Floor-Aware Adaptive Block Drafting

Integrate this into speculative decoding immediately after the target model verifies a draft block, using target logits already computed for accepted and rejected positions. For each anchor context, collect B short target rollouts with different sampled within-block paths. For every conditioning order r, form the target distributions p_{Z_r^(b)} at the next draft slot and estimate the shared proposal with q_hat_r(x)=B^{-1} sum_b p_{Z_r^(b)}(x). Compute the estimated floor F_hat_r as the mean total-variation distance to q_hat_r. Evaluate the current drafter distribution q_theta on the same paths and compute R_hat_r=B^{-1} sum_b TV(p_{Z_r^(b)},q_theta), then set G_hat_r=max(0,R_hat_r-F_hat_r). The mathematical quantities are F_r, R_r, and G_r; target rollouts, q_hat_r, and thresholds are empirical estimates. At inference, maintain an offline lookup table of F_hat_r indexed by domain, token position, target entropy, and anchor length, or update an exponential moving average online. Before generating a block, choose the largest parallel span whose predicted floor is below tau_F. If all longer spans exceed the threshold, sample one target token, append it to the prefix, and recompute the next span. First test on a 1B--3B target/drafter pair with standard speculative decoding on WikiText-103 and HumanEval. Compare fixed block lengths against this adaptive policy at equal target forward calls. Success means more accepted tokens per target evaluation with unchanged perplexity; a useful target is 10% fewer target evaluations at equal output quality.