acceptodds
Under review as a conference paper at ICLR 2027

Value-Informed Schedule Trimming for Accelerated Sequential Monte Carlo Guidance in Discrete Diffusion

Abstract

Reward-guided discrete diffusion enables test-time alignment with downstream objectives, but applying Sequential Monte Carlo (SMC) guidance at every denoising step is computationally expensive. Active guidance steps require additional reward evaluations and, for gradient-based proposals, value-gradient computations. Sparse guidance reduces this cost by guiding only a subset of timesteps, but commonly used schedules are fixed heuristics whose effectiveness can vary substantially across tasks. We propose VISTA, a training-free method that automatically selects guided timesteps under a fixed budget. VISTA formulates sparse guidance as approximation of the ideal full-guidance probability path and derives an estimable upper bound on \(\ell_1\) path deviation. Using value statistics collected from a small number of full-guidance warmup generations, VISTA optimizes the resulting empirical surrogate exactly via dynamic programming. The warmup generations can be retained as model outputs, and after schedule selection VISTA has the same per-generation guidance complexity as other sparse schedules with the same budget. Across text, biological sequence, and image generation tasks, VISTA discovers task-dependent schedules and outperforms hand-designed sparse schedules in most tested settings, while approaching or exceeding full-guidance performance on several task metrics. In toxic-text generation, for example, VISTA attains a 0.99 unique-toxic ratio compared with 1.00 for full guidance while reducing post-selection per-generation latency from 204.23 s to 38.16 s, a 5.35 speedup.

Primary area:
generative models
Keywords:
Discrete Diffusion Models, Sequential Monte Carlo, Inference-Time Guidance, Reward-Guided Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dgEZOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes VISTA, a training-free scheduler that automatically selects which timesteps to apply reward guidance in reward-guided Sequential Monte Carlo (SMC) for discrete diffusion models. The key idea is to formulate sparse guidance as an approximation problem over the ideal full-guidance probability trajectory {πt∗}\{\pi_t^*\}. The authors derive an estimable upper bound, UTD(G)\mathrm{UTD}(\mathcal{G}), on the ℓ1\ell_1 trajectory deviation TD(G)\mathrm{TD}(\mathcal{G}), expressed in terms of expected value functions. Using only a small number MM of full-guidance warmup runs to estimate V^t\hat{V}_t, VISTA then finds the optimal guidance schedule under a fixed budget constraint ∣G∣=T′|\mathcal{G}| = T' via dynamic programming. The method is evaluated on Binary MNIST, toxic text generation, CIFAR-10, DNA, and protein fitness tasks, and compared against manually designed scheduling strategies such as Interval-kk, Uniform, Top-VV, and Top-ΔV\Delta V.

Critical strengths
  1. The problem formulation is well-motivated. Determining when to apply guidance in sparse SMC remains a relatively underexplored research direction, and the observation that existing schedules largely rely on fixed heuristics is valid.

  2. Approximating the entire probability trajectory rather than only the final marginal distribution is conceptually meaningful. The intuition that correcting only the final timestep may lead to problems under a finite particle budget is also convincing.

Critical weaknesses
  1. UTD(G)\mathrm{UTD}(\mathcal{G}) is monotonically decreasing with respect to ∣G∣|\mathcal{G}|, meaning that the unconstrained optimum of the proposed objective is always full guidance. However, the experimental results show the opposite trend: in four out of six benchmarks, full guidance, which is treated as the gold standard, ranks among the worst-performing methods. If the optimum of the proposed objective corresponds to one of the worst empirical outcomes, the claim that better performance arises from closely approximating the full-guidance trajectory is difficult to justify. In Section 5.5, the authors briefly attribute this discrepancy to the fact that overly frequent correction can reduce particle diversity. However, this explanation fundamentally calls into question the validity of the objective function underlying the entire paper.

  2. The evaluation metric is circular and fails to reflect actual performance.The S^(G)\hat S(\mathcal{G}) columns in Tables 3/4/8/9/10/11 and the UTD columns in Tables 6/7 directly report the objective optimized by VISTA. Moreover, the same estimated V^\hat V is reused for evaluation, introducing in-sample optimization bias. Excluding these circular metrics, only runtime and Unique Valid remain as independent evaluation metrics. More importantly, the proposed surrogate fails to predict actual performance rankings. These ranking reversals raise a fundamental concern: the surrogate objective, whose optimization constitutes the core methodology, does not consistently correlate with actual performance.

