acceptodds
Withdrawn from ICLR 2027

Predictive Speculative Coupling: Distribution-Preserving Acceleration of Autoregressive Visual Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Autoregressive visual generators produce images and videos one token at a time, making sequential target-model evaluations the dominant inference cost. Draft-free speculative Jacobi decoding evaluates a block in parallel, but it initializes each new round with distributions recycled from an earlier speculative context. This one-round lag lowers proposal–target overlap, and the loss compounds along the verified prefix as the window grows. We introduce our PSC method, a training-free framework that separates proposal optimization from sampling correctness. PSC extrapolates aligned proposal drift to form a transported distribution, couples the existing proposal sample to that distribution, and then couples the transported sample to the target through exact maximal correction. The composed decoder preserves the complete autoregressive joint distribution for any non-anticipating transport policy. Across Lumina-mGPT-7B, Janus-Pro-7B, and Cosmos-AR-4B, the exact vectorized path reduces target evaluations and end-to-end latency relative to SJD, with the largest exact endpoint reaching 3.07×NFE and 2.51×latency speedups. The evaluated online selector overwhelmingly chooses the β = 0 reference; the results therefore validate the exact decoding interface and its implementation, but do not establish a benefit from nonzero predictive transport.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
autoregressive visual generation, speculative decoding, coupling, distribution transport, lossless acceleration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_reuJOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PSC aims to accelerate autoregressive image and video generation without an auxiliary draft model or a change to the target sampling distribution. Its starting point is that proposal distributions recycled from earlier speculative rounds lag behind the current context, reducing proposal–target overlap. PSC extrapolates changes in aligned distributions to predict a better proposal. A first maximal coupling transports an existing sample to this proposal, and a second corrects it to the target distribution. The method combines this mechanism with vectorized decoding and adaptive prediction strength. The paper presents a distribution-preservation argument and evaluates Lumina-mGPT-7B, Janus-Pro-7B, and Cosmos-AR-4B, reporting fewer target-model evaluations and lower latency than SJD while maintaining comparable measured generation quality.

Critical strengths

Vectorization yields substantial acceleration. Table 4 reduces latency from 13.8091 to 6.8262 seconds with little NFE change, and the no-prediction control helps identify its source. This is the clearest demonstrated practical value.

Critical weaknesses
  1. Prediction has limited demonstrated value beyond vectorization. On Janus, NFE is almost unchanged and prediction is faster at window 64 but slightly slower at window 32. On Cosmos, PSC is 9%–15% slower than MC at all four windows; Lumina lacks a matched vectorized no-prediction control. Moreover, at Janus window 32, MC NFE changes from 241.4704 in Table 2 to 153.3814 for Scalar MC in Table 6, while PSC NFE remains 152.7910; PSC timings also differ. These issues make it difficult to attribute the reported acceleration to the central predictive mechanism.

  2. The distribution-preservation argument leaves a conditional step unexplained. Composing the two maximal couplings is valid under the required assumptions. The unresolved point is why a reused token still follows q conditional on adaptive predictor history and selection. Appendix A.2 asserts this conditional validity without establishing it. Because exact preservation is a central claim, this step matters to the overall assessment; it is a proof-completeness concern, not evidence of demonstrated sampling bias.

Questions for authors
  1. When does prediction outperform matched β=0 decoding? Please extend the control to Lumina and Cosmos and report tokens committed per target call, total tokens processed, and the frequency and benefit of nonzero β choices.

  2. Please reconcile Tables 2, 4, and 6, including differences in samples or timing conditions. Repeated timings and a breakdown of target-forward, prediction, and coupling costs would clarify whether prediction offsets its overhead.

  3. Why does the reused token retain the conditional law required by Theorem 2 under adaptive history and selection? Please specify the dependencies and show how the condition is maintained across rounds.

Reviewer_7xRsOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Predictive Speculative Coupling (PSC), a training-free method for accelerating autoregressive visual generation. It addresses the distributional lag caused by reusing historical proposal distributions in Speculative Jacobi Decoding (SJD). PSC combines historical distribution extrapolation, two-stage maximal coupling, and adaptive selection of prediction strength. Experiments on Lumina-mGPT-7B, Janus-Pro-7B, and Cosmos-AR-4B report speedups over SJD of up to 3.07× in the number of target-model evaluations (NFE) and 2.51× in end-to-end latency while preserving the target sampling distribution.

