What if LLMs could think in weights, not tokens?

Meet HyperThink: a hypernetwork writes a per-question weight update to the LLM for it to skip the long thinking trace, a step toward accurate System 1 thinking.

How it works: from the question trained hypernetwork predicts a tiny set of bias offsets (under 0.02% of the weights) for the frozen LLM's later layers through a vector-quantized codebook, and the adapted LLM writes a short step-by-step answer with no thinking trace.

On math, HyperThink beats every baseline at comparably low latency, even the model's own extensively post-trained native non-thinking mode. It also approaches full thinking mode with up to 3.9× fewer FLOPs and generalizes to benchmarks held out from training.