Diffusion large language models (dLLMs) predict multiple token positions per forward pass, yet block-wise decoders serialize generation and repeatedly recompute stable predictions. We attribute this inefficiency to coupling a prediction's readiness for irreversible commitment with its utility as context. We introduce Tentative Context Pipelining (TeCoP), a training-free decoder that decouples these roles through staggered block-parallel pipelining and draft-and-verify. TeCoP reuses stable but uncommitted predictions as provisional cross-block context, with commitment governed by confidence and stability checks. Across four LLaDA and Dream dLLMs on GSM8K, Minerva MATH, and MBPP, TeCoP reduces number of forward evaluations by up to 87% and achieves up to 7.50-fold throughput without KV caching and 11.77-fold with caching over the uncached original decoder. It outperforms Fast-dLLM and DSB in throughput across all 12 model-benchmark settings, with and without caching, while keeping average cached accuracy within one percentage point of the original decoder.
Figures & tables
Figure 1: Visualization of pending-position confidence during masked diffusion decoding. Left: In the baseline, many positions outside the active block window achieve high confidence but cannot be committed. Right: Our pipeline method progressively expands the active block region, enabling these predictions to be committed earlier and reducing wasted forward passes.
Figure 2: Overview of the decoding efficiency of Tentative Context Pipelining (TeCoP). Left: Decoding trajectories of the original block-wise decoder, Fast-dLLM ( Wu et al., 2025b ) , DSB ( Luo et al., 2026a ) , and TeCoP, colored by the step at which each token is committed. TeCoP activates multiple blocks in parallel and finishes in fewer number of forward evaluations. Right: Throughput–accuracy trade-off on GSM8K with KV caching; speedups are relative to the uncached original decoder.
Figure 3: Illustration of Tentative Context Pipelining (TeCoP) mechanism. The pipelined block schedule keeps multiple blocks simultaneously active and progressively opens new blocks according to the activation threshold α (set to 75% here for illustration). The draft-and-verify procedure recycles uncommitted predictions as tentative drafts, uses them as provisional context in the next round, and commits them through a confidence or stability check.
Figure 4: Draft-and-verify analysis on 100 GSM8K samples. Left: commit lag by token position. Right: token confidence over decoding steps.
Method
GSM8K
Minerva MATH
MBPP
Acc. (%)
NFE ↓
TPS ↑
Spd. ↑
Acc. (%)
NFE ↓
TPS ↑
Spd. ↑
Acc. (%)
NFE ↓
TPS ↑
Spd. ↑
LLaDA-8B-Instruct
Original
76.8
256.0
16.8
1.00 ×
33.2
256.0
24.1
1.00 ×
40.2
256.0
16.2
1.00 ×
Fast-dLLM
77.5
78.2
54.7
3.25 ×
33.0
97.1
63.5
2.63 ×
40.6
70.1
59.1
3.65 ×
DSB
77.7
74.5
57.4
3.41 ×
33.1
93.6
65.6
2.72 ×
41.8
67.2
60.2
3.72 ×
TeCoP
76.7
61.8
71.7
4.26 ×
32.5
77.4
81.8
3.40 ×
41.2
53.1
75.8
4.68 ×
Table 2: Main results on GSM8K, Minerva MATH, and MBPP. Cached variants use DualCache (Fast-dLLM), DSB’s cache (DSB), and prefix KV cache (TeCoP). Spd. denotes TPS speedup over uncached Original. For Acc., bold and underline mark the top two among Original and cached variants (ties included). For NFE, TPS, and Spd., bold and underline mark the best across all methods.
Method
Δ Acc. (pp)
Spd. ↑
Fast-dLLM + cache
− 1.02
4.50 ×
DSB + cache
+0.20
4.72 ×
TeCoP + cache
− 0.98
6.65 ×
Table 3: Average accuracy change and geometric-mean speedup over original across 4 models and 3 benchmarks with KV caching. See Appendix Table 9 for per-benchmark results.
K
Acc. (%)
NFE ↓
TPS ↑
Spd. ↑
1 (no pipelining)
78.2
72.4
61.0
3.63 ×
2
76.7
61.8
71.7
4.26 ×
3
75.6
46.8
95.1
5.66 ×
4
74.3
41.8
105.9
6.30 ×
Table 4: Effect of pipeline depth K (LLaDA-8B-Instruct, GSM8K). Larger K raises throughput at the cost of accuracy. Spd. is the TPS speedup over Original.
τd
Acc. (%)
NFE ↓
TPS ↑
Spd. ↑
0.6
75.2
42.1
104.3
6.20 ×
0.7
77.1
48.7
90.9
5.41 ×
0.8
76.5
57.4
77.2
4.59 ×
0.85
76.7
61.8
71.7
4.26 ×
≥τ (no drafting)
77.4
66.8
66.2
3.94 ×
Table 5: Effect of draft threshold τd (LLaDA-8B-Instruct, GSM8K). Lower τd raises throughput, but overly low values make accuracy unstable.
Slot content
Positions
Admitted
Rejected
Genuine draft
37,091
97.31%
2.69%
Random token
37,091
16.46%
83.54%
Table 6: Verification diagnostic on GSM8K. Random token replacements cause most drafts to be rejected.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Meaning
b
Block size.
K
Max number of concurrent active blocks.
α
Threshold for opening the next block.
τ
Commit threshold for safe-commit.
τd
Threshold for admitting tentative drafts.
τv
Threshold for committing stable tentative drafts.
Appendix
Table 7: Notation of decoding parameters used in TeCoP.
Model
b
K
α
τ
τd
τv
LLaDA-8B-Instruct
32
2
0.80
0.90
0.85
0.80
LLaDA-1.5
32
2
0.80
0.90
0.85
0.80
Dream-v0-Base-7B
16
2
0.85
0.90
0.85
0.70
Dream-v0-Instruct-7B
16
2
0.85
0.90
0.85
0.70
Appendix
Table 8: Model-family-specific decoding configuration for TeCoP.
Method
Δ Acc. (%)
Spd. ↑
GSM8K
Fast-dLLM + cache
− 0.35
4.80 ×
DSB + cache
− 0.70
5.02 ×
TeCoP + cache
− 0.35
7.08 ×
Minerva MATH
Fast-dLLM + cache
− 0.80
3.40 ×
Appendix
Table 9: Average accuracy change and geometric-mean speedup relative to Original .
Method
Acc. (%)
NFE ↓
TPS ↑
Spd. ↑
LLaDA-8B-Instruct
Original
77.6
512.0
8.0
1.00 ×
Fast-dLLM
76.9
97.0
44.1
5.50 ×
DSB
77.7
97.8
43.2
5.39 ×
TeCoP
76.2
75.6
57.5
7.18 ×
LLaDA-1.5
Appendix
Table 10: Results on GSM8K with generation length 512 , without KV caching. TeCoP uses the length- 256 configurations without retuning. Spd. is the TPS speedup over Original, computed from unrounded TPS.
Model
Original / Fast-dLLM
TeCoP
LLaDA-8B-Instruct
18.53
16.17
LLaDA-1.5
18.51
16.16
Dream-v0-Base-7B
15.21
16.13
Dream-v0-Instruct-7B
15.22
16.15
Appendix
Table 11: Peak allocated GPU memory (GiB) on GSM8K without KV caching. Original and Fast-dLLM have identical peak memory in this setting.