Proposes a feature learning theory for single- and multi-modal contrastive learning, highlighting signal-to-noise ratio (SNR) as key to generalization.

Wei Huang, Andi Han, Yongqiang Chen et al.

MM-Embed leverages multimodal large models for universal retrieval, surpassing SOTA with 52.7% R@5 on M-BEIR.

Sheng-Chieh Lin, Chankyu Lee, Mohammad Shoeybi et al.

This paper introduces a nonlinear transformer framework with MLP layers that efficiently learns low-dimensional target functions in-context, requiring prompt length dependent only on the target function's subspace dimension r, not the ambient d.

Kazusato Oko, Yujin Song, Taiji Suzuki et al.

Proposes a recurrent-based variable-length image tokenizer, adapting token count (32-256) to image complexity for efficient compression.

Shivam Duggal, Phillip Isola, Antonio Torralba et al.

CRMArena benchmarks LLMs on nine realistic CRM tasks using Salesforce-based data; top models achieve less than 58% success, highlighting room for improvement.

Kung-Hsiang Huang, Akshara Prabhakar, Sidharth Dhawan et al.

Hunyuan-Large is Tencent's open-source MoE model with 52 billion activated parameters, handling up to 256K tokens.

Xingwu Sun, Yanfeng Chen, Yiqing Huang et al.

Proposes SlowFast framework with a state space model for 2ms low-latency speech enhancement, reducing computation by 70%.

Longbiao Cheng, Ashutosh Pandey, Buye Xu et al.

TAMA uses large multimodal models for few-shot time series anomaly detection, improving accuracy.

Jiaxin Zhuang, Leon Yan, Zhenwei Zhang et al.

ProD-RL leverages reinforcement learning to hierarchically decompose proofs, improving theorem proving success rate from 40.8% to 45.5%.

Kefan Dong, Arvind Mahankali, Tengyu Ma

Defines SLM by task capability and resource limits; analyzes techniques, applications, and future directions.

Fali Wang, Zhiwei Zhang, Xianren Zhang et al.

Achieving domain-independent certified robustness via knowledge continuity, independent of input domains.

Alan Sun, Chiyu Ma, Kenneth Ge et al.

Introduces 'Diversity Progress' to enhance goal selection in RL, accelerating skill learning.

Erik M. Lintunen, Nadia M. Ady, Christian Guckelsberger

Outer-PPO introduces non-unit learning rates and momentum, improving performance by 5-10% in complex environments.

Charlie B. Tan, Edan Toledo, Benjamin Ellis et al.

Proposed a resolvent-type method using finite-dimensional approximation to learn generators of unknown continuous-time dynamical systems.

Yiming Meng, Ruikun Zhou, Melkior Ornik et al.

EVSNet leverages event-based motion features with a lightweight fusion framework, achieving 34.1% mIoU on low-light VSPW, 11× parameter efficiency over SOTA.

Zhen Yao, Mooi Choo Chuah

Study reveals VLMs' limitations in multi-object reasoning tasks due to the binding problem.

Declan Campbell, Sunayana Rane, Tyler Giallanza et al.

EgoMimic leverages egocentric human videos and 3D hand tracking with cross-domain alignment and joint training, boosting manipulation task success by 34-228%.

Simar Kareer, Dhruv Patel, Ryan Punamiya et al.

This study enhances offline RL agents using diverse, informative language feedback, achieving over 10% performance gains on benchmarks.

Jiajun Xi, Yinong He, Jianing Yang et al.

TabM employs parameter-efficient ensembling of MLPs, achieving superior tabular data deep learning performance.

Yury Gorishniy, Akim Kotelnikov, Artem Babenko

NoPoSplat reconstructs 3D Gaussian scenes from sparse unposed images in real-time, surpassing pose-required methods.

Botao Ye, Sifei Liu, Haofei Xu et al.