Robot control model anticipates motion without generating future video
A preprint describes a vision-language-action model that learns compact physical tokens about how scenes will change, sidestepping the delay of generating video.
arXiv preprint · 27 October 2026
Tencent's Hy-Nomas generates video worlds that respond to text and camera commands
A video world model from Tencent Hunyuan follows successive typed instructions and camera moves while keeping track of what has already happened on screen.
arXiv preprint · 27 October 2026
Surface-shape maps help an AI video model keep a subject consistent from every side
A KAIST AI preprint steers video generation with geometry maps rebuilt from reference photos, so a subject stays the same as the camera circles it.
arXiv preprint · 26 October 2026
Leaner streaming model picks one voice out of crowded, noisy scenes
A preprint offers a more realistic test for audio-visual voice extraction, plus a streaming model the authors say is far smaller and faster than an existing baseline.
arXiv preprint · 23 October 2026
Teaching diffusion vision-language models to look at the image only when needed
A decoding method from KAIST AI checks whether the picture would change the answer before processing it, cutting reported decoding computation to about a third.
arXiv preprint · 27 October 2026
Benchmark finds AI agents remember conversations but struggle to infer user habits
A test built on hundreds of synthetic sessions finds the best agent memory method answering only about a third of habit questions correctly.
arXiv preprint · 20 October 2026
Robots learn faster by rewinding imagined futures inside a world model
A preprint describes a training method that replays a robot's decisions in a learned simulator, tries alternatives, and credits the choices that made the difference.
arXiv preprint · 26 October 2026
Tencent's single model parses documents, images, audio and video into one format
A preprint describes one shared encoder and one output schema for ten parsing tasks, with the authors reporting leading scores on document and multimodal benchmarks.
arXiv preprint · 29 October 2026
A scheduler that plans around AI agents' tool calls cuts serving delays
TAO, described in a preprint, picks which agent sessions get GPU time by weighing when their tools will return, what is cached where and how many tokens they carry.
arXiv preprint · 27 October 2026
Self-training AI decides for itself when its practice tasks have gone stale
A preprint from Xiaomi Research describes a training loop that watches its own learning signal and swaps in fresh, checkable questions, with no human-labelled data.
arXiv preprint · 28 October 2026
Hybrid model predicts car-following on snowy Norwegian highways more accurately
Researchers combine two physics-based driver models with a Transformer-LSTM network and report lower acceleration errors on Norwegian winter driving data.
PLoS ONE · 30 October 2026