acceptodds
Under review as a conference paper at ICLR 2027

DIMOS: Doob-Guided Inference-Time Multi-Objective Search for Scientific Design

Abstract

Scientific design often requires jointly satisfying multiple objectives and constraints. Pretrained masked diffusion models provide a generative foundation for this task, but fine-tuning them to meet these objectives and constraints incurs additional training costs, motivating inference-time guidance with frozen models. However, such guidance faces two challenges: pass-or-fail constraints and black-box reward models may provide no useful gradients, while jointly satisfying multiple requirements can leave a small feasible region, making feasible designs difficult to find within a limited inference budget. To address these challenges, we introduce DiMOS, a training-free framework for multi-objective scientific design. Using joint rewards from candidate completions, DiMOS performs approximate Doob-guided local resampling without requiring reward gradients. To allocate computation efficiently, it uses budget-efficient trajectory search to focus computation on promising continuations. Across six DNA, protein, and RNA tasks, DiMOS attains the highest joint success rate at comparable generation times, up to the strongest baseline on DNA and protein, while maintaining high sequence uniqueness and naturalness. Our code is available at https://anonymous.4open.science/r/DiMOS/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Masked Diffusion Model, Training Free, Sequence Design

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_M24QOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This is a paper develops an inference time methods for generating DNA/RNA/proteins using doob h-transform where the doob transition kernel is estimated using monte carlo samples. To speed-up monte carlo sample generation, the author took the denoiser in one step as a surrogate.

Critical strengths
  1. well written with nice illustrations.
  2. results in table 1 and 2 appear to have improved upon baseline methods.
Critical weaknesses
  1. Insufficient benchmarking and missing literature, the author should explain the relation/difference between their methodologies and TFG-Flow.

    Haowei Lin, Shanda Li, Haotian Ye, Yiming Yang, Stefano Ermon, Yitao Liang, and Jianzhu Ma. TFG-Flow: Training-free guidance in multimodal generative flow. In International Conference on Learning Representations, 2025. URL https://arxiv.org/abs/2501.14216

    Doob-h transform has been used to training time finetuning of masked diffusion models for example in this paper which the authors should discuss relations with. Both papers appears to target the same tilted target distribution but from different sides.

    Yuyuan Chen, Shiyi Wang, Peter Potaptchik, Jaeyeon Kim, and Michael S. Albergo. Discrete tilt matching. In Proceedings of the 43rd International Conference on Machine Learning, volume 306 of Proceedings of Machine Learning Research, pp. 17141–17164.

    There is also the concurrent work of D-DOIT which the author should explain connections:

    Jieke Wu, Qijie Zhu, Weimin Wu, Zeqi Ye, Minshuo Chen, and Han Liu. D-DOIT: Training-free adaptation of discrete diffusion via Doob’s h-transform. arXiv preprint arXiv:2610.04938, 2026. URL https://arxiv.org/abs/2610.04938

    PMLR, 2026.2. I don't remember the original SVDD paper did experiments for table 1 and 2. The authors should sweep parameters when adapting SVDD to these new settings to make comparisons fair.

  2. Inference methods always have reward hacking concerns. In table 1 did the authors score ATAC with the oracle that guides the model? If so how does the authors mitigate for reward hacking.

Questions for authors

Per Weakness 1,2,3 above.

3 comments
AuthorsOpenReview ↗

Response to M24Q [1/2]

We sincerely thank the reviewer for recognizing the clear presentation, illustrations and improvements in Tables 1 and 2. We also appreciate the suggestions on related work, baseline adaptation and independent evaluation. We address these three concerns below. New experiments are marked (new), and manuscript revisions are shown in blue.

W1 and Q1. Relation to TFG-Flow, Discrete Tilt Matching and D-DOIT

We thank the reviewer for highlighting these related works. We have incorporated all three into the revised Related Work section. DiMOS focuses on joint feasibility across multiple biological objectives, combining a coordinate-wise enumerated transition with trajectory search to generate sequences that satisfy several property requirements simultaneously.

TFG-Flow. Lin et al. [1] estimate guided discrete transition rates by weighting clean samples with a target predictor. DiMOS holds a proposed context fixed, enumerates distinct token alternatives at one newly revealed position, and selects using explicit prior-and-value weights. Each candidate is evaluated against the combined multi-objective reward, concentrating candidate evaluations on different choices for the same local decision.

