Eagle 3.1: Collaboration Between the EAGLE Team, vLLM Team, and TorchSpec Team

EAGLE 3.1, a collaboration between the EAGLE, vLLM, and TorchSpec teams, addresses robustness issues in speculative decoding by introducing FC normalization and post-norm hidden-state feedback to combat "attention drift"—a phenomenon where the drafter becomes unstable at deeper speculation depths. These architectural improvements result in significantly better long-context robustness, with up to 2× longer acceptance lengths compared to EAGLE 3, while maintaining backward compatibility with existing EAGLE 3 checkpoints. The integration is now available in vLLM and supported by TorchSpec's efficient training framework.