Han-Yang Peng

These findings support long-context inference in which pretrained models guide their own access to the information they retain, by introducing On-Demand Attention (ODA), a local-first decoding method that uses a lightweight recall head to selectively invoke global attention as its predicted benefit changes during generation.