Discrete Tilt Matching. Chen et al. [2] learn the reward-tilted unmasking posterior through a cross-entropy training objective. DiMOS constructs an approximate guided transition at inference time while keeping the pretrained model frozen. New combinations of objectives and feasibility thresholds can therefore be supplied through the reward without retraining the generator.

D-DOIT. The concurrent work of Wu et al. [3] samples candidate next states from a reverse-transition proposal and selects among them using completion-based values. Its optional search branches into independent continuations and selects the best terminal output. DiMOS changes both candidate construction and search allocation. It enumerates distinct tokens within a shared context and repeatedly selects and replicates promising partial trajectories during denoising. This directs subsequent computation toward trajectories with higher estimated joint feasibility.

Controlled Evidence. Appendices N and O examine these contributions through controlled SVDD comparisons. Replacing DiMOS's enumerated candidates with i.i.d. draws reduces joint success by 47.61 and 38.28 percentage points under the soft and hard gates. Transferring DiMOS's search to SVDD improves that baseline, while DiMOS retains a further 35.37- and 35.21-point advantage at comparable generation time. These results support the enumerated local transition and its combination with trajectory search as the algorithmic contributions to multi-objective sequence design.

References

[1] Haowei Lin et al. TFG-Flow: Training-Free Guidance in Multimodal Generative Flow. ICLR, 2025.

[2] Yuyuan Chen et al. Discrete Tilt Matching. ICML, 2026.

[3] Jieke Wu et al. D-DOIT: Training-Free Adaptation of Discrete Diffusion via Doob's hh-Transform. arXiv:2610.04938, 2026.

AuthorsOpenReview ↗

Response to M24Q [2/2]

W2 and Q2. SVDD Adaptation and Hyperparameter Analysis

We appreciate the request for parameter sweeps when adapting SVDD. We rerun SVDD on the DNA and protein tasks with the same frozen priors, task rewards and output counts as DiMOS. Appendix O adds a candidate-width and selection-rule comparison, together with proposal-temperature alignment on DNA. Appendix R further sweeps the selection temperature.

Candidate Width, Selection and Temperature Controls (new). We sweep M=4,12,20M=4,12,20 on DNA and M=4,6,12M=4,6,12 on protein under both gates. Each width is tested with greedy and stochastic selection, and the stochastic arm uses βSVDD=β\beta_{\rm SVDD}=\beta, equal to 0.250.25 on DNA and 0.020.02 on protein on the paper's reward scale. DNA additionally uses both the original proposal temperature and DiMOS's proposal temperature. The table below reports three-seed means for greedy SVDD, which is stronger than its stochastic variant at every tested width at the matched reward temperature. Table 20 includes both DNA proposal temperatures. Both DNA SVDD rows below use the DiMOS-matched temperature, η=3\eta=3. Time is in milliseconds per design.

DomainMethodMSoft SR (%)Hard SR (%)Time (ms)
DNASVDD1260.3649.01246
DNASVDD2064.7457.29435
DNADiMOS498.3987.55235
ProteinSVDD1240.3635.26247
ProteinDiMOS470.6458.38158

Increasing the candidate width improves SVDD but also increases its generation cost. Even the wider settings above remain below DiMOS in both domains and both modes while using more time. Appendix O, Table 20 gives the full width and selection comparison.

Selection-Temperature Sweep (new). We further vary the stochastic selection temperature βSVDD\beta_{\rm SVDD} in the rule pm∝exp⁡(Rm/βSVDD)p_m\propto\exp(R_m/\beta_{\rm SVDD}) over three seeds, including the matched value βSVDD=β=0.02\beta_{\rm SVDD}=\beta=0.02 and values both above and below it. All SVDD settings below take approximately 246--247 ms per design.

SVDD selection-temperature sweep on protein ΔΔG\Delta\Delta G + SeqRec. Joint SR is mean ±\pm standard deviation over three seeds, following Appendix R.

