1 articles with this tag
Transformers with Temporal Middle-Layer Recurrence (T2MLR) enables persistent intermediate reasoning by fusing cached middle-layer states, outperforming baselines and allowing efficient model retrofitting.
