HyperThink aims to let LLMs skip long thinking traces with per-question weight updates

A HyperThink co-lead introduced the project as a way for LLMs to “think in weights, not tokens.” They say a hypernetwork writes a per-question weight update intended to let a model skip a long thinking trace. The post described this as a step toward accurate “System 1” thinking and listed a COLM 2026 presentation for October 6.