SelectionSoft SR (%)Hard SR (%)
Argmax40.36±0.0040.36\pm0.0035.26±0.2135.26\pm0.21
βSVDD=0.005\beta_{\rm SVDD}=0.00540.43±0.5840.43\pm0.5832.83±0.0832.83\pm0.08
βSVDD=0.01\beta_{\rm SVDD}=0.0140.41±0.1340.41\pm0.1332.94±0.1432.94\pm0.14
βSVDD=0.02\beta_{\rm SVDD}=0.0238.85±0.1538.85\pm0.1532.57±0.1732.57\pm0.17
βSVDD=0.04\beta_{\rm SVDD}=0.0438.24±0.4038.24\pm0.4032.07±0.1632.07\pm0.16
βSVDD=0.08\beta_{\rm SVDD}=0.0833.81±0.2733.81\pm0.2731.81±0.4131.81\pm0.41

As βSVDD→0\beta_{\rm SVDD}\to0, stochastic selection concentrates on reward-maximizing candidates. The sweep yields similar soft-gate performance for low-temperature stochastic selection and argmax, while argmax exceeds the best tested stochastic setting by 2.32 percentage points under the hard gate. Neither closes the gap to DiMOS. Appendix N further isolates the local transition by transferring DiMOS's search layer to SVDD. Together, these comparisons test the adapted baseline beyond its original operating point.

W3 and Q3. Independent ATAC Evaluation

We thank the reviewer for raising this concern. Yes, the original experiments use a fixed pretrained ATAC predictor to define the accessibility objective and score the generated designs. Applying this predictor consistently across methods provides a common criterion for comparing their ability to satisfy the specified objective.

Cross-Predictor Evaluation (new). To examine whether the gains extend beyond this predictor, we trained a new ATAC evaluator on a separate ATAC-seq experiment and rescored the same generated sequences from every method. The new evaluator never participates in guidance. Appendix M, Table 17 reports the full comparison, with three-seed means for SVDD and DiMOS reproduced below. Retention is the fraction of originally accepted designs that remain accepted under the new evaluator.

MethodGateOriginal SR (%)New evaluator SR (%)Retention (%)
SVDDSoft55.1648.3982.77
DiMOSSoft98.3992.1993.70
SVDDHard45.0542.5088.77
DiMOSHard87.5584.5896.61

DiMOS retains the highest joint SR and acceptance retention among all methods in both gate modes. These results show that its performance advantage transfers beyond the predictor used for optimization.

Dear Authors,

Thank you for the answers, which have addressed my concerns. I continue to recommend accept.

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

This paper proposes DiMOS, a training-free framework for multi-objective biological sequence design using pretrained masked discrete diffusion models. The method combines approximate Doob-guided local resampling with a budget-efficient trajectory search strategy. The method is evaluated on six DNA, protein, and RNA design tasks involving multiple property objectives and constraints. The authors report substantial improvements in Joint Success Rate over several baselines, with up to 1.98× improvement at comparable generation times.

Critical strengths
  • The paper demonstrates substantial empirical improvements across six DNA, protein, and RNA design tasks. In particular, DiMOS achieves significantly higher joint success rates than SVDD under comparable inference-time budgets, while generally maintaining high sequence uniqueness and competitive biological quality metrics.
Critical weaknesses
  • The paper presents Doob-guided inference as a central methodological contribution. However, the underlying Doob h-transform formulation is mathematically equivalent to the soft-value-based transition reweighting already established in SVDD. Furthermore, the completion-based approximation of the h-function closely resembles the posterior-mean value estimation used in SVDD-PM.
  • The main differences therefore appear to lie in candidate construction, stochastic resampling, and trajectory search rather than in a fundamentally new guidance formulation. The paper does not sufficiently clarify this distinction, potentially overstating its methodological novelty.
  • The paper's central effectiveness claim, including the reported up to 1.98× improvement, relies primarily on Joint Success Rate, which is defined using fixed, task-specific acceptance thresholds. However, the manuscript does not sufficiently justify how these thresholds were selected or whether they correspond to biologically meaningful criteria. This threshold-dependent binary metric, relatively modest changes in continuous property scores can potentially lead to disproportionately large differences in success rates.
Questions for authors
  • What is the fundamental methodological novelty of the proposed Doob-guided framework compared with SVDD?
  • Could the authors provide both biological justification and the selection process for the thresholds used in calculating the joint Joint Success Rate?
  • Could the authors provide controlled comparisons using stochastic SVDD sampling and independently ablate candidate construction and selection strategies under matched computational budgets?
2 comments
AuthorsOpenReview ↗

Response to HqFS [1/2]

We sincerely thank the reviewer for recognizing the substantial improvements across DNA, protein and RNA, together with high sequence uniqueness and competitive biological quality. We address the distinction from SVDD and the interpretation of the success criteria below. New experiments are marked (new), and manuscript revisions are shown in blue.