  3. The 5.35$\times$ speedup claimed in the abstract is not attributable to VISTA and excludes warmup overhead. In Table 8 (T′=15T'=15), Interval-1 through Interval-5 take 33.98--34.82 seconds, whereas VISTA takes 38.16 seconds, making it slower under the same guidance budget. The speedup comes entirely from sparse guidance, which all baselines share, rather than from VISTA's scheduling strategy. Moreover, the reported speedup excludes the additional cost of full-guidance warmup runs.

  4. Presentation and Writing Issues. The paper contains several inconsistencies and proofreading errors:

  • Line 1022 contains an uncorrected LaTeX artifact ("'latex).
  • Figure 4 discusses CIFAR-10 but refers to ``unique toxic samples'' in its caption.
  • Table 9 contains a formatting error: ``20.43 ±\pm 027.02''.
  • The Full Guidance row in Table 4 omits uncertainty reporting (3.12).
  • Sections 5.4 and B.3.1 describe nearly identical settings but report inconsistent values for α\alpha (0.3 vs. 0.2) and warmup time (19.5s vs. 18.7s).
  • ``TDS'' appears in Table 11 without any definition or citation.
  • The main text provides almost no quantitative results, relegating them to the appendix, while the axes in Figures 2/3/4/6 are difficult to read.

These issues collectively undermine the clarity and overall presentation quality of the paper.

Questions for authors

See the weakness.

5 comments
AuthorsOpenReview ↗

Intro & W1. The full-guidance reference and finite-particle outcomes

Thank you for examining the objective, independent evidence and runtime accounting. We have added direct path-fidelity measurements and a three-trial compute-budget comparison, and brought quantitative results into the main text. Revisions are in blue. Locations below refer to the revised PDF.

VISTA solves the fixed-budget problem in §4.2. Full guidance has zero population path deviation. With only T′T' guided steps available, the question is which placements best approximate that path. Theorem 4.3 bounds this deviation, and DP optimizes the resulting fitted objective.

The reference πt∗\pi_t^* is induced by the reward-regularized control objective in §3.2. Distinct-valid yield also depends on the joint law of a finite particle batch. Even two equiprobable valid states give 1.5 expected distinct states under two independent draws and 1.0 under two identical copies, with the same individual marginals. In the B200 protein rerun, 16-step VISTA produced 255 distinct valid candidates versus 147 for Full across 20 batches. The gain comprised 20 more valid outputs and 88 fewer duplicates. Under uniform diffusion, Weighted VISTA produced 180 versus 58, despite 11 fewer valid outputs, with 133 fewer duplicates. These counts quantify the endpoint difference. They leave its ancestry mechanism open.

We have clarified Definition 3.1, Figure 1 and the bound's population scope (pp. 4–5), and added the counts in §5.5 and Appendix D.4, Table 15 (pp. 9, 34–35).

AuthorsOpenReview ↗

W2. Independent evidence for the scheduling criterion

We have labeled the fitted S^\hat S/UTD columns as selection diagnostics. We then enumerated every schedule in ten small diffusion models and evaluated its actual path deviation. At 3/4/6 steps, the mean Spearman correlations were 0.872/0.746/0.572, with selected-schedule mean ranks 3.6/220, 24.0/495 and 351.8/924. The bound is loose in magnitude (12–19×12–19\times). Its ranking is most informative at small budgets. The finite-particle reward and distinct-yield rankings are reported alongside it (Appendix D.2, Table 13, pp. 33–34).

We also scored frozen masked-protein schedules on disjoint warmup traces. VISTA did not have a consistent held-out surrogate advantage over Top-ΔV\Delta V or Uniform in that check. We have reported it separately from held-out generation quality in Appendix D.3 (p. 34). The measured DNA output below provides an independent practical evaluation.

