SEED replaces agent distillation's frozen teacher model

SEED replaces the frozen teacher model in agent distillation with one that keeps training alongside the student, per the arXiv paper. That's the gap between a distilled agent good enough for a demo and one that can actually replace the model it's trained from, and it lands first on teams building smaller, cheaper agents for production tool use.

The standard setup freezes the teacher once training starts, so the student chases a fixed target it can approach but never pass. Running training longer doesn't fix that, because the ceiling is set by the checkpoint, not by how much the student is capable of learning. SEED lets the teacher keep improving too, so the target keeps moving and the student's performance ends up bound by its own capacity instead of a number someone froze on day one.

Whether SEED becomes a standard training recipe or a research footnote depends on a number the paper doesn't publish: how much more compute an evolving teacher burns compared to a frozen one. If that cost is small, expect it to show up in agent training pipelines fast. If it's steep, the plateau problem stays a research curiosity for anyone without spare compute to burn.

It's sitting at 104 upvotes on Hugging Face's papers board as of July 19, the only signal so far on how much attention it's getting outside the paper itself.