Follow the latest research
alphaXiv connects papers, researchers, and organizations, grounding its answers in the underlying work.
CMU
USC
Training looped language models to settle into fixed points enables terminal attention-cache sharing with little accuracy loss, reducing decoding memory as recurrence depth grows.
Harvard
Sampling expert examples to better match a model’s distribution lets supervised fine-tuning learn new skills while often preserving prior capabilities better than on-policy methods.
Stanford
Author-validated judgments from 184 researchers let teams test whether AI can find papers that genuinely inspire research, not merely match its topic.
Researchers to follow
View 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
MITMultimodal agents can solve unfamiliar visual games more effectively when they can revisit past images and inspect details as they reason.
Sensetime
Tsinghua
Repeating shared Transformer blocks lets text-to-image models refine visual constraints internally, improving generation quality without adding model parameters.
UC Berkeley
Robots can improve reusable manipulation skills through simulated practice and failure feedback, raising success across 22 tasks before deployment on physical hardware.
OpenHands
UT Austin
Agents can improve reusable instructions from past successes and failures, enabling skill refinement when task replays are costly or impossible.
Apple
Using an earlier recurrent prediction to guide token selection lets looped language models improve accuracy or match full-depth performance with fewer iterations.
UW–Madison
Pre-trained language models can solve agentic tasks that post-trained versions miss when given enough attempts, revealing a coverage cost to post-training’s higher reliability.
NJU
Shanghai AI Lab
A streaming video model can retain time-stamped descriptions of past events, letting it answer later questions without revisiting those frames or losing recent visual context.
Training robots alongside human egocentric videos improves out-of-distribution manipulation as data scales, revealing a path to transfer physical knowledge across embodiments.
Researchers to follow
View all
Andrew Ng
Managing Partner
AI Aspire, Managing General Partner @ AI Fund, Founder @ DeepLearning.AI, Adjunct Professor, CS @ Stanford University, Chairman and Co-Founder @ Coursera
UMD
LLNL
A single autoencoder can generate sharp ImageNet images in one pass, offering a simpler, faster alternative to many-step diffusion models.
A single motion model can predict 3D scenes, generate robot actions, and retarget human movement by conditioning on whichever trajectories are known.
Google
This survey maps how tokenization shapes language-model quality, multilingual fairness, efficiency, and security, giving researchers a guide to methods, evaluation, and open problems.
Columbia
NYULearning which directions matter to a network’s outputs preserves language-model quality under aggressive compression, where conventional low-rank methods often fail.
Coding agents that learn when to summarize stale history and what to preserve solve more repository tasks, even before their context windows fill.
DeepMind
Early testers can use Argon to find and validate serious software vulnerabilities, including a flaw exposing hospital patients’ sensitive information that earlier frontier models missed.
Adaption Labs
Generating training data from task descriptions alone could help researchers build high-quality, diverse datasets for capabilities with no existing examples.
Biological models can score well without relying on sequence data, so benchmark accuracy alone cannot show whether predictions reflect biological inputs.
UC San Diego
Robots can turn failed manipulation attempts into reusable recovery skills, improving task success and reducing the need for human intervention.
