Tree‑based speculative decoding verifies many draft continuations in a single target‑model pass, yet finite trees built from draft scores suffer a fundamental draft‑target mismatch. By adopting a target‑flow perspective we identify a canonical exit law and prove that $1+\text{target coverage}$ sharply bounds the expected output‑block length (including the bonus token) of any exact path verifier.
All optimal verifiers share the same exit‑node and bonus‑token law, a property already realized by representative predraw‑and‑follow and sequential residual verifiers.
From this insight we introduce Tree Exit Verification (TEV), an exact, level‑parallel verification procedure that requires only one exit‑node decision and one bonus‑token decision.
The exit law also reveals the missing target probability, providing node‑level feedback for Exit‑Guided Draft‑Tree Training (ExitTrain), which adapts inference‑time draft trees to improve acceptance.
Experiments on dialogue, code generation and mathematical reasoning tasks show that, under a fixed‑tree setting, ExitTrain increases the average output‑block length by roughly 13%, while TEV reduces verifier‑stage latency by about 15%, yielding an overall end‑to‑end speedup of ~14% over DDTree.
The implementation is available at https://github.com/hsj576/TEV.
Review