HyperThink: Text-to-Parameter Hypernetworks for Efficient Reasoning
Long-form thinking traces can substantially improve the multi-step reasoning performance of large language models (LLMs), but they introduce high inference-time overhead, with late...