PAC Confidence Sets combine temperature scaling with a binomial generalization bound, yielding finite-sample coverage on ImageNet, tracking, and Half-Cheetah.

Sangdon Park, Osbert Bastani, Nikolai Matni et al.

Proposed Face X-ray detects face forgery by identifying blending boundaries, trained without fake images, achieving over 98% AUC on multiple datasets.

Lingzhi Li, Jianmin Bao, Ting Zhang et al.

LayoutLM integrates text, layout, and visual features for document understanding, achieving 79.27% F1 on FUNSD.

Yiheng Xu, Minghao Li, Lei Cui et al.

Survey of 2000 Americans reveals high concern for 13 AI governance issues; trust in institutions remains low, with no direct link to support for AI development.

Baobao Zhang, Allan Dafoe

Derived a spike-time-based learning rule for LIF neurons enabling error backpropagation, validated on BrainScaleS-2 with 97.1% accuracy on MNIST.

Julian Göltz, Laura Kriener, Andreas Baumbach et al.

Proposes TCoN with cross-domain co-attention for temporal feature alignment, achieving 15%+ accuracy boost in cross-domain action recognition.

Boxiao Pan, Zhangjie Cao, Ehsan Adeli et al.

Pretrained Audio Neural Networks (PANNs) trained on AudioSet achieve 0.439 mAP, surpassing previous models, demonstrating strong transferability across audio tasks.

Qiuqiang Kong, Yin Cao, Turab Iqbal et al.

Using Adversarial Representation Active Learning, achieve highest classification accuracy with minimal labels on datasets like MNIST.

Ali Mottaghi, Serena Yeung

Introduces Temporal Fusion Transformer (TFT), combining multi-scale temporal modeling with interpretability, achieving 20% accuracy improvement in multi-horizon forecasting.

Bryan Lim, Sercan O. Arik, Nicolas Loeff et al.

Proposes bias-corrected estimators for smooth functionals in high-dimensional normal models, achieving asymptotic efficiency and minimax optimality.

Vladimir Koltchinskii, Mayya Zhilova

PEGASUS introduces Gap Sentences Generation (GSG) as a pretraining objective, significantly improving abstractive summarization performance.

Jingqing Zhang, Yao Zhao, Mohammad Saleh et al.

Proposes P-CapsNets, removing routing, replacing convolution with capsule layers, and packaging tensors, achieving high efficiency with fewer parameters.

Zhenhua Chen, Xiwen Li, Chuhua Wang et al.

A vision-cone, position-attraction model produces worms, aggregates, and stable nematic bands without velocity alignment.

Lucas Barberis, Fernando Peruani

Proposes Differentiable Volumetric Rendering (DVR) for implicit 3D shape and texture learning from RGB images without 3D supervision.

Michael Niemeyer, Lars Mescheder, Michael Oechsle et al.

Constructs approximate Fekete points via truncated orthogonal expansion by maximizing kernel Gram matrix determinant, achieving super-exponential convergence with Gaussian kernels.

Toni Karvonen, Simo Särkkä, Ken'ichiro Tanaka

OpenAI Five used large-scale deep RL, training 10 months to beat Dota 2 champions, overcoming environment complexity.

OpenAI, :, Christopher Berner et al.

Proposed cascade cost volume improves high-res multi-view stereo accuracy by 35.6%, reducing GPU memory and runtime by over 50%.

Xiaodong Gu, Zhiwen Fan, Zuozhuo Dai et al.

Proposes a GNN-based decentralized multi-robot path planning framework, combining CNN feature extraction with imitation learning, achieving near-optimal success rates in complex environments.

Qingbiao Li, Fernando Gama, Alejandro Ribeiro et al.

Proposes federated learning algorithms with privacy and efficiency improvements; experiments show 20% accuracy gain on CIFAR-10.

Peter Kairouz, H. Brendan McMahan, Brendan Avent et al.

Deep symbolic regression employs RNNs and risk-seeking policy gradients to recover mathematical expressions, outperforming traditional methods on benchmark problems.

Brenden K. Petersen, Mikel Landajuela, T. Nathan Mundhenk et al.