Proposes ATLAS, a latent-space transformer framework combining diffusion, stochastic interpolants, and CRPS for state-of-the-art probabilistic medium-range weather forecasting, outperforming ECMWF IFS.

Jean Kossaifi, Nikola Kovachki, Morteza Mardani et al.

Introduced TAC and Soft-TAC methods, significantly enhancing reward function design efficiency and accuracy.

Calarina Muslimani, Yunshu Du, Kenta Kawamoto et al.

Jet-RL framework improves RL training efficiency with unified FP8 precision flow, achieving 33% rollout speedup.

Haocheng Xi, Charlie Ruan, Peiyuan Liao et al.

Proposes coherence optimization as a unified framework for self-improvement, proving its equivalence to description-length regularization, and introduces Gibbs sampling for scalable optimization.

Tianyi Qiu, Ahmed Hani Ismail, Zhonghao He et al.

Introduces RTCE benchmark using lossless compression algorithms to evaluate LLMs' bidirectional code reasoning and invertibility, revealing fundamental limitations.

Nickil Maveli, Antonio Vergari, Shay B. Cohen

FastAV reduces AV-LLM inference computation by over 40% using a two-stage pruning strategy.

Chaeyoung Jung, Youngjoon Jang, Seungwoo Lee et al.

EMoE introduces eigenbasis-guided routing, balancing expert utilization and diversity, achieving 88.14% Top-1 accuracy on ImageNet.

Anzhe Cheng, Shukai Duan, Shixuan Li et al.

Unified coverage framework Uρ interpolates between KL divergence, average, and minimax exploration via parameter ρ, optimizing state-action visitation in reward-free MDPs.

Xihe Gu, Urbashi Mitra, Tara Javidi

Proposes new long-context robust activation probes, improving misuse detection accuracy and efficiency.

János Kramár, Joshua Engels, Zheng Wang et al.

Proposes LD-GCN, an encoder-free GNN framework for low-dimensional modeling of parameterized time-dependent PDEs, improving interpretability and efficiency.

Lorenzo Tomada, Federico Pichi, Gianluigi Rozza

This paper proves that supervised fine-tuning (SFT) and reinforcement learning (RL) in post-training are inherently coupled; RL degrades SFT, and vice versa.

Xueyan Niu, Bo Bai, Wei Han et al.

ALVGL employs sparse+low-rank decomposition with ADMM to enhance causal discovery efficiency and accuracy in high-dimensional and confounded data.

Pingchuan Ma, Qixin Zhang, Shuai Wang et al.

RLVR training exhibits strong linear trajectories in parameters and outputs, enabling model extrapolation and 6.1× speedup via low-variance noise filtering.

Tianle Wang, Jiayu Liu, Zhongyuan Wu et al.

Proposed feature-dependent noise models tested in DMControl and Meta-world; PbRL without explicit denoising outperforms noise-robust methods.

Yuxuan Li, Harshith Reddy Kethireddy, Srijita Das

Using OpenForesight dataset, retrieval, and reinforcement learning, trained OpenForecaster 8B achieves competitive future prediction accuracy.

Nikhil Chandak, Shashwat Goel, Ameya Prabhu et al.

Proposes a rubric-guided reinforcement learning framework to enhance AI research plan generation, achieving 70% expert preference and 12-22% improvements across domains.

Shashwat Goel, Rishi Hazra, Dulhan Jayalath et al.

MFT method surpasses LoRA in VLMs without updating weights, enhancing performance.

Mingyuan Zhang, Yue Bai, Yifan Wang et al.

EnFlow combines flow-based conformer generation with explicit energy modeling, enabling low-energy molecular structure discovery with 1-2 ODE steps.

Guikun Xu, Xiaohan Yi, Ziqiao Meng et al.

Co-GRPO optimizes both model and inference schedule, enhancing MDM generation quality.

Renping Zhou, Zanlin Ni, Tianyi Chen et al.

kooplearn is a Scikit-Learn compatible library for evolution operator learning, supporting linear, kernel, and deep learning estimators.

Giacomo Turri, Grégoire Pacreau, Giacomo Meanti et al.