From Chains to Trees: Parent-Conditioned Drafting for Semi-Autoregressive Speculative Decoding

TL;DR - PCTree turns a semi-autoregressive speculative decoder's linear draft chain into a verification tree by reusing the drafter's already-pretrained Markov head, delivering faster LLM inference with no retraining and no extra backbone passes.

• Diagnoses DSpark's core limitation: a token block decoded as a single chain means one early mismatch invalidates the whole remaining suffix, capping gains from large draft blocks.
• PCTree scores alternative children separately per concrete parent using the existing Markov head, then allocates a fixed verification budget to the most probable paths — an inference-only change preserving one-pass parallel backbone drafting.
• Across Qwen3-4B/8B/14B and nine benchmarks at block size B=7, speedup gains over autoregressive decoding improve 3.1%–29.5% relative to matched DSpark.
• On Qwen3-4B GSM8K at B=16, mean acceptance length rises from 9.41 to 11.16 and three-run mean AR speedup from 6.14× to 6.60×.