e3 trains models with chain-based exploration, enabling extrapolation to twice the training inference length, significantly improving reasoning performance beyond training limits.

Amrith Setlur, Matthew Y. R. Yang, Charlie Snell et al.

ATRADA framework enhances aircraft trajectory datasets by generating high-quality synthetic data in latent space.

Seokbin Yoon, Keumjin Lee

Introduces SamS, an adaptive sample scheduling algorithm that improves large language model preference alignment by over 12% using model state feedback.

Zixuan Huang, Yikun Ban, Lean Fu et al.

EF-VFM extends variational flow matching with exponential family distributions, achieving state-of-the-art tabular data generation.

Andrés Guzmán-Cordero, Floor Eijkelboom, Jan-Willem van de Meent

TimeMCL employs WTA loss in multi-head neural networks for diverse probabilistic time series forecasting, achieving high-quality multi-modal predictions with theoretical guarantees.

Adrien Cortés, Rémi Rehm, Victor Letzelter

A delayed-exploration MTS algorithm matches the best predictor within O(OPT^{2/3}), with a nearly matching lower bound.

Matei Gabriel Coşa, Marek Eliáš

UNO algorithm uses orthogonalization for fast unlearning in generative models while maintaining generation quality.

Pinak Mandal, Georg A. Gottwald

Ether0, a 24B-parameter scientific reasoning model trained via reinforcement learning, outperforms existing chemistry models on molecular design tasks with high data efficiency.

Siddharth M. Narayanan, James D. Braza, Ryan-Rhys Griffiths et al.

Systematic data recipes improve reasoning models, achieving 53% on AIME 2025, 51% on LiveCodeBench, and 54% on GPQA, surpassing previous models.

Etash Guha, Ryan Marten, Sedrick Keh et al.

Proposes angle concentration as an intrinsic model signal, enabling GAIN-RL to accelerate training over 2.5× with dynamic data scheduling.

Qinsi Wang, Jinghan Ke, Hancheng Ye et al.

SynthRL uses verifiable data synthesis to generate challenging training samples, boosting visual reasoning model performance by +1-2% on benchmarks.

Zijian Wu, Jinjie Ni, Xiangyan Liu et al.

Introduces Attention-Indexed Model (AIM) to predict Bayes-optimal generalization error using statistical mechanics.

Fabrizio Boncoraglio, Emanuele Troiani, Vittorio Erba et al.

A unified ODE/SDE tutorial on flow matching and diffusion models

Peter Holderrieth, Ezra Erives

C3PO introduces a central path-inspired modification to PPO, improving constraint satisfaction and reward performance in constrained RL.

Nikola Milosevic, Johannes Müller, Nico Scherf

PointODE leverages Neural ODE for lightweight point cloud feature extraction, with only 0.58M parameters, enabling FPGA acceleration for edge devices.

Keisuke Sugiura, Mizuki Yasuda, Hiroki Matsutani

PhySense: Principle-based physics reasoning benchmark reveals LLMs' reasoning flaws in physics problems.

Yinggan Xu, Yue Liu, Zhiqiang Gao et al.

BinConv uses Cumulative Binary Encoding to model ordinal info, significantly improving time series forecasting accuracy.

Andrei Chernov, Vitaliy Pozdnyakov, Ilya Makarov

Object-Centric Concept Bottlenecks (OCB) integrates pretrained object detection and concept discovery to enhance performance and interpretability in complex visual tasks, achieving 68.84% accuracy on COCOLogic.

David Steinmann, Wolfgang Stammer, Antonia Wüst et al.

AReaL employs a fully asynchronous RL system, achieving up to 2.77× training speedup while maintaining or improving model performance, by decoupling generation and training.

Wei Fu, Jiaxuan Gao, Xujie Shen et al.

K2VAE combines KoopmanNet and KalmanNet for efficient long- and short-term probabilistic time series forecasting, outperforming state-of-the-art methods.

Xingjian Wu, Xiangfei Qiu, Hongfan Gao et al.