PaperlayerView exact revision on arXiv

Source-linked AI summary

Planned Diffusion

Daniel Israel, Tian Jin, Ellie Cheng, Guy Van den Broeck, Aditya Grover, Suvinay Subramanian, Michael Carbin

arXiv:2510.18087v2cs.AI

TL;DR

Autoregressive models generate tokens sequentially, while discrete diffusion needs a denoising order that is difficult to choose and can create a quality–latency trade-off. Planned diffusion has a single model plan semantically independent chunks autoregressively and denoise all chunks in parallel. On AlpacaEval, it reports speedups over autoregressive generation with a limited win-rate drop, and its quality continues to improve with more finetuning compute while the baseline plateaus.

  • Problem

    Discrete diffusion generates multiple tokens per step, but choosing a denoising order is difficult and existing heuristics create a steep quality–latency trade-off.

    Source: §1 · p. 1

  • Method

    Planned diffusion uses one model to autoregressively plan semantically independent chunks and then denoise all chunks in parallel.

    Source: §1 · p. 1

  • Results

    On AlpacaEval, planned diffusion achieves 1.27x–1.81x speedup over autoregressive generation with a 0.87%–5.4% drop in win rate, respectively.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Planned diffusion expands the latency–quality Pareto frontier and continues to improve with more training, while the autoregressive baseline plateaus.

    Source: §5 · pp. 7, 9

    The method conditions both its autoregressive and diffusion distributions on prior context tokens.

    Source: §3 · p. 4

Abstract

from arXiv · show

Most large language models are autoregressive: they generate tokens one at a time. Discrete diffusion language models can generate multiple tokens in parallel, but sampling from them requires a denoising order: a strategy for deciding which tokens to decode at each step. Determining a good denoising order is difficult, and existing approaches use heuristics that create a steep trade-off between quality and latency. We propose planned diffusion, a system that trains the model to determine its own denoising order. Planned diffusion uses a single model that transitions between autoregressive and diffusion-based generation: first, the model autoregressively generates a plan that partitions the response into semantically independent chunks; second, the model denoises all chunks in parallel. The autoregressive plan enables the model to define the denoising order itself. On AlpacaEval, planned diffusion achieves 1.27x to 1.81x speedup over autoregressive generation with only 0.87% to 5.4% drop in win rate, establishing a new Pareto frontier for parallel generation with discrete diffusion. Additionally, planned diffusion's instruction following quality continues to improve with more finetuning compute, while the autoregressive baseline plateaus. Our implementation provides simple runtime knobs that offer tunable control over the quality-latency trade-off.

1 INTRODUCTION

Autoregressive generation is sequential, while discrete diffusion requires a difficult-to-choose denoising order that can create a quality–latency trade-off. Planned diffusion has one model plan semantically independent chunks autoregressively, then denoise them in parallel; it reports faster generation with a small win-rate drop on AlpacaEval.

1 INTRODUCTION

  • Autoregressive models generate one token at a time, while discrete diffusion generates multiple tokens per step but requires a denoising order.The paper describes heuristic ordering as difficult and as creating a steep quality–latency trade-off.Source: §1 · p. 1
  • Planned diffusion uses autoregressive planning to partition responses into semantically independent chunks, then denoises tokens in different chunks in parallel.The plan’s control tags define the structure and length of chunks; the runtime scaffolds each with mask tokens before parallel denoising.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • 1.27x–1.81x speedup over autoregressive generation comes with a 0.87%–5.4% drop in win rate, respectively, on AlpacaEval.The paper reports a Pareto-optimal quality–latency trade-off on 805 instruction-following prompts.Source: §1 · p. 2
  • The paper also reports that instruction-following quality improves with more finetuning compute while the autoregressive baseline plateaus.Its contributions further describe runtime knobs for tuning the quality–latency trade-off.Source: §1 · p. 2

2 RELATED WORK

Planned diffusion is positioned among diffusion language models, semantic parallelism, and other parallel-generation techniques. Its distinction is to denoise semantically independent chunks with diffusion, while related semantic-parallel methods use autoregressive decoding within each chunk.

