LEAP

LEAP achieved substantial early-exit efficiency at the recommended threshold while maintaining STS-B quality near the baseline. LEAP trains intermediate representations toward both the teacher final embedding and the student final embedding. LEAP modifies training by encouraging intermediate layers to approximate final-layer representations while preserving embedding quality. LEAP requires retraining and cannot be added post hoc to an existing distilled checkpoint. LEAP is introduced to make distilled embedding transformers compatible with convergence-based early exit without adding inference-time parameters.