Proposes Token Assorted, mixing latent and text tokens to improve reasoning, reducing trace length by 17% and boosting performance by up to 13.3%.

DiJia Su, Hanlin Zhu, Yingchen Xu et al.

CARROT predicts model cost and accuracy, achieving minimax optimal routing; outperforms baselines on SPROUT dataset.

Seamus Somerstep, Felipe Maia Polo, Allysson Flavio Melo de Oliveira et al.

iVISPAR uses a sliding geometric puzzle to evaluate VLM spatial reasoning across 3D, 2D, and text modalities, highlighting current limitations.

Julius Mayer, Mohamad Ballout, Serwan Jassim et al.

Unified intrinsic motivation as constrained maximum entropy inference, integrating active inference, empowerment, and maximum occupancy.

Alex B. Kiefer

Mosaic3D introduces a novel pipeline and model, generating 5.6M high-quality 3D mask-text pairs, achieving SOTA open-vocabulary 3D segmentation.

Junha Lee, Chunghyun Park, Jaesung Choe et al.

Flow Q-Learning (FQL) uses flow matching to model complex action distributions, avoiding recursive backpropagation, and achieves state-of-the-art results on 73 tasks.

Seohong Park, Qiyang Li, Sergey Levine

Proposes a similarity evaluation framework for pretraining data selection, revealing off-the-shelf embeddings underperform compared to specialized models.

Dylan Sam, Ayan Chakrabarti, Afshin Rostamizadeh et al.

MotionLab introduces a unified human motion generation and editing framework based on the Motion-Condition-Motion paradigm, leveraging rectified flows for multi-task learning, achieving state-of-the-art results.

Ziyan Guo, Zeyu Hu, De Wen Soh et al.

AStar uses retrieved thought cards to boost training-free multimodal reasoning, reaching 53.9% on MathVerse.

Jinyang Wu, Mingkuan Feng, Guocheng Zhai et al.

DIME integrates diffusion models into MaxEnt RL, greatly enhancing high-dimensional control performance.

Onur Celik, Zechu Li, Denis Blessing et al.

Soup-of-Experts method enables pretraining of specialist models via parameter averaging for quick deployment.

Pierre Ablin, Angelos Katharopoulos, Skyler Seto et al.

Introduces a non-isotropic wrapped Gaussian on SPD manifolds using exponential maps, enhancing geometric modeling for high-dimensional data.

Thibault de Surrel, Fabien Lotte, Sylvain Chevallier et al.

PRIME introduces implicit process rewards for online reward model updates, boosting multi-step reasoning by 15.1% on benchmarks.

Ganqu Cui, Lifan Yuan, Zefan Wang et al.

Proposes InfoBridge, an unbiased mutual information estimator using diffusion bridge matching, effective for high-dimensional complex data.

Sergei Kholkin, Ivan Butakov, Evgeny Burnaev et al.

BVINet achieves blind video inpainting with zero annotations, significantly improving restoration performance.

Zhiliang Wu, Kerui Chen, Kun Li et al.

ConditionNET uses vision-language transformers to learn action preconditions and effects, boosting real-time robot execution monitoring.

Daniel Sliwowski, Dongheui Lee

CRISP leverages sequential convex programming to robustly solve contact-implicit motion planning without constraint qualifications, achieving high success rates even from naive initializations.

Yulin Li, Haoyu Han, Shucheng Kang et al.

CAIMAN enhances sample efficiency in legged robots using causal action influence, significantly improving object-pushing skills.

Yuanchen Yuan, Jin Cheng, Núria Armengol Urpí et al.

Proposes a dynamic KL-weighted diffusion model guided by large language models, achieving superior text-to-image synthesis with FID 30.5 and human score 4.6.

Julian Perry, Frank Sanders, Carter Scott

Proposes an efficient Orlicz-Sobolev transport method for unbalanced measures on graphs, using binary search for fast computation.

Tam Le, Truyen Nguyen, Hideitsu Hino et al.