AuthorsOpenReview ↗

W3. What VISTA achieves at the reported cost

At 15 guided steps on toxic text, VISTA achieved 0.99 Unique Toxic Ratio in 38.16 s, versus 1.00 in 204.23 s for Full. The faster schedules at that budget achieved at most 0.89. Sparsity supplies the saving over Full, and placement determines the quality attained with those calls. The 5.35×5.35\times ratio is post-selection latency. We have reported the 639.6 s startup and the startup-inclusive comparison in §5.2 and Appendix D.6 (pp. 7, 37).

On DNA, every method paid for its own tuning and any required warmup within a nominal 1,800 GPU-second budget. VISTA outproduced Uniform in all three trials, by 2.10×2.10\times, 1.37×1.37\times and 1.68×1.68\times. The ratios to Top-ΔV\Delta V were 1.33×1.33\times, 0.65×0.65\times and 1.39×1.39\times. Both value-based methods paid for the same warmup and received its outputs. Rates count distinct valid candidates within each batch and sum across batches. We have included all trials in §5.4, Table 3 (pp. 8–9), and measured sampling-time accounting in Appendix D.1, Table 12 (p. 33).

Reviewer_S8xNOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses when to apply SMC guidance in reward-guided discrete diffusion. Applying guidance at every denoising step is expensive, and existing sparse schedules (fixed intervals, uniform spacing) are heuristics. VISTA formulates sparse guidance as approximating the full-guidance probability path, defines Total Deviation as the average ℓ1 distance between the full-guidance and reduced-guidance marginals, and proves an upper bound expressible through value expectations along the full-guidance path. These expectations are estimated from M full-guidance warmup runs, and the resulting budgeted objective is maximized exactly by dynamic programming in O(T²T′). The method is training-free and proposal-agnostic. Experiments cover Binary MNIST, toxic text, CIFAR-10, DNA design, and protein fitness, against Interval-k, Uniform, Top-V, and Top-ΔV at matched budgets, plus a weighted extension for non-masked diffusion.

Critical strengths
  1. The question is well-posed and under-explored. Most work on SMC guidance studies how to propose and weight particles; when to spend the guidance budget is a reasonable orthogonal axis. And the derivation from Total Deviation to an estimable surrogate is clean, requires neither the intractable marginals nor Rmax, and the budgeted problem admits an exact DP solution rather than a greedy heuristic.
  2. This method is training-free, proposal-agnostic, and cheap to apply, with warmup generations retained as outputs. Also, the empirical scope is broad.
Critical weaknesses

I am mainly worried about optimization target which could be contradicted by the paper’s own experimental results. VISTA’s entire derivation rests on treating the full-guidance path as the approximation target. Definition 3.1 calls π*_t “the gold-standard path that VISTA aims to approximate,” and both the Total Deviation objective and the bound in Theorem 4.3 inherit that premise.

However, the experiments repeatedly show full guidance is not the best configuration: Protein Fitness (Table 4): full guidance gives 9.00 unique valid, while VISTA at T′=16 gives 14.15, and even the naive Interval-1 gives 13.20. Uniform-diffusion protein fitness (Table 10): full 3.15 vs VISTA 9.95. CIFAR-10 (Table 9): full 2.18 vs VISTA 4.28 at T′=15. So on three of five tasks, closer approximation of the “gold-standard” path corresponds to worse downstream performance. Section 5.5 notes this and attributes it to diversity loss under frequent correction, but the consequence is not addressed: if the full-guidance path is not the configuration one should aim at, then minimizing deviation from it is not the right objective, and VISTA’s wins on those tasks cannot be explained by the mechanism the paper claims. The method may work for reasons orthogonal to its derivation.

This is consequential because the derivation, the theorem, and the DP are the paper’s primary contribution. If the target they optimize is not the right one, what remains is a heuristic scheduler whose margin over Top-ΔV is small and untested for significance.

Questions for authors
  1. Several reported margins are small relative to the stated variability, e.g. Protein Fitness at T′=32 (12.00 ± 3.32 for VISTA vs 11.80 ± 2.82 for Uniform over 20 runs). Could you report paired per-seed tests for VISTA against Top-ΔV and Uniform at every budget? In how many of the tested settings does the advantage remain after correcting for multiple comparisons?