Diffusion Language Models

  • Planned diffusion complements diffusion acceleration methods and can integrate any diffusion sampling strategy; it targets semantic parallelism, unlike block diffusion and planned denoising as characterized here.The paper says block diffusion enforces an autoregressive structure over blocks and planned denoising learns an adaptive denoising schedule.Source: §2 · p. 3

Semantic Parallelism

  • Diffusion models parallelize token-level denoising, but do not learn to exploit independence across larger semantic chunks.The paper defines semantic parallelism as techniques that parallelize over semantically independent chunks of tokens.Source: §2 · p. 3
  • Prior semantic-parallel methods decode autoregressively within each chunk, whereas planned diffusion denoises each chunk with diffusion.The paper characterizes planned diffusion as, to its knowledge, the first to use diffusion for this purpose.Source: §2 · p. 3

Other Parallel Generation Techniques

  • Insertion-based methods parallelize token insertions, while speculative decoding drafts and verifies tokens; planned diffusion instead identifies independent chunks and needs no separate draft model.A single model handles autoregressive planning and parallel generation.Source: §2 · p. 3

3 PRELIMINARIES

The preliminaries contrast autoregressive token-by-token generation with discrete diffusion, which learns to reverse a masking corruption process. They define denoising orders as ordered partitions that determine which positions decode together and explain uniform and entropy-based ordering.

Autoregression

  • Autoregressive decoding sequentially samples each token conditioned on all previously generated tokens.The autoregressive distribution is described as a product of token conditionals.Source: §3 · p. 3

Discrete Diffusion

  • Discrete diffusion learns to reverse a corruption process that incrementally masks tokens, with training based on a variational lower bound on log-likelihood.The forward process independently masks each position at a noise level; the true posterior is intractable.Source: §3 · pp. 3–4
  • For any disjoint observed positions O and queried positions Q, the conditional distribution over Q factorizes under the diffusion training objective.The passage states this factorization follows because the objective decomposes over individual positions.Source: §3 · p. 4

Denoising Order

  • Planned diffusion conditions both autoregressive and diffusion distributions on prior context tokens.Source: §3 · p. 4
  • A denoising order is an ordered partition of token positions: positions in one subset decode in parallel, while subsets decode sequentially.The partition determines how the joint distribution factors across decoding steps.Source: §3 · p. 4
  • Uniform ordering draws a random permutation and decodes one position per step; in practice, the paper says it underperforms entropy-ordered unmasking.The uniform order is implied by the training objective.Source: §3 · p. 4
  • Entropy ordering selects the position with the lowest predicted entropy at each step, so its order depends on tokens revealed so far.The passage identifies this as the default inference strategy of Ye et al. (2025).Source: §3 · p. 4

4 PLANNED DIFFUSION

Planned diffusion alternates autoregressive planning, which partitions a response into semantically independent chunks, with parallel diffusion denoising. Its implementation uses control tags, an attention mask, and a variable-length denoising schedule.

4 PLANNED DIFFUSION

  • A plan partitions the response into semantically independent chunks, allowing tokens in different chunks to be denoised in parallel.The method alternates autoregressive planning with diffusion denoising; within-chunk denoising follows a separate inner order.Source: §4 · p. 4

4.1 FORMAL DESCRIPTION

  • A plan generated from prior tokens specifies K semantically described chunks and lengths, inducing contiguous token ranges whose composed schedule reveals positions across chunks simultaneously.The method decouples the choice of each chunk’s inner denoising order from the outer order over chunks.Source: §4.1 · p. 5

4.2 IMPLEMENTATION

  • Control tags describe chunk topics and predicted lengths, mark asynchronous content, and use <sync/> to defer further planning until dependent chunks are filled.Control tags are stripped from final outputs.Source: §4.2 · pp. 5–6
  • SlimOrca responses are annotated with control tags, with stochastic padding inside chunks to support variable-length generation; malformed annotations are discarded.The dataset is annotated by prompting a GEMINI model with the tags’ syntax and semantics.Source: §4.2 · p. 6
  • Training maximizes joint plan/content probability, uses causal attention for planning and bidirectional attention for denoising, and shares noise level t across chunks over iterations.Planned diffusion sparse attention restricts bidirectional dense attention within each chunk, enforcing full independence between chunks.Source: §4.2 · p. 6
  • The steps ratio r sets denoising steps to r times the longest chunk length; higher ratios improve quality but slow generation.Chunks are denoised in parallel, so the step count depends on the longest chunk.Source: §4.2 · p. 6

5 EXPERIMENTAL EVALUATION

On AlpacaEval, planned diffusion expands the latency–quality Pareto frontier relative to autoregressive and other diffusion-based approaches. Its shorter critical path supports faster generation, while its quality continues to improve with training as the autoregressive baseline plateaus.

