HarvardMIT

Yilun Du

A single model can generate video and infer depth and camera motion from partial observations, including just two frames.

Northwestern UniversityCMU

Qineng WangLi Fei-Fei

The benchmark lets researchers compare coding agents across diverse simulated robots and realistic sensing constraints, revealing how rarely today’s strongest agents succeed without privileged simulator access.

Hao KangWeili XuIon Stoica

Adapting compute allocation and cache reuse to agent behavior cuts request latency while helping LLM servers handle changing workloads without costly restarts.

Kempner InstituteUIUC

Audrey HuangNan JiangSham M. Kakade

Resetting the regularization reference as policies improve can prevent training from losing access to promising responses, enabling efficient convergence from capable base models.

SEUCUHK

Qiyuan ZhuangQi DOU

LeaP, a learnable, proprioception-conditioned stochastic prior, improves generative robot policies by replacing the standard Gaussian action generation source with an informed yet uncertain starting point. This method achieved an 81.6% average success rate across 15 simulated manipulation tasks and 80.0% in real-world experiments, outperforming existing baselines with minimal overhead.

Georgia TechHarvard

Bo DaiYilun Du

This tutorial connects generative modeling, representation learning, and planning through a shared framework, showing how separately learned models can be combined for new tasks.

TsinghuaWestlake University

Jun Zhu

Sharing the generator with its fake-distribution estimator reduces training memory and iteration time while preserving strong instruction following in long video generation.

UC Berkeley

Ion Stoica

A lightweight layer over existing agent harnesses can cut end-to-end execution costs by up to 35% while largely preserving task accuracy.

PKUCUHK

Shanghang Zhang

Robots can use predicted interaction cues and task progress to guide precise, long-horizon manipulation across simulated and real-world tasks.

Oxford

Minghao ChenPhilip Torr

Video world models can extract actions from unlabeled footage and reproduce demonstrated movements in new scenes, improving transfer across robotics, games, and driving.

Shanghai AI Lab

Weida WangDahua Lin

By tracking evidence, assumptions, and unresolved questions, an AI research system can adapt its next investigations and connect proposals to experimental validation.

Princeton

William Bialek

Biological systems can perform near optimally despite substantial variation in their underlying parameters, because many parameter changes barely affect performance.

KAIST

Sung Ju Hwang

Linking reusable lessons to their original agent runs lets agents check questionable guidance, improving task success while reducing failures caused by memory.

ZJUUCAS

Shanghang Zhang

Human videos without robot-action labels can improve robot policies’ performance on unfamiliar tasks, with greater gains as video training data increases.

Shanghai AI LabFDU

Jiahang LinXuanjing Huang

Research agents often satisfy the immediate request while missing constraints carried over from earlier decisions and existing project files.

SJTUShanghai Innovation Institute

cewu lu

Combining egocentric videos with human manipulation demonstrations helps robots control objects in unfamiliar scenes, extending video pretraining toward generalizable robot actions.

Qinfeng ZhuWeiguang Zhao

A visual safety layer that reviews robot goals, actions, and critical operations reduced harm across tested models while largely preserving task completion.

Texas A&M UniversityJHU

Daniel Khashabi

Models can recognize promising research ideas when given candidates, but often cannot generate them, suggesting external ideas may improve automated research systems.

NYUAMI Labs

Yann LeCun

A learned score signal lets self-supervised image models preserve diverse representations without explicit density calculations, while also guiding label-free adaptation to image corruptions.

CUHKSZUIUC

Zhizheng Wu

The 229,000-hour multilingual corpus pairs minute-long speech and dialogue with speaker, style, and vocalization annotations, enabling training and evaluation of controllable speech generation.

Georgia TechUIUC

Hao KangWeili XuHanchen Li

By adapting compute allocation and preserving reusable context as agent workloads shift, this serving system lowers request latency across coding, math, and computer-use tasks.