Definition
Plain language
Adjusting a model's weights on the fly while it works on the current task.
As stated in the literature
A family of approaches (e.g., TTRL, Akyürek et al.) that update model parameters at inference time using task-specific signals; one of two communities SIA proposes to unify with scaffold optimization.
Also called: TTT
Why it matters: It lets models adapt to genuinely novel tasks at inference, sidestepping the limits of what their frozen weights already encode.
For example, when given a new visual puzzle, a model might briefly fine-tune itself on a few transformed copies of that puzzle before answering.
Heard on the show
“… the concept pages that link this over to the other episodes we've done on agent memory and test-time training — that all lives on paperdive dot AI. …”Episode 114 — Agents That Rewrite Their Own Weights Instead of Just Taking Notes