W1, W2, Q1 and Q3. Local Transition and Controlled SVDD Comparisons

We thank the reviewer for requesting a clearer account of the algorithmic contribution and controlled comparisons. We first distinguish the shared guidance principle from DiMOS's algorithmic design, then use controlled comparisons to examine the sources of its improvement. Within the Doob/soft-value framework, DiMOS combines the following three components for multi-objective sequence design.

  1. Candidate enumeration. DiMOS evaluates distinct tokens at one selected position within a shared context CC, while SVDD samples complete next-state proposals independently.
  2. Prior-weighted local resampling. DiMOS samples the enumerated alternatives using the token-prior and completion-reward weights of Eq. 13, while greedy SVDD ranks its sampled candidates solely by completion reward.
  3. Trajectory search. Stage-wise branching and intermediate reward-guided replication allocate computation to continuations with higher estimated joint feasibility.

Candidate Construction and Selection Ablations (new). Table 21 separately tests candidate construction and stochastic selection on protein ΔΔG\Delta\Delta G + sequence recovery. Selection →\to MAP takes the highest-weight entry of the same enumerated menu, retaining both the prior and reward terms. Candidates →\to i.i.d. replaces the menu with token draws at the same position, keeping scoring and resampling unchanged. All three configurations take approximately 156--158 ms per design. The table below reports three-seed means.

ConfigurationSoft SR (%)Hard SR (%)
DiMOS70.6458.38
Selection →\to MAP70.4957.86
Candidates →\to i.i.d.23.0320.10

Under the same scoring and resampling rule, replacing the enumerated menu causes a substantial loss, while switching stochastic selection to MAP has little effect in this comparison. These results support the enumerated candidate construction used by DiMOS.

Stochastic and Temperature-Matched SVDD (new). Appendix O, Table 20 compares greedy SVDD with stochastic selection at βSVDD=β\beta_{\rm SVDD}=\beta, equal to 0.250.25 on DNA and 0.020.02 on protein. It covers three candidate widths in each domain under both gates and includes DNA proposal-temperature alignment. Greedy SVDD is stronger at every tested width at the matched reward temperature. Appendix R.3 further sweeps the selection temperature on protein, where low-temperature stochastic selection and argmax yield similar soft-gate performance, while argmax leads the best tested stochastic setting by 2.32 percentage points under the hard gate. Neither closes the gap to DiMOS.

Search Transfer. Appendix N, Table 18 transfers DiMOS's search layer to SVDD. With MM, KK and the split matched, DiMOS still improves joint SR by 35.37 and 35.21 percentage points under the soft and hard gates, respectively, at comparable generation time.

AuthorsOpenReview ↗

Response to HqFS [2/2]

W3 and Q2. Threshold Selection and Continuous Property Evaluation

We appreciate the request to explain the thresholds and examine whether binary scoring amplifies the gains. Appendix F explains their selection, Appendix J examines performance across six thresholds, and the main tables report complementary continuous property scores.

Biological Rationale and Threshold Selection. We select thresholds using biological reference points and empirical calibration. For protein stability, zero ΔΔG\Delta\Delta G uses the wild type as the reference, consistent with DRAKES and Nisonoff et al. [1,2]. For criteria requiring calibration, we examine score distributions and sweep candidate thresholds so that both objectives constrain the designs while baseline success remains measurable. Ctrl-DNA illustrates how dataset score distributions affect the difficulty of off-target thresholds [3]. Appendix F reports calibration examples, and all methods and both gate modes share the same thresholds within each task.

Threshold Sensitivity and Continuous Property Improvements. Joint SR measures how often generated designs satisfy all specified requirements. The reported improvement therefore concerns joint feasibility, while continuous property scores describe the quality of the generated outputs.

To examine sensitivity to the acceptance threshold, Appendix J, Table 14 evaluates six ΔΔG\Delta\Delta G thresholds on the three-objective task. DiMOS achieves the highest joint SR at every level, reaching 3.03.0 to 7.27.2 times that of the strongest baseline. Its advantage persists as the feasible set becomes rarer, showing that the ordering in this comparison does not depend on a single threshold.

