Follow the latest research
alphaXiv connects papers, researchers, and organizations, grounding its answers in the underlying work.
StanfordLearning visual tokenizers alongside image generators can improve image generation at equal training compute compared with generating directly from pixels.
CMU
Robots can adapt world-model predictions to hidden physical conditions from one or two interactions, improving action choices in simulated and real-world tasks.
A loss learned from image captions improves text-to-image generation by emphasizing semantically meaningful differences while preserving pixel-level detail.
Researchers to follow
See all
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
Are you a researcher? Find your profile
CUHK
UMich
Learning from research histories helps AI agents produce more testable scientific hypotheses, with improvements concentrated in falsifiability and evidence planning.
GPT-6-Astra approaches saturation on several robot benchmarks, but real-world trials reveal that reliable, efficient deployment remains a central challenge.
UC Berkeley
Overlapping forecast errors can make online ensembles chase noise; scaling updates to account for that overlap lets them adapt across horizons without dataset-specific tuning.
Harvard
NVIDIA
Robot agents acquire manipulation skills through online trials and turn successful strategies into reusable programs, avoiding teleoperation data for many tested tasks and robot types.
Agents can prepare reusable, linked workspaces from document collections, helping language models answer future cross-document questions without searching from scratch each time.
Tsinghua
Ant Group
A compact, accumulating 3D scene memory helps video generators maintain visual and geometric coherence across hundreds of frames and large viewpoint changes.
NYUObject-navigation agents often reach a named object but fail to stop there; explicitly estimating its visibility and distance improves success across six benchmarks.
In small dialogue experiments, a curiosity-driven system learned and later recalled word definitions from a single teaching exposure, without pretrained language models.
Researchers to follow
See all
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
The smallest finite sets meeting every lattice direction have sharply separated sizes, and each extremal size forces a specific geometric shape.
Decentralized agent swarms can learn both how to act and when to participate, improving performance across crafting, robot navigation, and research tasks.
Z.ai
Image-generation training can teach multimodal models facts absent from captions, but those facts often remain difficult to retrieve through text-only questions.
Synchronized reference-camera videos let a world model track moving objects beyond the current view, so they reappear consistently as the camera moves.
Reconstruction pretraining lets generators learn high-quality one-step image synthesis through simpler distributional training, without relying on a pretrained diffusion model.
SJTU
Reusing the target model’s contextual representations lets speculative decoding handle long, evolving agent conversations without maintaining a separate draft-model history.
PKU
A robot that tracks semantic task progress can monitor whether each step succeeded, revise its plan, and complete longer manipulation sequences more reliably.
WebCoder: Scaling General Frontend Development with Artifact-Grounded Agentic Reinforcement Learning
Training agents against an application's evolving, verified state enables them to coordinate design reproduction, feature building, and repairs while preserving existing behavior.
Tencent
Branching agent rollouts let language models learn which interaction turns help, while bounded targets keep updates stable across sequential training batches.