  2. Full guidance is outperformed by sparse schedules on Protein Fitness (Table 4: 9.00 vs VISTA’s 14.15 and Interval-1’s 13.20), uniform-diffusion Protein Fitness (Table 10: 3.15 vs 9.95), and CIFAR-10 (Table 9: 2.18 vs 4.28). Given this, in what sense is π*_t the right approximation target on these tasks? If closer path fidelity is not what produces VISTA’s gains there, what mechanism do you believe is responsible?

2 comments
AuthorsOpenReview ↗

Intro & Q2. Why approximate the full-guidance path?

Thank you for recognizing the importance of deciding when to guide, and the value of exact schedule optimization without training. We have directly tested path fidelity, finite-batch diversity and the practical comparison with strong schedules. Revisions are in blue. Locations refer to the revised PDF.

The reference πt∗\pi_t^* comes from the reward-regularized control objective in §3.2. VISTA asks where a limited number of corrections can best approximate that population path. Distinct-valid yield additionally depends on how the particles in a finite batch co-occur. For two equiprobable valid states, two independent draws yield 1.5 expected distinct states, while two identical copies yield 1.0, despite identical individual marginals. A higher distinct count can therefore coexist with a population reference defined by full guidance.

We quantified that difference in the B200 protein rerun, with counts summed over 20 batches of 32.

BackboneScheduleValidDistinct validDuplicates
MaskedFull, 128 steps501147354
MaskedVISTA, 16 steps521255266
Uniform diffusionFull, 128 steps42958371
Uniform diffusionWeighted VISTA, 16 steps418180238

For uniform diffusion, 133 fewer duplicates account for the higher distinct yield despite 11 fewer valid outputs. Resampling can replicate lineages. These endpoint counts quantify repetition, while the specific ancestry mechanism remains unmeasured. In the submitted CIFAR comparison at 15 steps, VISTA yielded 4.28 distinct valid images versus 2.18 for Full. Ancestry was also unmeasured there. We have clarified the reference in Definition 3.1 and Figure 1 (pp. 4–5), and added the decomposition in §5.5 and Appendix D.4, Table 15 (pp. 9, 34–35).

The path criterion has independent numerical support. On ten enumerable models with kernel-consistent exact values, the selected three-step schedule ranked 3.6 out of 220 by actual path deviation, with mean UTD–TD correlation 0.872. At four/six steps the correlations were 0.746/0.572. The bound is loose in magnitude, and particle-level reward and diversity rankings differ. Appendix D.2, Table 13 (pp. 33–34), reports all budgets and finite-particle outcomes. This establishes what the objective captures while quantifying its relationship with the downstream metrics.

AuthorsOpenReview ↗

Q1. Paired tests against Top- and Uniform

We have reported all available paired records, including the comparisons below. Entries are VISTA minus baseline with nominal paired 95% CIs. Protein uses 20 B200 seeds. Text uses 15 prompt-level pairs, averaging two runs per prompt. CIFAR uses ten paired seeds.

TaskStepsΔ\Delta vs Uniform [95% CI]Δ\Delta vs Top-ΔV\Delta V [95% CI]
Masked protein16+1.15 [-0.42, +2.72]-0.15 [-1.41, +1.11]
Masked protein32-0.35 [-2.37, +1.67]-1.45 [-3.09, +0.19]
Uniform-diffusion protein16+2.15 [-0.17, +4.47]+1.80 [-0.64, +4.24]
Uniform-diffusion protein32+3.95 [+1.30, +6.60]+3.60 [+0.41, +6.79]
Toxic text5+0.37 [-0.92, +1.65]+0.13 [-1.57, +1.83]
Toxic text10+0.13 [-0.26, +0.53]+0.47 [-0.12, +1.05]
Toxic text15+0.40 [-0.33, +1.13]+0.27 [-0.06, +0.60]
CIFAR follow-up150-2.40 [-6.86, +2.06]+1.40 [-3.68, +6.48]
CIFAR follow-up200-3.80 [-10.44, +2.84]+1.70 [-1.60, +5.00]
CIFAR follow-up250-1.80 [-3.26, -0.34]-1.10 [-2.43, +0.23]

Among these 20 contrasts, point estimates favor VISTA in 13. The uniform-diffusion protein comparison at 32 steps against Uniform survives Holm correction across all available baselines in that setting (adjusted p=0.0228). Treating the 20 requested comparisons as one family gives no rejection (minimum adjusted p=0.114). Correcting the two requested contrasts within each setting gives two VISTA gains at 32-step uniform-diffusion protein and one Uniform gain at 250-step CIFAR. We have stated the family explicitly in Appendix D.5, Table 16 (pp. 35–36).

