CHORD framework harmonizes SFT and RL via dynamic weighting, improving LLMs' performance.

Wenhao Zhang, Yuexiang Xie, Yuchang Sun et al.

Proposes a discrete-time linear causal model for latent process detection in Hawkes processes, enabling structure recovery without prior knowledge.

Songyao Jin, Biwei Huang

This review unifies Memory-Augmented Transformers through three taxonomies, but reports no new dataset-level benchmark.

Parsa Omidi, Xingshuai Huang, Axel Laborieux et al.

Pass@k training leverages the Pass@k metric as reward, with analytical derivation of advantage functions, significantly enhancing large models' exploration, outperforming traditional Pass@1 training.

Zhipeng Chen, Xiaobo Qin, Youbin Wu et al.

Πnet introduces an orthogonal projection layer for neural networks, enabling fast, reliable hard constraint satisfaction with orders-of-magnitude training speedup.

Panagiotis D. Grontas, Antonio Terpin, Efe C. Balta et al.

Proposes Noise Hypernetwork to learn reward-biased noise modulation, reducing inference costs while maintaining quality in diffusion models.

Luca Eyring, Shyamgopal Karthik, Alexey Dosovitskiy et al.

Hi-Vec uses hierarchical layers with dynamic selection and task vectors to improve test-time adaptation, boosting robustness against complex distribution shifts.

Sameer Ambekar, Marta Hasny, Laura Daza et al.

This paper systematically evaluates RL techniques in LLM reasoning, revealing that combining group normalization and high clipping surpasses existing strategies.

Zihe Liu, Jiashun Liu, Yancheng He et al.

Proposes a taxonomy for unsupervised VLM adaptation based on unlabeled data availability, covering four paradigms, improving generalization.

Hao Dong, Lijun Sheng, Jian Liang et al.

Theorem 3 proves GELU-like activations achieve Sobolev super-convergence at error rate (NL)^−2(n−m)/d.

Yahong Yang, Juncai He

Goedel-Prover-V2 achieves new breakthroughs in automated theorem proving with scaffolded data synthesis and self-correction, reaching 88.1% accuracy on MiniF2F.

Yong Lin, Shange Tang, Bohan Lyu et al.

EvaDrive employs adversarial multi-objective reinforcement learning, achieving 94.9 PDMS on NAVSIM and diverse driving styles without external preferences.

Siwen Jiao, Kangan Qian, Hao Ye et al.

RecoMind framework optimizes in-session user satisfaction in recommendation systems using RL, boosting video watch time by 15.81%.

Mehdi Ben Ayed, Fei Feng, Jay Adams et al.

Introduced a causal explanation method for concept drift, enhancing model actionability.

David Komnick, Kathrin Lammers, Barbara Hammer et al.

Proposes TP-GRPO with process rewards, achieving +4.32% to +6.67% in Pass@1 on math reasoning models with fewer samples.

Tao He, Rongchuan Mu, Lizi Liao et al.

Spec-VLA employs relaxed acceptance speculative decoding to accelerate VLA models, achieving 1.42× speedup with 44% longer acceptance length.

Songsheng Wang, Rucheng Yu, Zhihang Yuan et al.

W-CFM uses Gibbs kernel weighting to approximate Entropic OT, improving path straightness and sample quality efficiently.

Sergio Calvo-Ordonez, Matthieu Meunier, Alvaro Cartea et al.

NeuCoReClass AD integrates multi-task self-supervised learning with neural transformations, boosting time series anomaly detection by 15% on benchmarks.

Aitor Sánchez-Ferrera, Usue Mori, Borja Calvo et al.

This paper analyzes provable in-context learning of Transformers on nonlinear L-Lipschitz regression, highlighting the impact of Lipschitz constant L on convergence dynamics.

Hongbo Li, Lingjie Duan, Yingbin Liang

Introduces SLDAgent, an evolution-based AI system that autonomously discovers scaling laws surpassing human expertise, validated on SLDBench with high accuracy.

Haowei Lin, Haotian Ye, Wenzheng Feng et al.