HyperThink Amortizes Reasoning Latency via Text-to-Parameter Hypernetworks

HyperThink uses a lightweight hypernetwork to predict parameter updates for a base LLM, converting long-form thinking traces into a single query-conditioned update. This approach eliminates sequential decoding of intermediate reasoning steps at inference time while maintaining reasoning performance.