The continuous-score comparison below provides complementary evidence. DiMOS improves both property medians on the main DNA and protein tasks, showing that the joint-SR gains are accompanied by improvements in the underlying properties. The soft-gate medians are computed over all outputs, averaged over three seeds and reproduced from the main tables.

DomainPropertySVDD MedianDiMOS Median
DNAActivity4.665.95
DNAATAC0.830.99
ProteinΔΔG\Delta\Delta G0.380.51
ProteinSequence recovery0.530.60

We further replace the gated guidance reward with a weighted sum of standardized continuous property scores in Appendix Q, Tables 26 and 27. DiMOS achieves higher joint SR and higher medians for both properties than SVDD at all three tested weights on the DNA and protein tasks. These results extend the observed advantage to continuous-score guidance.

References

[1] Chenyu Wang et al. Fine-Tuning Discrete Diffusion Models via Reward Optimization with Applications to DNA and Protein Design. ICLR, 2025.

[2] Hunter Nisonoff et al. Unlocking Guidance for Discrete State-Space Diffusion and Flow Models. ICLR, 2025.

[3] Xingyu Chen et al. Ctrl-DNA: Controllable Cell-Type-Specific Regulatory DNA Design via Constrained RL. NeurIPS, 2025.

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

The paper proposes DiMOS that guides frozen masked diffusion models toward biological sequences that jointly satisfy several property thresholds, using reward evaluations rather than reward gradients. It constructs a scalar joint reward from hard or soft property gates and approximates a Doob transition by completing candidate partial sequences and evaluating their terminal rewards. At each guided update, it enumerates distinct alternatives at one selected token position and samples using both prior probabilities and estimated future rewards. Across six DNA, protein, and RNA tasks, the paper reports comparisons, with additional predictive checks and approximately comparable generation times.

Critical strengths
  1. Experiments show improvements against baseline. A useful controlled comparison separates local guidance from search. There are more credible evaluation than scoring with the optimized predictor.

  2. The theory exposes an approximation that can otherwise be overlooked.

Critical weaknesses
  1. The novelty is narrower than the Doob framing suggests. SVDD already guides frozen continuous and discrete diffusion models using future soft values without reward gradients. The paper should identify the distinct enumeration and selection rule as the main algorithmic advance, and compare its precise transition against a stochastic, temperature-matched variant.

  2. The local theorem does not justify the distribution of the final algorithm. Appendix explicitly excludes branch-level pruning from the bound. Replacing every branch with the best branch and returning the best terminal design changes the distribution again. Consequently, the exact reward-tilted target in Proposition 1 is a motivation rather than a sampling guarantee for DiMOS.

  3. Experiments do not fully support claims. The approximation diagnostics cover only late decoding regime. Also, several principal comparisons are not actually equal in wall time. Best-of-N and SVDD are explicitly budget calibrated, whereas other methods largely retain defaults, and does not establish the best tuned SMC/FK frontier. “Multi-objective” is demonstrated with two requirements. There is no study of three, or more conflicting constraints.

Questions for authors
  1. How do full-algorithm terminal distributions deviate from the nominal reward tilt after repeated best-branch replication?

  2. Can authors report early/middle-time value-estimation errors and protein error quantiles, rather than only three-mask medians?

  3. What happens with at least three conflicting requirements and a controlled sweep of feasible-set rarity?

  4. Do the gains survive equal-time tuning of SMC/FK-Steering and a reward-guided remasking sampler?

  5. Can authors provide pairwise sequence diversity, and the relationship between guidance and evaluation predictor training data?

5 comments
AuthorsOpenReview ↗

Response to SoEF [1/4]

We sincerely thank the reviewer for the careful and constructive review. We address each concern below. New experiments are marked (new) below, and manuscript revisions are shown in blue.

W1. Algorithmic Contribution and Comparison with SVDD

We thank the reviewer for requesting a precise comparison with SVDD. DiMOS combines coordinate-wise candidate enumeration, prior-weighted local resampling, and budget-efficient trajectory search for multi-objective sequence design. We first explain how these components construct and use the approximate Doob transition, then present controlled experiments that separate the effects of candidate construction, selection and search.

