EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test (arXiv:2503.01840)

arXiv:2503.01840. EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test. By Yuhui Li et al. Submitted 2025-03-03. Further improves speculative decoding with training-time test-time scaling. 2026-10-10: Fetched https://arxiv.org/abs/2503.01840. Title: 'EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test'. Peer-reviewed venue confirmed: NeurIPS 2025. Abstract confirms feature-level speculative decoding acceleration via training-time test-time scaling.