FlashMap: Flow Map Speculative Decoding

Jeff Shen; Zhejian Zhou; Ryan Lee; Jonathan May; Shirley Ho

cs.CLOct 17, 2026

FlashMap: Flow Map Speculative Decoding

Authors: Jeff Shen, Zhejian Zhou, Ryan Lee, Jonathan May, Shirley Ho

Organizations: Polymathic AI · Princeton University · USC Information Sciences Institute · New York University ·

Abstract

Speculative decoding accelerates large language model (LLM) inference by using a lightweight drafter to propose a block of tokens that the target model verifies in one parallel pass. Autoregressive drafters generate this block one token at a time, whereas block diffusion drafters such as DFlash generate it in one pass, making drafting faster. However, these diffusion drafters predict each token independently, ignoring dependencies between tokens in the block. We introduce FlashMap, a continuous flow drafter that conditions a flow map language model on the target model's hidden states and generates the whole block in one forward pass while capturing inter-token dependencies. On Qwen3-8B, FlashMap is faster than a matched DFlash baseline on every benchmark, with 7% higher average speedup and up to 6.1x lossless acceleration.

Figures & tables

Panel of Figure 1 — Figure 1: (a) One round of speculative decoding: the drafter reads hidden features from the target and proposes the next $K$ tokens in parallel with a flow map, and the target verifies them in one pass. (b) Sampling two correlated slots in the few-step regime (adapted from Figure 3 of FMLM): DFlash predicts each slot independently, so half of its drafts pair tokens from different modes, while FlashMap can represent a joint over all slots. (c) Speedup ( $\times$ ) on Qwen3-8B, with FlashMap drafting in 1 or 2 flow-map steps. Bold marks the best speedup.

Temp.	Method	Math			Code			Chat
Temp.	Method	GSM8K	MATH	AIME25	MBPP	HumanEval	LCB	MT-Bench	Alpaca	Arena-Hard
$T=0$	EAGLE-3 (16)	6.75	6.43	5.96	5.38	6.04	6.16	3.75	3.71	3.97
	DFlash	7.43	7.09	5.92	4.77	5.74	6.35	3.14	2.96	3.38
	FlashMap (1 step)	7.69	7.24	6.00	4.95	5.82	6.65	3.34	3.14	3.71
	FlashMap (2 steps)	8.31	7.69	6.36	5.24	6.26	7.19	3.40	3.20	3.80
$T=1$	EAGLE-3 (16)	6.15	5.55	4.61	4.94	5.29	5.37	3.43	3.25	3.20

Table 1: Acceptance length

\gamma

on Qwen3-8B. For each temperature and benchmark, bold marks the best result. Speedups are in Figure 1 (c).

Figure 2: Acceptance length

\gamma

against the proposal length

K

used at inference on Qwen3-8B at

T=0

, for FlashMap with one and two steps and for DFlash, retrained with the same data and settings, on GSM8K, HumanEval, and MT-Bench (80 prompts each). All drafters were trained at

K=16

and run at smaller

K

without retraining. Each point is a single run, so no uncertainty is shown.

Figure 3: Conditional acceptance rate on Qwen3-8B at

T=0

for the drafters of Figure 2 : the probability that the draft token at position

k

of a

K=16

block is accepted given that all earlier tokens in the block were accepted, pooled over the math, code, and chat tasks. Positions are 0-indexed, so position

k

is slot

k+1

; shaded bands are 95% Wilson intervals.

	Math			Code			Chat
Drafter	Shared	Shuffled	$\Delta$	Shared	Shuffled	$\Delta$	Shared	Shuffled	$\Delta$
DFlash	7.29	7.29	0.00	6.07	6.07	0.00	2.92	2.92	0.00
Flow (1 step)	6.56	6.56	0.00	5.46	5.46	0.00	2.88	2.88	0.00
Flow (16 steps)	7.85	7.03	0.82	6.70	6.11	0.59	3.32	3.03	0.29
FlashMap (1 step)	7.05	6.73	0.32	5.92	5.78	0.13	3.04	2.96	0.08
FlashMap (2 steps)	7.70	6.92	0.78	6.57	6.11	0.47	3.27	3.03	0.25

Table 2: Shuffle test on Qwen3-8B at

T=0

: accepted draft tokens per block, for a draft with shared noise across slots (shared) and for a draft with independent per-slot noise (shuffled).

\Delta

is the difference between the two, with a larger value indicating a greater acceptance due to joint structure. Results for Qwen3-4B are in Table 5 .

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Temp.	Method	Math			Code			Chat
Temp.	Method	GSM8K	MATH	AIME25	MBPP	HumanEval	LCB	MT-Bench	Alpaca	Arena-Hard
Speedup ( $\times$ )
$T=0$	EAGLE-3 (16)	2.66	2.64	2.47	2.20	2.33	2.44	1.49	1.37	1.64
	DFlash	5.18	5.23	4.56	3.72	4.00	4.27	2.43	2.28	2.58
	FlashMap (1 step)	5.14	5.02	4.47	3.71	4.04	4.28	2.46	2.28	2.66
	FlashMap (2 steps)	4.75	4.57	4.00	3.42	3.76	3.87	2.16	2.00	2.34

Appendix

Table 3: Speedup (

\times

) over autoregressive decoding with the target model (top) and acceptance length

\gamma

(bottom) on Qwen3-4B. For each metric, temperature and benchmark, bold marks the best result and underline the second best.

