PPLNs enhance event camera temporal vision inference using parametric piecewise linear functions, excelling in tasks like steering prediction.

Chen Song, Zhenxiao Liang, Bo Sun et al.

Proposed STAtten enhances Spike Transformers by integrating block-wise spatial-temporal attention, achieving significant accuracy gains on CIFAR and ImageNet datasets.

Donghyun Lee, Yuhang Li, Youngeun Kim et al.

RNG uses neural Gaussian points to achieve fast relightable 3D reconstruction, supporting complex materials and soft boundaries, with only 1.3 hours training.

Jiahui Fan, Fujun Luan, Jian Yang et al.

RoboNurse-VLA integrates SAM 2 and Llama 2 for high-precision surgical instrument handover, achieving over 90% success in unseen scenarios.

Shunlei Li, Jin Wang, Rui Dai et al.

Proposes a method to learn a bridge policy via reinforcement learning to enhance robot adaptability in novel environments.

Alicia Li, Nishanth Kumar, Tomás Lozano-Pérez et al.

TwinCL model uses twin encoders to optimize contrastive learning, improving recommendation accuracy by 5.6%.

Chengkai Liu, Jianling Wang, James Caverlee

Emu3 employs a single Transformer trained solely with next-token prediction, surpassing SDXL and LLaVA-1.6 in multimodal tasks.

Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo et al.

Proposes a decentralized multi-agent safety control framework combining black-box predictors, conformal decision theory, and control barrier functions, with theoretical safety guarantees.

Sacha Huriot, Hussein Sibai

MinerU employs PDF-Extract-Kit with multi-model fusion and rule-based post-processing for high-precision document content extraction.

Bin Wang, Chao Xu, Xiaomeng Zhao et al.

Fully integrating Flang with standard MLIR achieves up to three times performance improvement.

Nick Brown

Proposes a sparse, hierarchical Bayesian network world model supporting open-ended learning with interpretability and scalability.

Lancelot Da Costa

Search3D constructs hierarchical scene representations for open-vocabulary 3D segmentation, achieving 49.5 AP on MultiScan benchmark.

Ayca Takmaz, Alexandros Delitzas, Robert W. Sumner et al.

Embodied-RAG employs hierarchical semantic forests for non-parametric memory, boosting robot navigation and QA performance.

Quanting Xie, So Yeon Min, Pengliang Ji et al.

GRAPHEDX: neural set divergence-based estimator for general-cost graph edit distance.

Eeshaan Jain, Indradyumna Roy, Saswat Meher et al.

MoGenTS employs joint-level spatial-temporal VQ-VAE and 2D transformer with masking for high-quality text-driven human motion synthesis, reducing FID by 26.6%.

Weihao Yuan, Weichao Shen, Yisheng He et al.

Vulnerability-aware Adversarial Training (VAT) adaptively adjusts perturbation magnitudes based on user fit, significantly improving robustness against poisoning attacks.

Kaike Zhang, Qi Cao, Yunfan Wu et al.

Introduces AV-ALOHA, a low-cost bimanual robot system with 7-DoF active vision, significantly improving manipulation success rates in occlusion and detail-critical tasks.

Ian Chuang, Andrew Lee, Dechen Gao et al.

Proposes Schrödinger bridge-based deep generative model for efficient high-quality conditional sampling, outperforming existing diffusion methods.

Hanwen Huang

Molmo is an open-source vision-language model trained on PixMo, achieving 96.3 score on benchmarks with 72B parameters, surpassing closed models.

Matt Deitke, Christopher Clark, Sangho Lee et al.

EventHDR method reconstructs high-speed HDR videos from event sequences, providing the first real paired dataset.

Yunhao Zou, Ying Fu, Tsuyoshi Takatani et al.