OUTLETS Predicts LLM Output Length via Speculative Decoding

OUTLETS repurposes the draft decoder in speculative decoding frameworks like EAGLE-3 to predict generation length. By using existing latent representations from the speculative backbone, it eliminates the latency of external proxy models for more efficient cluster scheduling.