A new preprint on arXiv, Transformers as Adaptive Kernel Learners across Tasks for In-Context Learning, argues that transformers can be understood as adaptive kernel learners. The opening of the abstract notes that in-context learning (ICL) lets models adapt to new tasks by conditioning on prompts, with no parameter updates, and that the paper aims to understand this behavior.
The source, however, is incomplete: the abstract ends mid-sentence at "To understand the…", so the proposed mechanism, experiments, and conclusions are not visible in the available text. Because there is only this single truncated source, no concrete claims about the paper's findings can be reported here.
Readers should treat the headline as the paper's stated framing rather than a verified result. The full abstract or manuscript would be needed to assess how the authors formalize adaptive kernel learning and what evidence they provide.