Follow the latest research
alphaXiv connects papers, researchers, and organizations, grounding its answers in the underlying work.
Cursor
UC BerkeleyTextual feedback can teach coding agents hard-to-discover tools and styles while preserving coding performance during reinforcement learning.
Physical IntelligenceRobotic policies can improve beyond pretrained behaviors while retaining their useful guidance, achieving high success across 19 tasks on four benchmarks.
Ricursive IntelligenceResearchers can access an open-weight 65-billion-parameter model for agentic coding, trained from scratch at a reported $32 million development cost.
Researchers to follow
See all
Are you a researcher? Find your profile
Together AIA persistent inference engine lets optimization agents test 252 times more candidates, helping deployments keep pace with changing models and traffic.
NJU
Robots can improve demonstration-trained generative policies through task-reward feedback, with forward-process updates supporting manipulation in simulation and on physical robots.
StanfordGiving weak models partial solutions creates useful accuracy trends, letting researchers forecast larger models’ benchmark performance even when smaller models score near zero.
RUC
Training on tasks built from execution-witnessed tool sequences substantially improves agents’ ability to coordinate tools across environments over long workflows.
NVIDIALanguage models can maintain near-baseline accuracy on many long-running tasks with short contexts by compacting their history or selectively reading source documents.
The post demonstrates that a general-purpose language model can inspect network weights to infer behaviors and directly edit them to fix specific failures.
Subconscious
MITTraining draft models to trigger fresh, more precise decoding at key decisions improves coding agents’ long-horizon task success without changing their output distribution.
KAISTHuman videos can guide robots to reproduce demonstrated motions without paired human–robot training data or action labels.
Researchers to follow
See all
UMDAgents can reuse lessons from completed tasks to adjust their tools and instructions mid-run, without carrying task-specific changes into future runs.
Dense, time-aligned descriptions and explicit reference bindings help video generators follow complex scenes and preserve the right identities across shots.
ColumbiaA robot simulator that predicts geometry, images, and touch in a shared 3D scene can generate training data from camera views never recorded in reality.
CMULLM proposals often identify the problems researchers pursue but miss their concrete approaches, and expert reviewers judge unmatched proposals less likely to succeed.
FDUCoworkOS lets agents and people edit shared documents and designs while preserving work across interruptions, supporting complex projects that span multiple sessions.
IntologyLocus autonomously adapts molecular models and designs antibody workflows that perform better on held-out targets, often using fewer candidate molecules.
A low-cost wheeled robot can learn coordinated household tasks from human demonstrations, offering researchers an integrated platform for studying mobile manipulation.
As pretraining data grows, AdamW-trained language models can surpass Muon-trained models after fine-tuning, despite Muon achieving lower pretraining loss.
Reconstructing workspace geometry from human demonstrations lets robots learn whole-body motions that avoid obstacles while carrying out manipulation tasks.
TsinghuaA shared robot-control harness lets open-weight models manipulate objects in simulation and on physical robots without privileged simulator data.