  1. Coordinate-wise enumeration in place of full-state sampling. SVDD draws MM next states i.i.d. from the frozen transition and keeps the one with the highest completion reward. DiMOS draws one prior transition, selects the newly revealed position JJ with the largest denoiser entropy, and holds the shared context CC fixed while enumerating the MM most probable distinct tokens there, in place of MM draws that can coincide. The resulting menu contains distinct candidates that differ only at the selected position, directing reward evaluations to alternatives for the same local decision.
  2. Prior-weighted selection over the enumerated menu. DiMOS weights the enumerated alternatives by both token-prior probability and completion reward, selecting with pm∝μθ,tℓJ,η(am∣x)exp⁡ ⁣(R(x^0(y(m)))/β)p_m\propto\mu_{\theta,t_\ell}^{J,\eta}(a_m\mid x)\exp\!\bigl(R(\hat x_0(y^{(m)}))/\beta\bigr) (Eq. 13). SVDD samples candidates from the prior and ranks them solely by completion reward.
  3. Budget-efficient trajectory search. The local transition is combined with stage-wise branching and intermediate reward-guided branch selection, concentrating computation on continuations with higher estimated joint feasibility.

Candidate Construction and Selection Ablations (new). We separately test candidate construction and stochastic selection while keeping the remaining configuration fixed on protein ΔΔG\Delta\Delta G + sequence recovery. Selection →\to MAP takes the highest-weight entry of the same enumerated menu, retaining both the prior and reward terms. Candidates →\to i.i.d. replaces enumeration with MM token draws at the same position, keeping scoring and resampling unchanged. The three-seed means below use approximately 156--158 ms per design for all three configurations. Table 21 gives the complete results with standard deviations.

ConfigurationSoft SR (%)Hard SR (%)
DiMOS70.6458.38
Selection →\to MAP70.4957.86
Candidates →\to i.i.d.23.0320.10

Under the same scoring and resampling rule, replacing the enumerated menu causes a substantial loss, while switching stochastic selection to MAP has little effect in this comparison. These results support the enumerated candidate construction used by DiMOS.

Stochastic and Temperature-Matched SVDD (new). We replace argmax with stochastic selection proportional to exp⁡(R(x^0(y(m)))/βSVDD)\exp(R(\hat x_0(y^{(m)}))/\beta_{\rm SVDD}), using the completion reward on the scale defined in the paper. We set βSVDD=β\beta_{\rm SVDD}=\beta, matching DiMOS at 0.250.25 on DNA and 0.020.02 on protein. Appendix O, Table 20 compares both rules at M=4,12,20M=4,12,20 on DNA and M=4,6,12M=4,6,12 on protein under both gates over three seeds. Each pair shares the candidate count and value estimator.

Table 20 includes both DNA proposal temperatures. The Activity + ATAC comparison below uses the DiMOS-matched proposal temperature, η=3\eta=3, and reports three-seed mean joint SR and generation time in milliseconds per design, rounded to the nearest millisecond.

MethodMSoft SR (%)Hard SR (%)Time (ms)
SVDD (greedy)1260.3649.01246
SVDD (stochastic)1238.6532.86246
DiMOS498.3987.55235

As βSVDD→0\beta_{\rm SVDD}\to0, stochastic selection concentrates on reward-maximizing candidates. At the matched temperature, greedy SVDD achieves higher joint SR at every tested width. Appendix R.3 further sweeps the temperature over three seeds on protein. Low-temperature stochastic selection and argmax yield similar soft-gate performance, while argmax leads the best tested stochastic setting by 2.32 percentage points under the hard gate. Neither closes the gap to DiMOS.

Local Kernel and Search Contributions. Appendix N transfers the search layer to SVDD to separate its contribution from the local transition (Table 18). Search improves SVDD. With MM, KK and the split matched, DiMOS improves joint SR by a further 35.37 and 35.21 percentage points under the soft and hard gates, respectively, at comparable generation time. Together, these comparisons support the enumerated local transition and its combination with trajectory search as the algorithmic contributions to multi-objective sequence design.

AuthorsOpenReview ↗

Response to SoEF [2/4]

W2, Q1 and Q2. Local Theory and Empirical Diagnostics

We appreciate the request to connect the local analysis with the behavior of the complete algorithm. We have added terminal-distribution measurements and early/middle value diagnostics so that the local approximation and the final outputs can be examined together.