The CIFAR rows cover follow-up budgets 150/200/250, with 100/100/10 candidates per seed. Paired records for the original CIFAR budgets, Binary MNIST, DNA and continuous MNIST are outside this analysis. Appendix D.3 also reports the protein rerun alongside the submitted results, including the reversal between VISTA and Interval-1 at 16 masked steps (p. 34).

The practical benefit is supported by the compute-budget comparison and performance across settings. On DNA, VISTA delivered 1.37–2.10×1.37–2.10\times Uniform's distinct-valid rate in all three trials after charging warmup and tuning. Ratios to Top-ΔV\Delta V were 1.33, 0.65 and 1.39 (§5.4 and Appendix D.1, pp. 8–9, 33). Across eight settings with all nine policies, VISTA had the best mean rank, 1.88, and smallest observed maximum regret, 24.3%, versus 37.7% for Top-ΔV\Delta V and 41.4% for Uniform (Appendix D.5, Table 17, p. 37). These observations give a practitioner evidence for choosing VISTA before knowing which schedule will perform best on a new task.

Reviewer_aGAkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method VISTA that selects a fixed-budget subset of diffusion timesteps at which to perform expensive reward-guided Sequential Monte Carlo updates. It formulates scheduling as approximating the full-guidance sequence of marginal distributions and derives an upper bound on average L1 marginal deviation. A few full-guidance warmup runs estimate value statistics, after which dynamic programming optimizes an empirical scheduling surrogate. The resulting schedule can be reused with either base or gradient-guided SMC proposals without retraining the generator. Experiments on images, toxic text, DNA, and protein fitness are reported, although the learned schedule does not win every downstream comparison.

Critical strengths
  1. The paper isolates a concrete computational decision. Exact optimization is available for the selected surrogate. The dynamic program avoids an additional approximate optimizer over timestep subsets.

  2. The idealized bound has an understandable derivation.

Critical weaknesses
  1. The theory does not bound the error of the practical finite-sample scheduler. Appendix establishes guided marginals only as both particle count N and value-rollout count J tend to infinity, whereas text and DNA use N=8 and J=4. Theorem assumes the ideal marginals and true values. There is no concentration or optimization-error bound connecting the estimated schedule to the actual finite-N/J/M path. The upper bound can be too loose to explain finite-particle behavior. The bound uses a global Rmax and an expectation at the last guided time, rather than explicitly measuring weight variance, ancestry loss, or propagation over each skipped interval. Low temperature makes value weights highly concentrated, and DNA uses alpha=0.01.

  2. Experiments cannot fully support claims. Maximizing the surrogate can choose a worse schedule on the reported task metric like at DNA T'=24 VISTA has the higher surrogate score, and at protein T'=32 Interval-1 yields 12.6 v.s. VISTA's 12. These show that exact surrogate optimization is not evidence of downstream optimality. A fixed number of guided steps is not the same as an optimal fixed wall-clock budget. The method leaves all denoiser evaluations in place and compares schedules largely at fixed N, J, and T'. A sparse heuristic could spend its saved warmup time on additional particles, rollouts, or output candidates. The role of existing adaptive resampling is also not isolated. It's necessary to compare schedule gains under several ESS thresholds and resampling policies, with reward-evaluation and backward-pass costs reported separately.

Questions for authors
  1. How variable is the selected schedule and its held-out quality across independent three-run warmup sets?

  2. Can authors compare actual marginal deviation with UTD and downstream quality on an exactly tractable example?

  3. What is the total-time break-even point against Uniform and Top-Delta-V when warmup outputs are counted consistently?

  4. Does schedule selection still help after jointly tuning particle count, rollout count, and ESS thresholds under a common time budget?

