Pulkit Gopalani

Excited to announce our recent work on understanding training-time emergence in Transformers! Thread🧵(1/11)

What drives in-context learning in LLMs? New paper: Provable Low-Frequency Bias of In-Context Learning of Representations. We show LLMs have a low-frequency bias when learning representations in context, offering a theoretical answer to several previously open questions. 🧵👇

@scychan_brains We studied training-time emergence for algorithmic tasks in shallow Transformers: x.com/GopalaniPulkit…

Emergence in transformers is a real phenomenon! Behaviors and capabilities can appear in models in sudden ways. Emergence is not always just a "mirage". Compiling some examples here (please share any I missed): 🧵

Check out our paper for details and results on other algorithmic tasks / model configurations: arxiv.org/abs/2506.13688 GitHub repository: github.com/pulkitgopalani… (11/11)

Repetitive sequences are easy for Transformers: we show that training on sequences like ‘x_1, x_2, …, x_n, [sep] x_1, x_1, …, x_1’ (or other similar sequences) does not involve loss plateaus like other algorithmic tasks, and the loss converges in a few training steps. (10/11)

Excited to announce our recent work on understanding training-time emergence in Transformers! Thread🧵(1/11)