Theoretical Analysis and Full-Algorithm Evaluation (W2). The paper's theoretical claims concern the ideal Doob process and the approximation error of the local kernel, not exact sampling by the search-augmented algorithm. The appendix explicitly places branch-level pruning outside the bound. Proposition 1 characterizes the ideal reward-tilted terminal law, and Theorem 1 quantifies the local transition error through context reweighting, top-MM truncation and value estimation. To address the reviewer's question about the terminal distribution, we measure the full-sequence distribution discrepancy of the complete algorithm, including repeated best-branch replication and final selection.

Full-Algorithm Terminal Distribution Discrepancy (Q1, new). To measure the output law after search, we compare complete DiMOS outputs, including repeated best-branch replication and final-best selection, with rejection samples from the nominal reward-tilted target in Proposition 1. Following Appendix C, Table 6, we use unbiased squared MMD with a fixed full-sequence Hamming kernel, 256256 sequences per set, and report the mean and median over 500500 random subsamples of fixed cohorts.

DomainMean MMD$^2$Median MMD$^2$
DNA0.00056790.0005685
Protein0.00233800.0023089
RNA0.00757800.0075932

With both search operations present, mean MMD$^2$ stays below 0.0080.008 in these full-sequence comparisons. These results add a direct measurement of the complete output distribution to the single-step analysis of Theorem 1.

Early and Middle Value-Estimation Errors (Q2, new). The terminal measurement examines complete outputs, while this diagnostic examines the future values used during decoding. Appendix C, Table 4 reports the value-estimation TV errors, with the early and middle measurements reproduced below.

DomainEarly MedianEarly MeanMiddle MedianMiddle Mean
DNA0.08600.10000.08450.1047
Protein0.01020.02080.00070.0192
RNA0.01230.02130.00380.0567

In these measurements, protein and RNA remain below 0.060.06 on both statistics at both stages, and median TV stays below 0.090.09 in all three domains. Appendix C, Table 5 reports the quartiles: the error stays small over most of the sample in every domain, and on protein the lower quartile is below 0.00010.0001 and the upper quartile is 0.02510.0251. Together with the exact three-mask measurement in Appendix C, these results broaden the empirical analysis beyond late decoding.

AuthorsOpenReview ↗

Response to SoEF [3/4]

W3 and Q3. Three Objectives and Feasible-Set Rarity

We thank the reviewer for asking how DiMOS behaves with three competing requirements and rarer feasible sets. We address these together by adding a three-objective protein task and then varying its difficulty through a shared threshold.

Three-Objective Tradeoff Evaluation (new). Appendix I adds hydrophilicity to protein ΔΔG\Delta\Delta G and sequence recovery, giving all seven methods the same requirements. Within this task, ΔΔG\Delta\Delta G and hydrophilicity are negatively correlated in eight of twelve backbones. The prior joint feasible rate is also lower than the rate predicted under independence, at 0.152% versus 0.311% (Table 12). Under this tradeoff, DiMOS achieves 3.03.0 to 3.33.3 times the joint SR of SVDD, the strongest baseline, using 5.4% less generation time under the soft gate and 2.6% less under the hard gate, with 100% uniqueness in both modes (Table 13).

MethodSoft SR (%)Soft Time (ms)Hard SR (%)Hard Time (ms)
SVDD12.98246.010.03246.0
DiMOS42.90232.729.99239.6

Controlled Feasible-Set Rarity Analysis (new). Appendix J varies feasible-set rarity by changing the ΔΔG\Delta\Delta G threshold. Table 14 reports all six rarity levels, with the endpoints and standard threshold reproduced below.

ΔΔG\Delta\Delta G thresholdObserved prior feasibility (%)SVDD SR (%)DiMOS SR (%)
>−1.0> -1.01.6720.1061.00
>0.0> 0.00.1511.5942.08
>+1.5> +1.50.001.9714.19

DiMOS leads at every level and retains 100% uniqueness, including at the strictest level where no feasible design is observed in the prior calibration draw. Its relative advantage also grows with rarity, from 3.03.0 to 7.27.2 times the strongest baseline.

