News

Meet ReGE, which efficiently consolidates an agent’s past attempts at a task into guidance that helps the same model do better on later attempts, letting GPT-5.6-Luna match GPT-5.6-Sol at only 28% of Sol’s cost. Read the paper or visit the project page.

Released StopLoss, a small monitor that learns from finished runs to predict when a coding agent is about to fail, so doomed runs can be stopped early. Read the paper, browse the project page, or grab the models and training data.

New preprint: FIM-Midtraining mid-trains code LLMs to fill in masked functions, a task that mirrors how an agent calls a tool and uses the result, and lifts SWE-Bench-Verified by about 3 points. The paper is on arXiv, and the checkpoints and 2.6B-token corpus are on Hugging Face.

SWE-Next is out! It turns real merged GitHub pull requests into 2,308 self-verifying tasks for training software-engineering agents. See the paper, the project page, and the dataset.

Introducing VisPhyWorld, which tests whether multimodal LLMs really understand physics by asking them to rebuild a video as runnable simulator code. The paper and benchmark data are available now.

Our VisCoder2 models (3B to 32B) write, run, and self-debug visualization code in 12 programming languages. Details are in the paper; the models, the 679K-sample training set, and the VisPlotBench benchmark live in our Hugging Face collection.