The paper proposes RLVR, counterfactual events, and multi-source data; Polymarket training reduced R1-14B Brier score from 0.214 to 0.197.

Sang-Woo Lee, Sohee Yang, Donghyun Kwak et al.

Flow Matching (FM) offers efficient, stable generative modeling for biological sequences, molecules, and proteins, outperforming traditional methods in speed and accuracy.

Zihao Li, Zhichen Zeng, Xiao Lin et al.

TorchAO is a PyTorch-native framework supporting FP8, QAT, PTQ, and sparsity, enabling end-to-end model optimization from training to deployment.

Andrew Or, Apurva Jain, Daniel Vega-Myhre et al.

GUI-G² models GUI elements as Gaussian distributions, replacing sparse binary rewards with dense continuous signals, achieving 24.7% accuracy gain on ScreenSpot-Pro.

Fei Tang, Zhangxuan Gu, Zhengxi Lu et al.

This study reveals models transmit behavioral traits via hidden signals in unrelated data, highlighting risks for AI safety and model control.

Alex Cloud, Minh Le, James Chua et al.

Leveraging asynchronous cross-border market prices via ensemble models improves day-ahead electricity price forecasts in Belgium by 22% and Sweden by 9%.

Maria Margarida Mascarenhas, Jilles De Blauwe, Mikael Amelin et al.

This study evaluates whether non-equivariant CNNs trained with rotation augmentation can learn equivariance, showing comparable performance to equivariant models in molecule tasks.

Ewa M. Nowara, Joshua Rackers, Patricia Suriana et al.

Overtone employs cyclic patch size modulation via CSM and CKM modules, reducing spectral harmonic errors by up to 40%, enabling flexible, compute-adaptive PDE surrogates.

Payel Mukhopadhyay, Michael McCabe, Ruben Ohana et al.

IPPRO introduces a scale-invariant pruning criterion based on projective geometry, outperforming magnitude-based methods in neural network compression.

Jaeheun Jung, Jaehyuk Lee, Yeajin Lee et al.

ABench-Physics evaluates LLMs' physical reasoning via high-difficulty dynamic problems, revealing significant performance gaps.

Yiming Zhang, Yingfan Ma, Yanmei Gu et al.

Conformal Information Pursuit (C-IP) leverages prediction set sizes to estimate uncertainty, optimizing multi-round queries for large language models, improving efficiency and interpretability.

Kwan Ho Ryan Chan, Yuyan Ge, Edgar Dobriban et al.

Combines fast experiments with Bayesian optimization to enhance long-term outcomes, reducing experiment time by 60%.

Qing Feng, Samuel Daulton, Benjamin Letham et al.

Proposed a variational flow matching (VFM) framework for controlled molecular generation, incorporating equivariance to ensure invariance under rotations, translations, and permutations, achieving state-of-the-art results.

Floor Eijkelboom, Heiko Zimmermann, Sharvaree Vadgama et al.

RLPR leverages LLM's token probabilities as rewards, eliminating the need for verifiers, significantly boosting reasoning capabilities.

Tianyu Yu, Bo Ji, Shouli Wang et al.

TabArena is a continuously maintained benchmark for tabular ML, combining curated datasets and models, advancing evaluation standards.

Nick Erickson, Lennart Purucker, Andrej Tschalzev et al.

VideoPDE employs video diffusion transformers to solve PDEs as high-fidelity video inpainting, outperforming state-of-the-art methods.

Edward Li, Zichen Wang, Jiahe Huang et al.

ERL-VLM employs absolute ratings from large VLMs to improve reward learning efficiency in RL.

Tung Minh Luu, Younghwan Lee, Donghoon Lee et al.

Uni-DPO enhances LLM performance by dynamically reweighting samples, outperforming Claude 3 Opus by 6.7 points on Arena-Hard.

Shangpin Peng, Weinong Wang, Zhuotao Tian et al.

Identifies three common pitfalls in KL gradient estimation for RL, emphasizing the correct full-sequence approach.

Yunhao Tang, Rémi Munos

LPO combines information entropy and dynamic distance rewards to enhance GUI interaction accuracy, achieving state-of-the-art results.

Jiaqi Tang, Yu Xia, Yi-Feng Wu et al.