October 2026 Summaries

EAGLE-4 is a speculative decoding system designed to accelerate large language model responses, particularly prompt enhancement for increasingly fast image and video generation workflows. It uses a lightweight drafter that predicts token continuations from target-model hidden states and embeddings, adding sliding-window attention and a low-rank candidate head to reduce drafting costs. Rather than relying on a fixed token chain, its hardware-calibrated harness dynamically constructs branching verification trees, allowing the target model to check multiple plausible continuations in one pass while preserving its output distribution through rejection sampling. The accompanying Prompt Enhancement SpecBench dataset contains 400 creative examples across text-to-video, image-to-video, text-to-image, and image-to-image tasks, addressing a gap in speculative-decoding evaluation. Benchmarks on NVIDIA B300 GPUs found that branching improved acceptance length over chains, especially at larger verification budgets and for harder creative prompts, while EAGLE-4 Harness achieved the highest throughput among compared methods, reaching 660.6 tokens per second on Fireworks Bench and 622.9 on SpecBench at small batch sizes.