X Debate: Is Chain-of-Thought Prompting a Form of Parameter Reuse?

aryaman2020 · x · 2026-09-03

aryaman2020 and former Llama researcher Suchin Zeng clashed over what counts as parameter reuse: he argues that running the same params over more KV cache (as in layer looping) is a form of reuse, likening it to CoT, which she rejects. He later adds that under the 'KV as fast weights' view the label is shaky — activations fed into each loop step can also be seen as fast weights, motivating test-time training for more expressivity.

Original post →

More from Research

Research channel →