Decoding strategies

  • The comparison includes autoregressive and diffusion baselines, Fast-dLLM, Pasta-SFT, Skeleton-of-Thought, planned diffusion, and its sparse-attention variant.PDSA treats concurrently generated chunks as independent conditioned on the plan, using block-sparse attention.Source: §5 · p. 7

Benchmark and metrics

  • Evaluation uses AlpacaEval’s 805 instruction-following prompts, reporting mean response latency and length-controlled win rate against the 16-epoch autoregressive reference.The 16-epoch model was selected as the fixed reference after comparison of autoregressive variants.Source: §5 · p. 7

Quality and Latency

  • Planned diffusion is on the latency–quality Pareto frontier, achieving 1.27× speedup over autoregressive decoding at 49.2% length-controlled win rate versus 50.0% for AR; PDSA reaches 1.81× at 44.6%.Source: §5 · pp. 7–8

Speedup Analysis

  • Autoregressive decoding averages 367.3 critical-path steps, versus 160.0 for planned diffusion and 155.2 for PDSA.The realized speedup is smaller than the critical-path reduction because planned-diffusion steps do more work and reuse the KV cache less.Source: §5 · p. 8

Scaling

  • Diffusion rises from 41.1% to 52.6% LCWR between 2 and 16 epochs, while autoregressive training stays flat at 50.0%.Planned diffusion rises from 44.9% to 49.2%, and its sparse-attention variant from 40.2% to 43.7%.Source: §5 · p. 9

6 ADDITIONAL ANALYSIS

Additional analyses test plan components, chunk-length predictions, and runtime controls. Topic attributes support inference quality, predicted chunk lengths perform best without scaling, and step ratio and confidence threshold offer a smooth quality–latency trade-off.

Denoising Plan Ablation

  • Removing topic attributes cuts LCWR from 46.65% to 23.33%; removing sync tokens lowers it to 42.96% and raises latency from 3.23 s to 5.81 s.Both ablations were trained for four epochs; the authors conclude topic attributes are critical for quality and sync-token removal offers no latency benefit in this setting.Source: §6 · p. 9

Chunk Lengths

  • LCWR peaks at the predicted chunk length, with ±50% deviations reducing quality and no observed systematic over- or under-prediction.Increasing the length-scaling factor above 1.0 also raises latency; additional denoising steps did not improve accuracy.Source: §6 · pp. 9–10

Quality versus Latency Sweep

  • Sweeping step ratio r from 0.25 to 1 and confidence threshold τ from 0.4 to 0.9 produces a smooth quality–latency trade-off.Planned diffusion and PDSA yield higher quality at equal or lower latency for most operating points.Source: §6 · p. 10

7 CONCLUSION

Planned diffusion combines autoregressive planning with parallel diffusion execution to improve the latency–quality trade-off. It expands the Pareto frontier, with gains attributed to a shorter critical path and scalability with training.

7 CONCLUSION

  • Planned diffusion expands the latency–quality Pareto frontier with speedups over autoregressive and diffusion baselines and a minimal quality drop.The conclusion describes the approach as combining autoregressive planning with parallel diffusion-based execution.Source: §7 · p. 10
  • A shorter critical path and scalability with training drive the reported efficiency gain.The conclusion also says instruction-following quality continues to improve with more finetuning compute.Source: §7 · p. 10
  • Runtime controls provide fine-grained control over the speed–quality trade-off at inference time.Source: §7 · p. 10

A DATA ANNOTATION PROMPT

The annotation prompt identifies response segments suitable for parallel generation and marks dependencies with asynchronous and synchronization tags.

