirling jobs respectively, and centered kernel alignment scores ranging from 0.933 to 0.989. The gate‑twirling job shows the smallest deviation across all metrics, while the baseline job’s contrasts remain stable under deletion‑stable diagnostics; the study notes that sampling alone cannot explain the observed errors and that implementation fidelity and task relevance are distinct diagnostic axes.
By Rostyslav Sipakov
The paper introduces DODR, a deterministic operator‑driven reasoning architecture that models reasoning as graph computations in a high‑dimensional linear‑algebraic space, replacing token‑level sampling with matrix operations. Reasoning states are snapshot vectors of semantic units, and three trainable matrix operators—deduction, induction, and abduction—implement Peirce’s inference types. Experiments on 503 records demonstrate near‑perfect deduction, high generalization for induction, and significant gains for abduction, while the design guarantees zero hallucination and supports continual learning.
By Weicai Huang (Beijing MQPat Technologies, Co., Ltd.)
The paper documents a growing trend of removing safety guardrails from open-weight AI models, profiling the ecosystem that produces, redistributes, and applies these models. Between January 2024 and March 2026, 3,471 original uncensored models were identified on HuggingFace, each re‑packaged an average of 2.4 times, with three actors responsible for 52 % of all 8,164 compressed redistributions. After quantization and mirroring across platforms such as Ollama, these models persist even when upstream versions are removed, and 25 % of 1,643 GitHub applications that integrate uncensored large language models were classified as explicitly malicious.
By 10a Labs, :, Juliette Garcia, Hailey May, Bobby McKenzie, David Pham, Matthew Swain, Joshua Valdez, Corie Wieland, Zachary Yahn
AlcaTRAz is a prompt‑level defense that uses rule trees to insert controlled character‑level perturbations into input text, disrupting jailbreak attacks without modifying or retraining the target LLM. It operates solely on the input, making it suitable for black‑box deployments, and was evaluated on 33 open‑weight models and 22 jailbreak types, outperforming three baseline defenses in 73.4 % of model‑attack combinations. While it significantly reduces high‑severity jailbreak success, it does not eliminate it and is intended as one layer of a broader defense strategy.
By Jakub Re\v{s}, Petr Ka\v{s}ka, Martin Pere\v{s}\'ini, Martin Ukrop, Kamil Malinka
The paper introduces a Kolmogorov-Arnold network as an interpretable post‑hoc surrogate to assess the trustworthiness of YOLOv10 object detections, using seven geometric and semantic features. Its additive spline structure allows direct visualization of each feature’s influence, revealing when confidence scores are reliable or unreliable. A bootstrapped BLIP language‑image model generates scene captions, providing a lightweight multimodal interface that does not compromise interpretability. Experiments on COCO and University of Bath campus images show the framework accurately flags low‑trust predictions under blur, occlusion, or low texture, offering actionable insights for risk mitigation.
By Marios Impraimakis, Daniel Vazquez, Feiyu Zhou
The paper introduces Reflection-Aware GRPO (RA‑GRPO), a reinforcement‑learning framework that aligns diffusion generative models with human preferences. It uses Diffusion Reflection to correct intermediate sampling paths by reversing the diffusion process, and Counterfactual Path Synthesis to embed these corrected trajectories into the policy, avoiding extra inference cost. Experiments on text‑to‑image and text‑to‑video models show RA‑GRPO outperforms existing methods, reducing reward hacking and improving generalization while remaining architecture‑agnostic.
By Junlong Wu, Jiuzhou Lin, Jia Sun, Boheng Zhang, Huaiqing Wang, Dewen Fan, Houde Liu, Qianqian Gan, Fan Yang, Tingting Gao
The paper introduces BMT, a unified hierarchical Vision Transformer that jointly performs SAR-to-optical image translation and semantic segmentation. It incorporates a LocalViTBlock, an enhanced output module, a ControlNet-style conditional injection, and a bounded Kendall uncertainty weighting scheme to balance the two tasks. Experiments on paired and unpaired datasets demonstrate competitive performance in both translation quality and segmentation accuracy.
