A recall head lets a model decide when a long-context read is worth paying for.

arXiv 2609.20734 shows a pretrained model's decoding states already carry information predictive of whether a global read will help, before the read happens. Training only a small recall head to invoke global attention selectively, with pretrained weights untouched and the full historical KV cache available for later recall, plus GPU-side conditional execution implemented in vLLM, turns reduced global reads into real decoding speedups at long context. Across Qwen and Gemma including hybrid-attention backbones, selective recall recovered most of the accuracy lost under pure local attention. arXiv