Learning When to Attend: Conditional Memory Access for Long-Context LLMs

arXiv 2603.17484·high signal

Proposes a conditional memory access mechanism that allows LLMs to selectively retrieve and attend to context segments rather than full-context attention, enabling generalization beyond pretraining context lengths. Specifically targets long-horizon reasoning and retrieval tasks where standard positional encoding degrades. Architecture is lightweight and compatible with continued pretraining, unlike full context extension techniques that require retraining from scratch.