4 comments
AuthorsOpenReview ↗

Intro & Q4. Joint tuning under a common compute budget

Thank you for the precise questions about estimation, schedule variability and compute accounting. Schedule selection improved DNA output under the requested joint-tuning comparison in all three trials against Uniform. We have added that experiment and answered the four questions below. Revisions are in blue, and locations refer to the revised PDF.

At 24 guided steps, VISTA, Top-ΔV\Delta V and Uniform each received 1,800 GPU-seconds for sampling during tuning, any required warmup, and held-out generation. All searched N∈{4,8,16}N\in\{4,8,16\}, J∈{2,4,8}J\in\{2,4,8\} and ESS thresholds 0.50/0.98 with three seeds per cell. VISTA and Top-ΔV\Delta V selected schedules from the same three warmup runs, charged in full to each, with their outputs retained. We interleaved methods and repeated the protocol on three disjoint seed blocks.

VISTA relative toTrial 1Trial 2Trial 3
Uniform, tuned workload2.10×2.10\times1.37×1.37\times1.68×1.68\times
Top-ΔV\Delta V, tuned workload1.33×1.33\times0.65×0.65\times1.39×1.39\times
Uniform, matched configurations1.31×1.31\times1.44×1.44\times1.41×1.41\times

The metric sums within-batch distinct valid candidates per unit time. The matched comparison pools the same 18 tuning cells and excludes startup. Splitting those cells by ESS threshold preserves the VISTA/Uniform advantage at both levels. We have reported the sampler, hardware, timing boundary and all trials in §5.4, Table 3 and Appendix D.1, Table 12 (pp. 8–9, 33). The resampling rule was fixed to partial resampling. Timing includes in-pipeline reward and guidance operations as a combined cost. Other resampling rules and separate forward/backward timings remain open.

AuthorsOpenReview ↗

Q1. Independent warmup sets and held-out quality

On masked protein, five disjoint M=3M=3 warmups produced mean pairwise schedule Jaccards of 0.327/0.330 at 16/32 steps, versus 0.136/0.184 for Top-ΔV\Delta V. On the same 20 fresh evaluation seeds, VISTA's mean distinct-valid counts were 12.99 ±\pm 1.01 and 12.90 ±\pm 1.15, where SD is across fitted schedules. Uniform yielded 11.95/10.40, Top-ΔV\Delta V 13.87/12.03 and Interval-1 14.70/12.90.

A further five-fit sweep at each of M=3,6,12M=3,6,12 gave VISTA means of 13.52/13.47/13.69 at 16 steps and 12.75/12.89/13.00 at 32. Held-out quality varied less than the selected step sets. We have revised Appendix B.4 and added the measurements in Appendix D.3, Table 14 (pp. 24, 34). That appendix also reports the B200 replication. Masked VISTA at 16 steps changed from 14.15 to 12.75, and Interval-1 from 13.20 to 13.45, reversing their ordering. These executions do not isolate a hardware cause.

AuthorsOpenReview ↗

Q2 and W1. Exact deviation and finite-sample scope

We enumerated all schedules on ten length-five problems with horizon 12. Value functions were computed by backward recursion through the same reverse kernel used for the marginals and particle simulations. At 3/4/6 guided steps, mean Spearman correlations between UTD and actual TD were 0.872/0.746/0.572. The selected schedule ranked 3.6/220, 24.0/495 and 351.8/924 on average by TD. UTD/TD averaged 12.3–18.7.

With eight particles, 400 batches per schedule and 97 prespecified schedules per budget, UTD–distinct-yield correlations were −0.328/−0.183/−0.064, while UTD–mean-reward correlations were +0.557/+0.570/+0.484. Thus the experiment directly tests the path criterion and measures its downstream associations. We have reported the full protocol and independent validation of pilot-selected candidates in Appendix D.2, Table 13 (pp. 33–34).

Theorem 4.3 concerns population marginals. We have also added a conditional estimation result in §4.3 and Appendix A.5 (pp. 6, 18). If every timewise mean estimate is within ϵ\epsilon, population surrogate regret is at most 2(T−T′)ϵ2(T-T')\epsilon. A finite-N,J,MN,J,M path-error guarantee requires control of those estimation errors and particle effects. The toy temperature α\alpha=0.3 leaves DNA's concentrated-weight regime, α\alpha=0.01, open.

