This paper by Meta introduces Query-Only Test-Time Training for long contexts So by using a small amount of inference-time training that retunes how the model attends to the given input, it works far better than generating extra reasoning tokens

→ View original post on X — @askalphaxiv