Papers1055

Autoregressive (AR) models write one word at a time, which is accurate but slow, especially when your computer or GPU can’t keep many tasks in memory at once.

HyperVL is a small but smart model that understands images and text, designed to run fast on phones and tablets.

OpenDataArena (ODA) is a fair, open platform that measures how valuable different post‑training datasets are for large language models by holding everything else constant.

FINERWEB is a new, carefully built dataset pipeline that teaches computers to spot names of people, places, and more across 91 languages and 25 writing systems.

SAGE is a smart video-watching agent that decides when to answer quickly and when to take multiple steps, just like how people skim or rewind long videos.

LitePT is a new AI backbone for 3D point clouds that uses convolutions in early layers and attention in later layers to be both fast and accurate.

The paper tackles a paradox: visual tokenizers that get great pixel reconstructions often make worse images when used for generation.

Steer3D lets you change a 3D object just by typing what you want, like “add a roof rack,” and it does it in one quick pass.

Digital humans used to just copy motions; this paper makes them think, speak, and move in sync like real people.

RoboTracer is a vision-language model that turns tricky, word-only instructions into safe, step-by-step 3D paths (spatial traces) robots can follow.

The paper introduces Nemotron-Cascade, a step-by-step (cascaded) reinforcement learning recipe that trains an AI across domains like alignment, instructions, math, coding, and software engineering—one at a time.

LongVie 2 is a video world model that can generate controllable videos for 3–5 minutes while keeping the look and motion steady over time.