Beyond Parallel Blindness: Information Floors and Model Gaps in Block Drafting
The paper says block-drafting rejection can be split into an unavoidable information floor and a separate model-quality gap.
The authors define the floor as the minimum expected rejection at a given conditioning order, with any extra rejection counted as the model gap. Across target rollouts, the all-parallel floor reaches 0.286 in Qwen3-4B’s final slot, capping even an ideal proposal at 71% per-slot acceptance. Realising one token removes 86% to 100% of that floor. Current drafters still sit well above their floors, with final-slot model gaps accounting for large shares of DFlash and DSpark rejection. ArXiv · AI/CL/LG's note
The authors define the floor as the minimum expected rejection at a given conditioning order, with any extra rejection counted as the model gap. Across target rollouts, the all-parallel floor reaches 0.286 in Qwen3-4B’s final slot, capping even an ideal proposal at 71% per-slot acceptance. Realising one token removes 86% to 100% of that floor. Current drafters still sit well above their floors, with final-slot model gaps accounting for large shares of DFlash and DSpark rejection. ArXiv · AI/CL/LG's note
score 5