A DATA ANNOTATION PROMPT

  • The prompt asks annotators to identify whether a chatbot response may be generated in parallel and tag suitable segments.It is used to annotate training data with Gemini Flash 2.0 at temperature 1.0 and top-p 0.95.Source: A DATA ANNOTATION PROMPT · p. 16
  • Paired <async> tags mark sentences that can be generated independently of following text, with a concise topic description for continuity.Sentences needed as context for later sentences are not suitable for asynchronous generation.Source: A DATA ANNOTATION PROMPT · p. 16
  • A singleton <sync/> tag makes preceding content accessible to subsequent generation.The instructions say to use synchronization sparingly because it introduces significant slowdown.Source: A DATA ANNOTATION PROMPT · p. 16

B TABLE SUMMARY OF EVALUATIONS

The section points readers to a table summarizing Figure 2 results; its caption identifies latency and length-controlled win rate as comparison measures.

B TABLE SUMMARY OF EVALUATIONS

C FULL SPEEDUP DISTRIBUTION

The section reports the speedup distribution, describes KV-cache behavior in planned diffusion, and clarifies its definition of a hybrid model.

C FULL SPEEDUP DISTRIBUTION

  • Across 803 test samples, the mean speedup ratio is 3.08 with a standard deviation of 4.10, and the distribution is described as right-skewed.The passage attributes the skew to a subset of cases with exceptionally high speedups.Source: C FULL SPEEDUP DISTRIBUTION · p. 17, Figure 4 · p. 17

D KV CACHING IN PLANNED DIFFUSION

  • Tokens in an <async> chunk cannot be cached until denoising is complete, while later stages can attend to preceding cached stages.The planning stage uses causal attention and conventional KV caching; bidirectional attention within diffusion chunks does not support caching.Source: D KV CACHING IN PLANNED DIFFUSION · p. 17

E FURTHER DISCUSSION OF HYBRID MODELS

  • Here, a hybrid model means a single architecture capable of both pure left-to-right autoregression and pure discrete diffusion.This definition excludes works that interpolate between the two objectives.Source: E FURTHER DISCUSSION OF HYBRID MODELS · pp. 17–18

F DATASET PERFORMANCE BREAKDOWN

AlpacaEval win rates are reported across five instruction-dataset sources, and the authors observe little sensitivity to source.

F DATASET PERFORMANCE BREAKDOWN

  • The AlpacaEval breakdown covers self-instruct, openassistant, vicuna, koala, and hh-rlhf.The authors report length-controlled win rates and standard errors for each source.Source: F DATASET PERFORMANCE BREAKDOWN · p. 18
  • Win rate is not especially sensitive to the dataset source of the instruction.This is the authors’ reported observation across the five sources.Source: F DATASET PERFORMANCE BREAKDOWN · p. 18
  • Table 2 presents planned diffusion LC Win Rates across different datasets.Source: Table 2 · p. 18

G QUALITATIVE ANALYSIS

The qualitative analysis illustrates how planned spans can support a non-repetitive list when the planner identifies distinct topics, and how poor planning can instead yield generic spans. The attention-mask discussion contrasts the standard and sparse variants.

G QUALITATIVE ANALYSIS

  • The qualitative analysis examines raw outputs to show control tags and the interaction between planning and diffusion.The examples identify <promise>-<topic> as a span plan, <async>...</async> as diffusion-generated text, and [PAD] as unused allocated span capacity.Source: G QUALITATIVE ANALYSIS · p. 18

G.1 SUCCESSFUL EXAMPLE

  • Distinct topic spans for UN, World Bank, NGO, and other careers precede parallel generation, producing a comprehensive, non-repetitive list.The authors attribute the outcome to spans being disambiguated by the planning stage.Source: §G.1 · p. 18

G.2 FAILED EXAMPLE

  • Poor planning yields generic list-part tags instead of specific locations, with the example repeating New York City and Denver across generated entries.The authors present this as a failure mode that can result from ineffective semantic decomposition.Source: §G.2 · pp. 20, 22

H PLANNED DIFFUSION ATTENTION MASKS

  • Standard planned diffusion uses causal attention for planning and bidirectional attention within diffusion chunks; its sparse variant prevents asynchronous blocks from attending to one another.The standard mask allows concurrent chunks to cross-attend, while sparsity is described as enabling more compute-efficient training and inference.Source: H PLANNED DIFFUSION ATTENTION MASKS · p. 22
Loading 2510.18087v2…