Researchers have proposed MentorPulse, a method for refreshing cross-model latent guidance in long-form generation tasks. This approach addresses the limitation of existing methods that assume the guidance signal remains useful as the output grows. MentorPulse compresses mentor states into a capped slot memory and incrementally processes newly generated tokens to update the guidance signal. The authors claim this method outperforms other approaches, such as C2C, T2T, and LoRA, on thirteen datasets with an average improvement of 52.2% in closing the mentor-student gap. --- Why it matters: This matters because long-form generation tasks often rely on frozen large models that provide guidance to smaller student models. However, existing methods fail to adapt this guidance as the output grows, leading to suboptimal performance. MentorPulse addresses this limitation by providing a practical and cost-effective way to refresh the guidance signal. Source: https://arxiv.org/abs/2608.20927