By Siyuan Liu, Xuze Zhang, Yongshun Wang, Licong Pan, Hang Liu, Huihui Li
BEAM3R is a dose estimation framework that operates in beam’s-eye-view (BEV) for both photon and proton beamlet calculations. It uses a Mamba‑3 state‑space depth‑sequence core combined with physics‑based transport conditioning, sharing a 2D CNN encoder‑decoder architecture to process per‑plane BEV slices. The system achieves high accuracy on the DoseRAD2026 challenge, with local gamma pass rates above 96% for CT‑based models and runtimes under 25 s for photon and 18 s for proton predictions.
By Chen Cheng, Michael Ferraro, James Grover, David E J Waddington, Emily Hewson
The paper introduces LARK, a two‑stage latent reasoning framework designed to mitigate cross‑modal dilution in multimodal recommendation systems. In the first stage, learnable latent tokens are interleaved with chain‑of‑thought reasoning and aligned with a frozen vision encoder to preserve visual details. The second stage projects these latent representations through a bridge MLP, employing item‑to‑item contrastive learning and aligning intermediate features with the first‑stage hidden states to anchor final embeddings to the model’s reasoning output. Experiments on three public benchmarks and an industrial dataset demonstrate that LARK achieves state‑of‑the‑art performance across multiple recommendation architectures, with ablation studies confirming the contribution of each component.
By Jiarui Jin, Anyang Ji
MURAL is a multimodal recommendation framework that replaces static similarity graphs with a dynamic topology discovery process. It uses an Adaptive Edge Learner to find latent item-item correlations efficiently and an Uncertainty-Aware Fusion module to down‑weight noisy modality signals based on aleatoric uncertainty. The model also incorporates a contrastive teacher‑student alignment to stabilize training and has been shown to outperform state‑of‑the‑art baselines on large‑scale TikTok and Amazon datasets, providing both higher accuracy and interpretability.
By Ahmad Mousavi (Department of Mathematics,Statistics American University), Majid Alikhani (Independent Researcher), Yeon-Chang Lee (Department of Computer Science,Engineering Ulsan National Institute of Science,Technology), Roberto Corizzo (Department of Computer Science American University), Yeganeh Abdollahinejad (Department of Biosystems,Agricultural Engineering Michigan State University)
The paper examines how large language models (LLMs) tend to overuse persona attributes in persona-based dialogue generation, producing unnatural responses. It identifies a systematic bias in LLMs to incorporate all provided persona details and shows that current metrics cannot assess contextual appropriateness. To address this, the authors introduce Self-CONtrastive Persona Overuse Suppression (SCONPOS), which intervenes in the prompt encoding stage to reduce overuse, and propose the Persona Appropriateness Score (PAS), a new metric that penalizes both overuse and underuse of persona attributes.
By Jongkyung Shin, Inkyu Lee, Chiehyeon Lim
The paper presents a scalable two‑stage clustering algorithm that guarantees per‑sample quality guardrails for large‑scale LLM‑based recommender systems. By first forming Mini‑batch K‑Means clusters and then greedily selecting representatives that meet user‑specified similarity and attribute constraints, the method ensures each sample inherits only relevant and safe outputs. Benchmarks show the approach runs faster and scales to millions of inputs, achieving a 50‑fold reduction in downstream LLM cost and runtime while maintaining personalization.
By Longshaokan Wang, Wai Tsang Keung, Punit Ghodasara, Roman Wang, Ali Dashti, Francesc Moreno-Noguer
The paper introduces GRACE, a framework that breaks down large language model (LLM) responses into atomic claims and grounds them against trusted knowledge priors using a weighted bipartite graph. Edge weights enable weighted centrality analysis to classify claims as Grounded, Refuted, or Boundary, identifying hallucinations and frontier knowledge. An objective called Return on Attention (RoA) prioritizes expert review only for high‑uncertainty claims, and verified claims become new evidence anchors, creating a loop that expands the knowledge base across iterations.
By John Seon Keun Yi, Joshua R. Minot, Dokyun Lee