Critical strengths
  1. Distribution-preserving speculative decoding The two-stage maximal coupling construction separates proposal prediction from target-distribution correction. Under the stated conditions, it preserves the target autoregressive sampling distribution without requiring additional target-model forward passes.

  2. Evaluation across multiple models The experiments cover two image generation models and one video generation model, reporting both NFE and end-to-end latency across different speculative window sizes. This provides reasonable experimental coverage.

Critical weaknesses
  1. The prediction mechanism is insufficiently validated. The paper fails to verify whether changes in historical proposal distributions can reliably predict the target distribution in the next verification round. Furthermore, the counterfactual overlap metric evaluates candidate transport strengths by reusing the target distributions evaluated on the actual transported prefixes; however, varying transport strengths may alter these prefixes and, consequently, the corresponding target conditional distributions. Thus, this metric serves merely as a proxy score under a fixed context, and the paper has not demonstrated its ability to reliably predict actual committed-prefix lengths.

  2. The incremental benefit is not established. In Tables 4 and 6, NFE (Neural Function Evaluations) for the full PSC (Predictive Speculative Coupling) method is nearly identical to that of the Vectorized MC (Maximal Coupling) baseline without predictive transport. Moreover, all tested fixed positive transport strengths increase NFE. In Table 4, vectorization alone reduces latency from 13.81 seconds to 6.83 seconds, while the full method achieves 6.28 seconds. These results are insufficient to substantiate that predictive transport enhances acceptance efficiency; it remains unclear whether the controller primarily falls back to zero transport strength to avoid harmful extrapolation.

  3. The efficiency claims lack consistent support. For Cosmos-AR-4B, across all four speculative window sizes, PSC has approximately 8.7%–15.1% higher end-to-end latency than MC, despite nearly identical NFE. Additionally, for Janus-Pro-7B with a speculative window size of 32, MC has an NFE of 241.4704 in Table 2, whereas Scalar MC has an NFE of 153.3814 in Table 6. This discrepancy fundamentally alters the assessment of PSC's advantage over the baseline. The authors must account for these inconsistencies and demonstrate the incremental gains of their approach using a unified experimental protocol and repeated latency measurements.

Questions for authors
  1. Please validate the predictive value of historical proposal changes and assess whether counterfactual overlap reliably predicts actual committed-prefix lengths.
  2. Please report transport-strength selection frequencies and clarify the independent benefit of predictive transport, given PSC's nearly identical NFE to Vectorized MC in Tables 4 and 6.
  3. Please reconcile the MC results in Tables 2 and 6 and provide repeated latency measurements under a unified protocol, explaining PSC's higher latency on Cosmos-AR-4B.
Reviewer_jc3gOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Predictive Speculative Coupling (PSC), a training-free framework for accelerating autoregressive visual generation while preserving the exact target distribution. PSC introduces predictive proposal transport based on historical distribution changes and employs two-stage maximal coupling to ensure distribution-preserving sampling. Experiments on image and video generation models demonstrate improvements in neural function evaluations (NFE) and inference latency compared with Speculative Jacobi Decoding (SJD).

Critical strengths
  1. The proposed two-stage maximal coupling framework allows the proposal distribution to be modified without changing the target sampling distribution. The theoretical analysis establishes that PSC preserves the complete autoregressive joint distribution under the stated assumptions.
  2. PSC does not require an additional draft model or model fine-tuning, making it applicable to existing autoregressive visual generators.
Critical weaknesses
  1. Limited effectiveness of predictive proposal transport. Table 4 shows that PSC achieves nearly identical NFE to Vectorized MC, while fixed transport strengths even degrade performance. This raises concerns about the actual contribution of the proposed prediction mechanism.
  2. Limited improvements over strong baselines. Although PSC significantly outperforms SJD, its improvements over MC and GS are inconsistent. Moreover, much of the latency reduction appears to come from vectorization rather than the proposed method itself.
Questions for authors

Refer to the weakness part.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.