Follow the latest research
alphaXiv connects papers, researchers, and organizations, grounding its answers in the underlying work.
FDU
The development record shows agents can propose methods and execute revisions, while human researchers retain responsibility for goals, evaluation, and research direction.
NVIDIA
Critic-free asynchronous training can learn from one rollout per prompt while remaining stable under substantial policy lag and preserving tool use.
Google
UMD
Replaying discovery histories lets agents improve exploration strategies cheaply before deploying them, reducing the cost of long-horizon algorithmic and scientific search.
Researchers to follow
View all
Are you a researcher? Find your profile
A recurrent decoder carries computation across every prompt and response token, while exact full-history replay keeps reinforcement-learning policy states aligned with current parameters.
Periodic LabsPeriodic Labs describes training an AI system on laboratory data to automate difficult, multiphase XRD interpretation and expand autonomous materials experiments.
UC San Diego
UIC
Agents can adapt to unfamiliar software by exploring, verifying outcomes, and reusing environment-specific memory without changing model parameters.
UPenn
Harvard
Sophia TangShiyi WangA teacher-free transport map enables language models to generate and iteratively self-correct complete sequences in one or a few evaluations.
University of Montreal
AI2
Adaptive resampling shifts reinforcement-learning compute from already-solved prompts toward harder ones, improving language models’ ability to solve difficult math and coding tasks.
Google Research
CMU
A staged multi-agent workflow lets language models explore, decompose, verify, and revise research proofs while preserving objections and partial progress.
Predicting structured motion, semantic, and geometric futures before actions improves robot manipulation and lets actionless human videos strengthen learning.
At equal response budgets, batch or group centering should be chosen from gradient structure and baseline quality, not difficulty distribution shape alone.
Researchers to follow
View all
Ion Stoica
Co-Founder & Executive Chairman
Anyscale, Co-Founder & Executive Chairman @ Databricks, Professor, CS @ UC Berkeley
Andrew Ng
Managing Partner
AI Aspire, Managing General Partner @ AI Fund, Founder @ DeepLearning.AI, Adjunct Professor, CS @ Stanford University, Chairman and Co-Founder @ Coursera
Controlled small-scale experiments show that per-token recurrence adds substantial training cost without improving language-model quality or long-range prediction.
The framework enables AI systems to revise research questions, representations, and experiments as external evidence reshapes what remains unknown.
UC Berkeley
UW
Across open-ended benchmarks, agents’ marginal gains eventually fall below independent sampling, enabling Elo curves to identify when parallel sessions become more effective.
Researchers can turn interactive scientific workflows into reusable tasks and rubrics that continually improve an agent’s procedures and problem-solving coverage.
SJTU
A shared simulation interface lets researchers compare visuo-tactile bimanual manipulation policies across diverse dexterous hands and controlled scene shifts.
Stanford
Benign-sounding reasoning injected into context can steer models toward harmful plans that they paraphrase as their own, evading chain-of-thought monitors.
UT Austin
Preference-based reward design can align reinforcement-learning objectives without costly or unsafe environment rollouts, using imagined trajectories and language-model-designed features.
MIT
Training-free value smoothing and direct probability encoding make low-bit attention both more faithful and faster for video diffusion models.
Robots can reuse interaction outcomes and fine-grained visual memories across rooms and tasks, avoiding repeated exploration during long-horizon manipulation.