W3 and Q4. Generation Budgets and Reward-Guided Remasking

We thank the reviewer for emphasizing generation cost and suggesting a remasking comparator. We supplement the generation-cost comparisons in Tables 9 and 10 and Figure 4 with new SMC and FK parameter sweeps in Appendix R, then examine whether remasking closes the performance gap.

Baseline Tuning and Generation Cost (new). We additionally sweep SMC's resampling temperature α∈{0.1,0.25,0.5,1,2}\alpha\in\{0.1,0.25,0.5,1,2\} and FK's guidance strength λ∈{10,20,50,100}\lambda\in\{10,20,50,100\} over three seeds. The strongest tested settings remain below DiMOS, despite requiring more generation time. The table below reports the setting with the highest mean SR for each baseline, which is the same under both gates. Appendix R gives the full sweeps.

Tuned baseline comparison on protein ΔΔG\Delta\Delta G + SeqRec. SR is mean ±\pm standard deviation over three seeds, following the appendix. Time is mean ms per design.

MethodParameterSoft SR (%)Hard SR (%)Soft time (ms)Hard time (ms)
SMCα=0.1\alpha=0.19.55±0.809.55\pm0.8010.81±0.3710.81\pm0.3783.3882.09
FK-Steeringλ=50\lambda=5010.94±0.0510.94\pm0.0510.03±0.0510.03\pm0.05192.55192.72
DiMOS--62.89±0.23\mathbf{62.89\pm0.23}51.91±0.19\mathbf{51.91\pm0.19}78.2178.35

DiMOS retains higher SR under both gates at lower generation cost after these sweeps. Tables 9 and 10 extend the lower-cost comparison to DNA and protein. Figure 4 plots DiMOS's empirical SR-time budget envelopes, where measured configurations reach SVDD's SR with approximately 63% less time on DNA and 74% less on protein.

Reward-Guided Remasking (new). We also adapt the remasking refinement of Uehara et al. [1] to SVDD, as reported in Appendix K, Table 15. The two main tasks are summarized below.

DomainGateSVDD + remasking SR (%)Time (ms)DiMOS SR (%)Time (ms)
DNASoft73.9124498.39235
DNAHard60.0524487.55235
ProteinSoft33.6216770.64158
ProteinHard28.1516758.38158

Even with Reward-Guided Remasking, SVDD remains below DiMOS on all four reported tasks under both gates, so the performance gap persists when SVDD is allowed to revisit committed tokens.

References

[1] Masatoshi Uehara et al. Reward-Guided Iterative Refinement in Diffusion Models at Test-Time with Applications to Protein and DNA Design. ICML, 2025.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and New Experiments

We thank all reviewers for their constructive feedback. Following their suggestions, we have added more than ten new experiments and analyses and clarified the algorithmic contribution. The main revisions are summarized below. Manuscript revisions are marked in blue in the revised PDF, and the individual responses discuss the evidence for each question.

Algorithmic Contribution and Baseline Comparisons

  • Clarified the enumerated local transition and its relation to SVDD (Appendix O).
  • Stochastic, temperature-matched SVDD comparisons across candidate widths, including DNA proposal-temperature alignment (Appendix O).
  • Separate controls for candidate construction and stochastic selection (Appendix O).
  • Three-seed sweeps of SMC guidance strength, FK guidance strength and SVDD selection temperature (Appendix R).
  • Comparisons with reward-guided remasking refinement (Appendix K).
  • Expanded discussion of TFG-Flow, Discrete Tilt Matching and D-DOIT (Section 2).

Local Approximation and Terminal Distributions

  • Early- and middle-time value-estimation diagnostics across DNA, protein and RNA, together with error quantiles (Appendix C).
  • Full-sequence MMD measurements of complete DiMOS outputs against the nominal reward-tilted target, including branch replication and final selection (Appendix C).

Multiple Objectives and Success Criteria

  • A three-objective protein evaluation combining stability, sequence recovery and hydrophilicity (Appendix I).
  • Controlled feasible-set rarity experiments across six stability thresholds (Appendix J).
  • Clarified the biological reference points and empirical calibration used to select thresholds (Appendix F).
  • Continuous-score guidance rewards across three objective weights on DNA and protein (Appendix Q).

Diversity and Independent Evaluation

  • Pairwise Hamming diversity and counts of distinct successful designs (Appendix L).
  • ATAC reevaluation using a new predictor trained on a separate ATAC-seq experiment (Appendix M).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.