Introduces DeepONet within a GAN framework for high-resolution, long-range precipitation nowcasting, improving stability and physical consistency.

Mohammad Kian Golkar, Luciano Alves de Oliveira, Mohammad Khanjani

ECCBench evaluates vision-language models' memory beyond accuracy.

Shmuel Berman, Jia Deng

This study reveals that contrastive preference optimization (e.g., DPO) unintentionally transfers sycophantic behaviors from teacher models, with the log-ratio of teacher sycophancy predicting student behavior.

Camila Blank, Zhuofan Ying, Christopher Potts et al.

Liquid Gated Attention (LGA) enables continuous-time modeling with input-driven gating, achieving linear temporal complexity.

Yiheng Jiang, Yuanbo Xu, Yongjian Yang

This paper systematically organizes 2025-2026 optimizers into a multi-dimensional framework, emphasizing matrix-awareness, temporal estimation, and system representation, highlighting no single optimizer dominates.

Ruoran Xu

Optimized Relief-based algorithms (e.g., MultiSWRF*, μ-Relief) significantly enhance detection of 2nd-order interactions with 10-35x faster runtime.

Kia Kazemi-Nia, Harsh Bandhey, Philip J. Freda et al.

DARTS uses entropy-weighted L1 and position biases to correct autoregressive decoder biases, improving multi-task model merging with only 0.1% parameter overhead.

Aaryan Ajay Sharma, Sai Nishanth Padala, Seganrasan Subramanian

Proposes GGEW, a Gibbs-based adaptive weighting framework, improving forecast combination by modeling model errors probabilistically.

Prasen R. Nuthanakaluva, Nava K. Gaddam

PhyMamba framework achieves robust battery health prognostics with 31.8% mean error reduction.

Sara Sameer, Yunyi Zhao, Wei Zhang et al.

Introduces Fast Weight Attention with normalized first-order updates, enhancing long-sequence modeling in continual learning scenarios.

Yifan Zhang, Steve Ta, Jasper Zhang et al.

Introduces the concepts of information floors and model gaps, quantifying the limits of parallel draft acceptance; results show models are far above their theoretical bounds.

Xinwei Qiang, Xiang Fang, Chang Chen et al.

TRACE-CRC employs trajectory-adaptive conformal risk control for multi-step CSI prediction, achieving 0.933 coverage with smaller uncertainty balls.

Kiarash Rezaei, Mehdi Sattari, Javad Aliakbari et al.

Introduces feature geometry-based covariance envelope bounds under volume sampling, with a sharp phase boundary defined by feature margin ν_A.

Kihun Rhee

VGAS achieves training-free reward alignment by reducing variance and adaptive selection, enhancing biological sequence generation.

Kwanyoung Kim

Proposes SVD-based channel group-shared low-rank approximation to compress pointwise convolutions in large-kernel CNNs, reducing parameters by over 81%.

Hao Luo, Yiting Yang, Wenyi Zhao et al.

Proposes Bellman calibration using monotone transformations to reduce occupancy ratio bias in offline RL.

Lars van der Laan, Nathan Kallus

Proposes a geometric framework analyzing robustness of neighborhood fairness audits, validated by experiments on benchmark datasets.

Binita Maity

ELR (effective learning rate) governs loss dynamics; matching ELR schedules collapses loss trajectories across configurations.

Zihan Liu, Ruiheng Zheng, Shaobo Zhang et al.

PhysicsBench provides a unified benchmark for generative and predictive models in engineering design, evaluating geometric and physical accuracy across multiple tasks and scales.

Sang Won Lee, Hyogu Jeong, Namwoo Kang

This study systematically analyzes how base model priors, reward density, and prompt diversity influence RL post-training success, validated via controlled experiments.

Donovan Clay, Saket Gollapudi, Sankar Harilal et al.