Meet EAGLE 3.1: The Speculative Decoding Algorithm That Fixes Attention Drift in LLM Inference

EAGLE team released EAGLE 3.1, an update to speculative decoding focused on reducing attention drift.

It adds FC normalization and post-norm hidden-state feedback to improve robustness across prompts and long contexts.

The upgrade is backward-compatible with EAGLE 3 checkpoints and integrates with TorchSpec and vLLM.

Overall, EAGLE 3.1 aims to make LLM serving and inference more stable and reliable in real-world workloads.