acceptodds
Under review as a conference paper at ICLR 2027

COMPUTING GUIDANCE SCHEDULES FROM CONDITIONAL INFORMATION IN MASKED GENERATIVE MODELS

Abstract

Masked generative models apply classifier-free guidance (CFG) repeatedly during parallel decoding, yet existing methods schedule guidance using a predefined function of normalized time. The appropriate time-based shape can vary across checkpoints, token grids, and reveal policies, requiring a new curve-family or exponent search for each setting. We introduce InfoProgress, an information-coordinate view of CFG scheduling. A short Monte Carlo probe of the frozen decoder measures the conditional-to-unconditional log-likelihood ratio carried by tokens committed at each round. Accumulating this evidence produces a checkpoint-specific information clock, from which dose normalization computes all per-round guidance weights in closed form. InfoProgress fixes a shared clock-to-guidance mapping and retains only the ordinary one-dimensional CFG-strength scan. On MaskGIT-256, it improves tuned constant CFG from FID 4.481 to 3.218 at 32 network evaluations and surpasses the checkpoint’s reported FID 4.17 at 64 evaluations. Across decoding budgets, token grids, model capacities, and MaskBit codebook widths, it consistently improves constant CFG and remains within 0.28 FID of the best hindsight-selected fixed curve.

Primary area:
generative models
Keywords:
masked generative models, classifier-free guidance, guidance scheduling, parallel decoding, information-based allocation, efficient generation.

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ZyTtOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes InfoProgress, a classifier-free guidance scheduling method for masked generative models. It estimates conditional evidence from accepted tokens during a calibration run and uses the accumulated evidence to construct a guidance schedule. Experiments on MaskGIT and MaskBit show improvements over constant guidance, with performance generally close to that of tuned fixed schedules.

Critical strengths
  • The method does not require model retraining and achieves noticeable improvements over tuned constant CFG. The authors also evaluate it across different decoding budgets, token grids, model sizes, and codebook settings.
Critical weaknesses
  • The proposed approach primarily replaces normalized decoding time with an estimated information-progress measure, followed by a fixed power mapping. The theoretical results are largely straightforward identities and do not establish why this construction should produce better guidance schedules than existing approaches.

  • InfoProgress does not outperform the best tested fixed schedule in any of the eight settings in Table 2, and the held-out results are nearly identical to those of a simple quadratic schedule. The experiments do not convincingly establish the additional value of measuring conditional information, especially given the required calibration.

Questions for authors
  • In Table 2, InfoProgress does not outperform the best fixed schedule in any setting. What is the actual advantage of using the information clock instead of a simple quadratic or linear schedule?

  • Could the improvements mainly come from applying stronger guidance in later decoding steps? How do we know that the conditional-information measure itself is important?

  • Since InfoProgress still requires calibration and guidance-strength tuning, does it actually reduce the overall tuning cost compared with using a fixed quadratic schedule?

Reviewer_aqf3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes InfoProgress, a method for constructing classifier-free guidance schedules for masked generative models. A calibration run measures conditional-to-unconditional log-probability ratios for newly accepted tokens. The accumulated evidence defines an information clock, which is mapped to guidance weights using a power function and normalized to separate schedule shape from average strength. The exponent is selected once and reused across the primary experiments, while the overall guidance strength is tuned for each setting. Experiments on MaskGIT and MaskBit cover different decoding budgets, token grids, model capacities, and codebook sizes. The method improves over tuned constant guidance and performs close to separately tuned fixed schedules. A held-out checkpoint experiment finds that the resulting schedules closely resemble a quadratic ramp and achieve similar FID.

Critical strengths

The construction of InfoProgress is simple and requires no retraining or additional model evaluations during generation after the calibration. The experiments provide evidence that the method can improve on tuned constant guidance in practice. But I think it does not outweigh the weaknesses below.

Critical weaknesses
  1. I am not convinced that the proposed calibration provides enough value over a simple fixed schedule. In Table 2, quadratic guidance has a lower reported FID in six of the eight settings and appears similarly robust across settings. It also avoids the calibration step while retaining the same one-dimensional strength search. The improvements over constant guidance therefore do not establish the benefit of the information clock itself. Given the relatively narrow scope of guidance scheduling, this missing advantage substantially limits the contribution.

  2. The connection between the measured evidence and the guidance allocation is insufficiently justified. Accumulated conditional evidence does not obviously indicate where additional guidance will improve generation quality, and the propositions do not establish that connection. They mainly describe identities and properties of the chosen construction. There is also a possible confounding factor: summing evidence over accepted tokens makes the clock depend on the number of tokens revealed at each round. Without a generation-quality comparison against a revealed-token-fraction clock, it remains unclear whether the conditional evidence contributes useful information beyond the reveal schedule.

Questions for authors
  1. What advantage does InfoProgress offer over fixing a quadratic schedule once and tuning only its strength on subsequent settings? Given Table 2, I would like to understand whether the advantage is generation quality, robustness, or lower total tuning cost.

  2. What motivates making guidance an increasing power of accumulated conditional evidence? Is there an argument or diagnostic showing that this quantity identifies where additional guidance is useful, instead of where the model has already accumulated substantial conditional evidence?

  3. How does InfoProgress compare with the same power mapping and normalization applied to the revealed-token fraction?

Reviewer_rX6tOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes InfoProgress, which schedules classifier-free guidance in masked generative models by an "information clock." A 500-image probe under constant guidance measures, per decoding round, the summed log-ratio log⁡pc/pu\log p_c/p_u of newly accepted tokens; the normalized cumulative profile ItI_t is mapped to weights wt∝(It+ϵ)1.4w_t \propto (I_t + \epsilon)^{1.4} with a fixed token-weighted mean ww, and only ww is tuned. Experiments cover MaskGIT budgets, sizes and token grids, five MaskBit codebooks, a held-out H-MaskGIT-S-384 checkpoint, and Show-o and Meissonic diagnostics.

Critical strengths

The evaluation is careful: every schedule gets its own bracketed strength search, boundary optima are flagged, each comparison stays on one machine, and the held-out protocol and screening rule were fixed in advance.

Critical weaknesses
  1. The method does not improve on a fixed quadratic ramp, disproving its own motivation. Specifically, the paper claims that time shapes must be re-searched for each setting in the abstract, yet one uncalibrated s2s^2 schedule is within 0.281 FID of the best tested schedule in all eight Table 2 settings and beats InfoProgress in six (by 0.008-0.223).
  2. The measured clock is not independently useful. In Figure 2a, the information clock nearly coincides with the revealed-token fraction, which the reveal schedule fixes without any probe, and in Figure 2b the resulting weights overlap s2s^2. No ablation checks the revealed fraction through the same map.
Questions for authors

N/A

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.