Latent

    models · language · multimodal

    A small object on a rotating turntable in a studio, surrounded by cameras at different angles
    Latent · Video generation

    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
    A person speaking in a busy café while blurred figures talk in the background
    Latent · Speech separation

    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
    Abstract illustration of a photograph dissolving into a grid of tokens beside lines of text on a screen
    Latent · Multimodal efficiency

    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
    A doctor reviewing patient test results on a computer screen in a hospital ward
    Latent · AI diagnosis

    AI modelled on a panel of doctors beats generalist physicians in a diagnosis benchmark

    In a benchmark where diagnosticians must decide what information to gather, a language-model system far outscored generalist doctors, under controlled conditions.

    Science · 22 October 2026
    A video frame of an indoor room overlaid with arrows and distance lines marking the positions of objects
    Latent · Spatial vision models

    AI that judges left from right may not learn how far away things are

    A preprint names a training pitfall called interface lock-in and reports that replaying mixed spatial data in time order improves video models on three benchmarks at once.

    arXiv preprint · 25 October 2026
    Pages of documents, an audio waveform and strips of video frames flowing into a single glowing processor
    Latent · Multimodal parsing

    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
    Rows of server racks in a dimly lit data centre with glowing status lights
    Latent · Agent serving

    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
    Abstract illustration of a computer system sorting stacks of practice questions made of images, charts and text
    Latent · Model self-improvement

    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
    Two abstract grids of points, one formed from words and one from image tiles, rotating into alignment
    Latent · Representation learning

    Good training pairs, not scale alone, bring AI models to a shared representation

    A preprint proves that one family of AI models learns the same representation across data types when trained on the right pairs, and calls next-token prediction an inefficient route.

    arXiv preprint · 29 October 2026
    A laptop screen showing a search bar and scrolling lines of text in a dimly lit room
    Latent · AI search agents

    Rewarding how AI searches, not just what it answers, yields sturdier search agents

    Grading the intermediate steps of an AI search agent, rather than only its final answer, produced agents that coped better with unfamiliar tasks, a study reports.

    Nature Communications · 28 October 2026
    An open book whose pages dissolve into lines of computer text on a desk
    Latent · Computational narratology

    AI story generation trails story understanding in its use of narrative theory

    A survey of language-model research that draws on narratology finds work on reading stories ahead of work on writing them, and argues against a single test of ‘narrative quality’.

    Artificial Intelligence Review · 30 October 2026

    Signals

    Latent