AuthorsOpenReview ↗

Q3. Startup-inclusive break-even

We have charged VISTA and Top-ΔV\Delta V the same startup and credited warmup outputs consistently. For toxic text at 15 steps, their post-selection batch times are 38.16 and 43.07 s, versus Uniform's 44.91 s. With the 639.6 s warmup counted as three delivered batches, VISTA first finishes the same number of batches sooner than Uniform at batch 78. The threshold is 95 if warmup outputs are discarded. Relative to Top-ΔV\Delta V, the common startup cancels and VISTA is faster from the first subsequent sparse batch.

For cumulative distinct-valid protein yield, we used measured warmups of 9.07/8.38 s for masked/uniform diffusion. On uniform diffusion, the projected crossover with Uniform occurs after 18/6 sparse batches at 16/32 steps. Against Top-ΔV\Delta V, the common startup again cancels, and VISTA's measured rates are 7.31 vs 5.73 and 6.27 vs 3.95 distinct valid/s. Masked protein has lower rates than Top-ΔV\Delta V at both budgets. The completion-time and yield projections, with their assumptions, are in Appendix D.6, Tables 18–19 (p. 37).

We have addressed surrogate optimality separately from generation quality. The paired tests and the eight-setting rank/regret comparison are in §5.5 and Appendix D.5 (pp. 9, 35–37). VISTA has the best observed mean rank, 1.88, and its equal-compute DNA advantage over Uniform remains after tuning the particle and rollout budgets.

Comments on the paper

1 comment

We thank the Area Chair and reviewers for examining the contribution and practical cost of VISTA. The paper addresses when to spend an SMC guidance budget. A correction can require several reward evaluations and a backward pass. VISTA selects those timesteps automatically from a few full-guidance warmup runs, optimizes the fitted schedule objective exactly by dynamic programming, and reuses the schedule without training the generator or a scheduler. Warmup generations can themselves be retained as outputs. We have revised the manuscript in blue. All locations below refer to the revised PDF.

The additional experiments support three contributions.

  • Useful output under a common compute budget. On DNA, each method received 1,800 GPU-seconds, including sampling time spent on its own tuning and any required warmup. VISTA delivered 1.37–2.10×1.37–2.10\times Uniform's distinct-valid yield rate in all three trials. VISTA and Top-ΔV\Delta V shared the same warmup, charged to each. Their comparison varied across trials, with VISTA/Top-ΔV\Delta V ratios of 1.33, 0.65 and 1.39. Counts are deduplicated within each batch and summed. The main text presents all three trials together in §5.4, Table 3 (pp. 8–9), with accounting in Appendix D.1, Table 12 (p. 33).
  • A computable criterion for choosing correction times. Theorem 4.3 bounds deviation from the reward-tilted population path, and the DP optimizes its fitted surrogate at the chosen budget. In ten enumerable models, the selected three-step schedule ranked 3.6 out of 220 on average by exact path deviation, with mean rank correlation 0.872 between the bound and actual deviation. The criterion tracks its intended target most closely at this tight budget. We report all budgets and finite-particle outcomes in Appendix D.2, Table 13 (pp. 33–34).
  • A practical choice across tasks. Among nine sparse policies in eight common task–budget settings, VISTA had the best mean rank, 1.88, and the smallest observed maximum regret, 24.3%, compared with 37.7% for Top-ΔV\Delta V and 41.4% for Uniform. These are descriptive comparisons. The complete paired tests and coverage are in Appendix D.5, Tables 16–17 (pp. 35–37). On toxic text, VISTA also attained a 0.99 unique-toxic ratio with 5.35×5.35\times lower post-selection batch latency than Full, which attained 1.00. We report the one-time 639.6 s startup and its reuse explicitly in §5.2 and Appendix D.6 (pp. 7, 37).

The revision connects the mathematical objective to measured path fidelity and to the useful output obtained under a sampling budget. It also distinguishes that population objective from finite-batch diversity and quantifies where startup is repaid. We believe these results substantiate VISTA's contribution as a training-free, reusable method for deciding when to guide.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.