Proposes bisimulation encoder for invariant visual representations, reducing latent size by 10×, enhancing robustness of JEPAs against background variations.

Leonardo F. Toso, Davit Shadunts, Yunyang Lu et al.

Information-theoretic analysis plus Oracle RL and conditional-MI rewards improves CoT monitorability and reduces reward hacking.

Usman Anwar, Tim Bakker, Dana Kianfar et al.

Proposes sinusoidal positional encoding to control symmetry breaking in graph flow models, balancing training speed and generalization.

Benjamin Honoré, Alba Carballo-Castro, Yiming Qin et al.

Gradient regularization (GR) mitigates reward hacking, improves reward model accuracy, outperforming KL penalty in RLHF tasks.

Johannes Ackermann, Michael Noukhovitch, Takashi Ishida et al.

Proposes MantisV2, using synthetic data and test strategies to close the zero-shot time series classification gap.

Vasilii Feofanov, Songkang Wen, Jianfeng Zhang et al.

Grassmannian MoE (GrMoE) uses matrix Bingham distribution's concentration parameters to smoothly control routing entropy, preventing expert collapse.

Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma

Reverso model achieves efficient zero-shot time series forecasting using long convolutions and linear RNN layers, reducing parameters by 100x.

Xinghong Fu, Yanhong Li, Georgios Papaioannou et al.

Flow Expander (FE) employs verifier-constrained entropy maximization via mirror descent to expand pre-trained flow models’ valid design space.

Riccardo De Santi, Kimon Protopapas, Ya-Ping Hsieh et al.

GLM-5 integrates DSA and asynchronous RL, achieving SOTA in long-text understanding and code generation with 20% improvement.

GLM-5-Team, :, Aohan Zeng et al.

Flowers employs multi-head pointwise warps for PDE solution operators, outperforming similar-sized models with only 17M parameters, achieving state-of-the-art results.

Till Muser, Alexandra Spitzer, Matti Lassas et al.

Introduces event-level augmentation protocol for IoT anomaly detection, evaluating 14 models under simulated real-world perturbations, revealing regime-dependent robustness.

Dmitry Zhevnenko, Ilya Makarov, Aleksandr Kovalenko et al.

This paper introduces size transferability of Graph Transformers using convolutional positional encodings, enabling models trained on small graphs to generalize to larger ones.

Javier Porras-Valenzuela, Zhiyang Wang, Xiaotao Shang et al.

BindCLIP combines contrastive learning with diffusion-based pose generation, enhancing interaction sensitivity and out-of-distribution generalization in virtual screening.

Anjie Qiao, Zhen Wang, Yaliang Li et al.

Proposes Categorical Flow Maps, leveraging continuous paths for rapid few-step categorical data generation with state-of-the-art results.

Daan Roos, Oscar Davis, Floor Eijkelboom et al.

SafeNeuron identifies safety neurons via ES and SAS metrics, freezes them, and fine-tunes remaining parameters, boosting robustness against neuron pruning with minimal utility loss.

Zhaoxin Wang, Jiaming Liang, Fengbin Zhu et al.

Proposes G-OPD framework using reward extrapolation to surpass teacher performance, validated on math reasoning and code tasks.

Wenkai Yang, Weijie Liu, Ruobing Xie et al.

NRT model enhances reasoning by self-generating reasoning paths without external verifiers, significantly improving complex reasoning tasks.

Yuanfu Wang, Zhixuan Liu, Xiangtian Li et al.

Krause Attention reduces attention sinks by promoting local synchronization, improving performance and reducing complexity.

Jingkun Liu, Yisong Yue, Max Welling et al.

Proposed retrieval-aware distillation retains only 2% of attention heads, recovering 95% performance.

Aviv Bick, Eric P. Xing, Albert Gu

pplx-embed combines diffusion pretraining with contrastive curricula, reaching 69.66 MTEB and 81.96 ConTEB nDCG@10.

Sedigheh Eslami, Maksim Gaiduk, Markus Krimmel et al.