Timer-S1 is a huge time-series model (8.3B parameters, only 0.75B used per step) that predicts the future by thinking step-by-step inside one forward pass.
The paper trains one model from scratch to both read text and see images/videos, instead of starting from a language-only model.
This paper builds a gigantic library of video puzzles (VBVR) so AI can practice not just making pretty videos, but actually thinking through what happens over time.
This paper shows, step by step, how to train a 1.36-billion-parameter science-focused language model directly from raw arXiv LaTeX files using only 2 A100 GPUs.
This paper builds a new audio tokenizer, called MOSS-Audio-Tokenizer, that turns sound into tiny tokens the way text tokenizers turn sentences into words.
Long tasks trip up most AIs because they lose track of goals and make small mistakes that snowball over many steps.
Metric Anything is a new way to teach AI real, ruler-like distances (metric depth) from very mixed and noisy 3D data.
The paper shows a simple way to teach AI models what not to learn by removing only the exact words (tokens) related to unwanted topics during pretraining.
SERA is a new, low-cost way to train coding helpers (agents) that learn the style and secrets of your own codebase.
TSRBench is a giant test that checks if AI models can understand and reason about data that changes over time, like heartbeats, stock prices, and weather.
This paper builds a fair, big playground (a benchmark) to test many EEG foundation models side-by-side on the same rules.
The paper shows that big language models often get stuck with weight sizes set by training hyperparameters instead of by the data, which quietly hurts performance.