HyperThink uses a hypernetwork to replace long chain-of-thought with one parameter update

Donggyun Kim · hf · 2026-10-14

HyperThink proposes a text-to-parameter approach that amortizes LLM reasoning computation.

Method: A lightweight hypernetwork reads the question and predicts updates to a small subset of the base LLM's parameters in a single forward pass; a vector-quantized decoder constrains updates to a finite set of reusable patterns for robustness and transfer.
Training: End-to-end on the base model's own thinking traces.
Results: At test time no long thinking trace is needed—the adapted model directly emits a concise step-by-step solution with far fewer tokens, improving the low-latency region of the accuracy-latency trade-off on math and general reasoning, with the strongest gains in the near-non-thinking regime.

More from Models