Graph-GRPO introduces analytical transition probabilities and reinforcement learning for efficient graph flow model training, achieving 95% validity in 50 steps.

Baoheng Zhu, Deyu Bo, Delvin Ce Zhang et al.

MIPO enhances LLM performance by maximizing mutual information between prompts and responses without additional data.

Hyunji Nam, Haoran Li, Natasha Jaques

TTC layer enhances LLM reasoning via LQR planning, achieving 27.8% improvement on MATH-500.

Peihao Wang, Shan Yang, Xijun Wang et al.

Entropy-Aware On-Policy Distillation balances mode-seeking and diversity, boosting math reasoning accuracy by +1.37 to +5.05 on benchmarks.

Woogyeol Jin, Taywon Min, Yongjin Yang et al.

Swimba enhances SSM capacity by mixing expert streams in parameter space, maintaining computational efficiency.

Zhixu Du, Krishna Teja Chitty-Venkata, Murali Emani et al.

Proposes a design procedure-based prompting strategy using LLMs to generate industrial-grade CAD programs with high geometric diversity for data augmentation.

Yan-Ying Chen, Dule Shu, Matthew Hong et al.

Proposes DistriVoting with GMM and SelfStepConf to improve confidence calibration, outperforming SOTA by 3.5%.

Xizhong Yang, Haotian Zhang, Huiming Wang et al.

Strategy-Guided Exploration (SGE) enhances RL in LLMs by generating high-level natural language strategies, improving exploration and task success rates.

Andrew Szot, Michael Kirchhof, Omar Attia et al.

Modular memory combines IWL and ICL to solve forgetting in continual learning.

Vaggelis Dorovatas, Malte Schwerin, Andrew D. Bagdanov et al.

Gome employs gradient-based optimization for ML engineering tasks, achieving 35.1% medal rate on MLE-Bench, surpassing tree search methods.

Yifei Zhang, Xu Yang, Xiao Yang et al.

VACO employs advantage alignment and TV filtering to mitigate policy lag in asynchronous RL, improving robustness and convergence.

Homayoun Honari, Roger Creus Castanyer, Michael Przystupa et al.

ReGFT leverages human reference solutions to synthesize positive trajectories, significantly alleviating reward sparsity and boosting RL-based mathematical reasoning performance.

Yangzhen Wu, Shanda Li, Zixin Wen et al.

SPLARE introduces a sparse latent retrieval model using pretrained SAEs, outperforming vocabulary-based methods in multilingual and out-of-domain tasks.

Thibault Formal, Maxime Louis, Hervé Dejean et al.

SOTAlign leverages minimal image-text pairs and optimal transport to achieve semi-supervised vision-language alignment, outperforming existing methods significantly.

Simon Roschmann, Paul Krzakala, Sonia Mazelet et al.

EvoX employs meta-evolution of search strategies, outperforming AlphaEvolve on nearly 200 tasks with significant efficiency gains.

Shu Liu, Shubham Agarwal, Monishwaran Maheswaran et al.

Applying Takeuchi Information Criterion (TIC) to evaluate DNN generalization near NTK regime; experiments show high correlation with actual gaps.

Hiroki Naganuma, Taiji Suzuki, Rio Yokota et al.

RAIN-Merging is a gradient-free method that enhances instruction-following in large reasoning models while preserving their thinking format, with over 8% improvement in benchmarks.

Zhehao Huang, Yuhang Liu, Baijiong Lin et al.

Proposes RoPE-ID method, enhancing long input handling by applying RoPE at high frequency to a subset of channels.

Davis Wertheimer, Aozhong Zhang, Derrick Liu et al.

VCDF framework enhances robustness of time series causal discovery by stability validation across data blocks, improving VAR-LiNGAM's F1 score by ~0.1.

Gene Yu, Ce Guo, Wayne Luk

L2G-Net employs Cauchy decomposition to enable efficient spectral graph transforms, capturing long-range dependencies with quadratic complexity, outperforming traditional full eigendecomposition.

Samuel Fernández-Menduiña, Eduardo Pavez, Antonio Ortega