Steps	Samples	Tree size	Math			Code			Chat
Steps	Samples	Tree size	GSM8K	MATH	AIME25	MBPP	HumanEval	LCB	MT-Bench	Alpaca	Arena-Hard
Speedup ( $\times$ )
1	1	16	6.34	5.57	4.90	3.98	4.47	4.71	2.62	2.41	2.48
	2	16	6.62	6.00	5.28	4.16	4.73	4.92	2.77	2.86	2.49
	4	16	6.84	6.07	5.29	4.25	4.78	5.73	2.85	2.98	2.58
	4	64	6.49	5.62	4.85	4.29	4.46	4.87	2.59	2.75	2.21

Appendix

Table 4: Multi-sample drafting with FlashMap on Qwen3-8B at

T=0

: speedup (

\times

) over the target and acceptance length

\gamma

for various numbers of steps, sample counts, and tree verification sizes. For each metric and benchmark, bold marks the best result. Calculated on one H100 SXM with 32 prompts per benchmark (30 for AIME25).

Figure 4: Same as Figure 3 but at $T=1$ .

Panel of Figure 5 — Figure 5: Same as Figure 3 , but for Qwen3-4B at $T=0$ (top row) and $T=1$ (bottom row).

	Math			Code			Chat
Drafter	Shared	Shuffled	$\Delta$	Shared	Shuffled	$\Delta$	Shared	Shuffled	$\Delta$
DFlash	7.44	7.44	0.00	5.74	5.74	0.00	2.88	2.88	0.00
Flow (1 step)	6.74	6.74	0.00	5.20	5.20	0.00	2.78	2.78	0.00
Flow (2 steps)	7.70	7.25	0.46	6.12	5.79	0.34	3.04	2.88	0.16
Flow (16 steps)	8.05	7.22	0.83	6.47	5.84	0.62	3.17	2.88	0.29
FlashMap (1 step)	7.07	6.71	0.36	5.60	5.45	0.15	2.86	2.77	0.08

Appendix

Table 5: Same as Table 2 but for Qwen3-4B.

Speculative Decoding

Speculative decoding accelerates generation by drafting multiple tokens and verifying them in parallel with a larger target model.

See all 308 papers

DFlare: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding
Jun 1, 2026Jiebin Zhang, Zhenghan Yu, Song Liu +9Speculative Decoding LLM Inference Acceleration
School of Computer Science, Peking University · Tencent
Accelerating Speculative Decoding with Block Diffusion Draft Trees
Apr 14, 2026Liran Ringel, Yaniv RomanoSpeculative Decoding Parallel Speculative Decoding
Department of Computer Science, Technion – Israel Institute of Technology. · Department of Electrical and Computer Engineering, Technion – Israel Institute of Technology.
DStill: Trajectory Consistency Distillation for Accelerated One-Step Drafting in Block Diffusion Speculative Decoding
Oct 19, 2026Shreyas Kulkarni, Uttam Gurram, Madhav Kanda +5Speculative Decoding for Diffusion Language Models LLM Inference Acceleration
AI Foundations, Capital One

Jun 1, 2026J/K move · Enter open · S save

DFlare: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding

Jiebin Zhang, Zhenghan Yu, Song Liu, Eugene J. Yu+8

School of Computer Science, Peking University · Tencent

Speculative Decoding LLM Inference Acceleration Diffusion Language Model Decoding

Open paper AskPDF arXiv

Block diffusion speculative decoding accelerates LLM inference by predicting all tokens within a block simultaneously for the target model to verify in parallel. Predicting an entire block at once requires a sufficiently capable draft model and effective utilization of the target model's internal knowledge. However, the state-of-the-art method DFlash constrains all draft layers to share a single fused representation derived from only a few target layers, limiting per-layer expressiveness and hindering further scaling of draft capacity. In this paper, we present \modelname, which flares out the narrow conditioning bottleneck of DFlash through a lightweight layer-wise fusion mechanism: each draft layer attends to its own learnable combination of a broad set of target layers at negligible overhead, simultaneously injecting richer target knowledge and providing every draft layer with a distinct input. This enhanced per-layer expressiveness enables scaling the draft model to deeper architectures with consistent gains. We further scale training data from 800K to 2.4M samples to fully exploit the enlarged capacity. On six benchmarks spanning mathematical reasoning, code generation, and conversation, \modelname attains average wall-clock speedups of 5.52x on Qwen3-4B, 5.46x on Qwen3-8B, and 3.91x on GPT-OSS-20B, improving over DFlash by roughly 11%, 8%, and 5% respectively. Our code is available at https://github.com/Tencent/AngelSlim.

Abstract

Speculative Decoding

Explore similar work

DFlare: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding

Accelerating Speculative Decoding with Block Diffusion Draft Trees

DStill: Trajectory Consistency Distillation for Accelerated One-Step Drafting in Block Diffusion Speculative Decoding