Scholar Feed

Speculative decoding

DFlash

DFlash: Block Diffusion for Flash Speculative Decoding

Superseded baseline#10 of 151 most-superseded · first seen Feb 5, 2026

Superseded — cited as a baseline and beaten by newer methods

4 papers critique it · 6 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites DFlash as a baseline.

Limitation. DFlash acceptance length is fundamentally bounded by the longest correct prefix: once a mismatch occurs at some position, all subsequent draft tokens are discarded regardless of their quality.
— D^2SD: Accelerating Speculative Decoding with Dual Diffusion Draft Models
“However, these one-shot drafters have a fundamental limitation: the predicted distribution for draft token $x_{t+i}$ is conditioned only on the prefix context $x_{ t}$, with no dependence on preceding drafted tokens. This non-autoregressive conditioning causes the drafter's distribution to increasingly diverge from the verifier's true autoregressive distribution as draft depth grows.”
— TreeFlash: Parallel AR-Approximation for Faster Speculative Decoding
“Vanilla DFlash, however, explores only one continuation per round.”
— Accelerating Speculative Decoding with Block Diffusion Draft Trees
“While recent approaches such as DFlash and DART mitigate this issue with position-aware decaying weights, their weights are fixed and primarily position-dependent.”
— PARD-2: Target-Aligned Parallel Draft Model for Dual-Mode Speculative Decoding

Beaten on benchmarks

Head-to-head results where a newer method reports beating DFlash. Values are copied from the source paper's tables — verify against the cited paper.

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.