最新の研究をフォロー
alphaXivは論文・研究者・組織をつなぎ、研究に基づいた回答を提供します。
UC Berkeley
A learned dynamics model lets a humanoid controller match large motion libraries with over 1,000 times fewer GPU-hours than cluster-scale training.
NVIDIA
Combining reusable code-based skill plans with visual foresight helps robots perform complex manipulation across parallel-gripper and dexterous-hand tasks.
KAIST
ETHZ
Forecasting persistent 3D point motion alongside video helps world models generate more coherent multi-view futures and improves robot manipulation in simulated and real settings.
おすすめの研究者
すべて見る
Yann LeCun
Executive Chairman
AMI - Advanced Machine Intelligence, Jacob T. Schwartz Professor, CS @ New York University
Li Fei-Fei
Co-Founder and CEO
World Labs, Founding Co-Director @ Stanford HAI, Sequoia Professor, CS @ Stanford University
研究者の方ですか？ 自分のプロフィールを探す
A single simulated robot policy can reproduce two-handed manipulation demonstrations across new trajectories and held-out objects without fine-tuning.
Princeton
Researchers can use this benchmark to test whether coding agents can improve published systems across computer science, against the authors’ code on held-out workloads.
SJTU
HKU
Giving humanoid robots five key body targets plus a learned motion prior lets them coordinate whole-body movement while completing varied manipulation tasks.
UPenn
Cambridge
Self-supervised image representations can better preserve data geometry and support stronger classification across Euclidean, hyperbolic, and equivariant vision models.
UMD
UW
Agents can use lessons from completed tasks to adapt their prompts and tools mid-run, while keeping each task’s changes from affecting later ones.
Shanghai AI Lab
The agent checks evidence and redirects stalled searches, helping it tackle long investigations without carrying unsupported conclusions forward.
NJU
Tencent
Predicting distant video chunks during training helps autoregressive generators produce more physically plausible continuations without changing how they generate videos.
RUC
Training agents across varied execution setups helps them handle unfamiliar harnesses, including external systems they never encountered during training.
おすすめの研究者
すべて見る
Yoshua Bengio
President and Scientific Director
LawZero, Founder and Scientific Advisor @ Mila - Quebec Artificial Intelligence Institute, Canada CIFAR AI Chair @ CIFAR, Full Professor, CS @ Université de Montréal
NTUA general coding agent can answer spatial questions across images and videos more reliably when measurements persist and arrive in plain language.
Stanford
Strong text-to-image generation can be trained directly from paired images and captions, without internet-scale language pretraining and with substantially less total pretraining compute.
EPFL
Microsoft AI
A diffusion language model that carries its own continuous predictions between steps solves Sudoku and mazes more accurately than prior diffusion models.
Image compression can discard redundant visual detail while keeping depicted subjects recognizable, because token allocation can prioritize semantic fidelity alongside pixel accuracy.
Amazon
A single language model can draft several tokens in parallel and verify them autoregressively, generating text faster without sacrificing comparable quality.
University of Toronto
University of Rochester
Researchers can use 6,191 community-authored riddles across 51 languages to test whether language models apply cultural knowledge, not just recall facts.
USTC
Robots can use execution history to track task progress, while rollout-based revisions improve a planner’s success on history-dependent manipulation tasks.
Tsinghua
The best mix of general and specialized training data depends on when the learning rate decays, so tuning both together can lower adaptation loss.
Predicting upcoming footsteps lets humanoid robots focus terrain sensing on likely landing spots, improving learning speed and traversal of sparse obstacles.
