GLM-5 is a new open-weight AI model that moves from 'vibe coding' (prompting the model to write code) to 'agentic engineering' (letting the model plan, build, test, and fix software on its own).
This paper teaches image models to copy a change shown in one image pair and apply it to a new image, like saying 'hat added here, add a similar hat there.'
DreamZero is a robot brain that learns actions by predicting short videos of the future and the matching moves at the same time.
The paper teaches small AI models to make high‑quality text embeddings by first copying a big expert model (distillation) and then practicing four jobs with special mini‑modules (LoRA adapters): retrieval, similarity, clustering, and classification.
TAROT teaches code-writing AI the way good teachers teach kids: start at the right level and raise the bar at the right time.
The paper finds a simple trick—randomly skipping some parameter updates—can train large language models better than fancy optimizers.
COMPOT is a training-free way to shrink Transformer models while keeping their smarts.
Panini is a way for AI to keep learning new facts without changing its brain by storing them as tiny linked Q&A facts in an external memory.
ResearchGym is a new "gym" where AI agents are tested on real research projects end to end, not just on toy problems.
The paper builds StarWM, a ‘world model’ that lets a StarCraft II agent imagine what will happen a few seconds after it takes an action.
This paper introduces Nexus Adapters, tiny helper networks that let a diffusion model follow both a text prompt and a structure map (like edges or depth) at the same time.
This paper builds a medical image segmentation system that uses both pictures (like X-rays) and words (short clinical text) at the same time.