Block Drafting Model Gaps Analysis
Beyond Parallel Blindness: Information Floors and Model Gaps in Block Drafting
The findings help developers understand why block drafters underperform and guide improvements to proposal quality in text generation systems.
The paper introduces an information floor concept to separate two losses in block drafters: missing within-block path information and imperfect observable information modeling. On Qwen3-4B, the all-parallel floor reaches 0.286 at the final slot, limiting best proposal acceptance to 71% per slot. One realized token removes 86-100% of this floor, as confirmed by mutual-information analysis. Current drafters show significant model gaps accounting for 43-64% of DFlash rejection and 85-92% of DSpark's oracle-conditioned rejection.
Beyond Parallel Blindness: Information Floors and Model Gaps in Block Drafting
Block drafters propose several tokens in one forward pass, before earlier target tokens are realised. Their rejection mixes two losses: missing within-block path information and imperfect modelling of observable information. Accepted length cannot distinguish them. We separate the two with an information floor, the minimum expected rejection at a specified conditioning order; rejection above this floor is the model gap. Estimating both from target rollouts across four domains, four open-weight targets, and a frontier API target yields three findings. First, the all-parallel floor reaches $0.286$ at the final slot on Qwen3-4B, limiting even the best proposal to $71\%$ per-slot acceptance. Second, one realised token removes $86$--$100\%$ of this floor, a locality also recovered by an independent mutual-information analysis. Third, current drafters remain far above their floors: the final-slot model gap accounts for $43$--$64\%$ of DFlash rejection and $85$--$92\%$ of DSpark's oracle-conditioned rejection. These findings separate the value of short-range conditioning from proposal quality.
