Follow the latest research
alphaXiv connects papers, researchers, and organizations, grounding its answers in the underlying work.
Oxford
Martian
This survey maps what mechanistic interpretability can reveal about model internals, proposes shared criteria for evaluating its claims, and clarifies where safety benefits remain unproven.
Hugging FaceResearchers can use this hands-on guide to diagnose MoE training failures, compare design choices, and carry practical recipes from cluster setup through serving and reinforcement learning.
UPenn
KAISTThis survey maps recent advances in continuous text generation, helping researchers compare design choices and identify open challenges in scaling, control, and safety.
Researchers to follow
See all
Are you a researcher? Find your profile
AnthropicThe release demonstrates that AI agents can move beyond software, designing functional electronic devices and controlling robots in physical-world tasks.
NUS
MITRobots can choose improvised tools from unfamiliar objects and locate the regions needed to use them, enabling tool use from a single RGB-D view.
NVIDIA
JHURobots can generate detailed spatial plans in fewer sequential decoding steps, with VFlash improving 3D trajectory accuracy while reducing generation latency.
Stanford
CMUA real-world dataset pairs human demonstrations of improvised tool use with multi-view annotations, enabling researchers to study how robots can adapt tools to diverse targets.
Brown UniversityA swarm of small language-model agents can discover stronger mathematical and scientific programs in minutes, using one GPU instead of frontier models.
UCF
Microsoft ResearchA coding agent can turn text prompts into editable 3D scene drafts that guide video generators toward more physically plausible motion and composition.
UIUC
UT AustinFlashState lets hybrid-model servers resume agent requests at exact prior endpoints, reducing repeated prefill and sustaining throughput as concurrent workloads strain cache capacity.
MATS
DeepMindTraining against monitors can teach language models to evade detection, but carefully designed interventions can reduce reward hacking across coding, medical advice, and biography tasks.
Researchers to follow
See all
Duke University
Scoring latent reasoning states lets models refine answers, stop when progress stalls, and choose among solutions using one learned objective.
Princeton
Turning papers into reproducible, scored research environments lets researchers test AI scientists on real experiments and use their traces to train future systems.
AppleGeneral-purpose AI models can control robots directly, completing diverse manipulation tasks without demonstrations or task-specific training, and adapting actions as challenges arise.
BAAI
TsinghuaTraining robots to align visual features with 3D geometry improves manipulation across changing scenes and robot embodiments without adding inference-time components.
NTU
Robot policies that retain access to intermediate visual features can control objects more robustly under changes in camera views and scene appearance.
EPFL
Offline distillation can prepare language models to benefit from on-policy teacher feedback, improving held-out math accuracy even when preparation alone adds little.
ZJUA consistent measure of 3D point spacing helps image-based scene reconstruction generalize across different scene scales and camera viewpoints.
Allocating visual tokens between broad video coverage and targeted high-resolution details improves question answering across long-video benchmarks without retraining the model.
AI agents can discover quantum algorithms for simulations beyond classical testing, while mathematical proofs certify their correctness and resource costs.
CornellContinuously auditing coding-agent benchmarks catches broken or unfair tasks, helping researchers distinguish agents’ capabilities from errors in the evaluation.
