acceptodds
Under review as a conference paper at ICLR 2027

Exploratory Biological Sequence Design on Masked Discrete Diffusion Models

Abstract

Biological sequence design under limited experimental budgets requires effective exploration of vast discrete sequence spaces under unreliable proxies trained from scarce data. Conservative search may remain concentrated around previously observed sequences, whereas broader exploration may reach regions where proxy predictions are uncertain. To reconcile conservative and broad exploration, we introduce a masked discrete diffusion framework Reward-tilted Diverse Discrete Diffusion (RD3). Our method consists of three components to precisely control exploration. First, it constructs local neighborhoods of high-scoring sequences by partial masking and denoising. Second, we fine-tune the pretrained masked diffusion model to tilt sequence towards high-scoring regions during denoising. Finally, we apply categorical particle guidance, a variant of particle guidance in continuous space to promote diversity in the batch. Experimental results across nine benchmarks spanning DNA, RNA, peptide, and protein design demonstrate that RD3 effectively discovers high-fitness sequences, outperforming existing methods in terms of maximum fitness. Fitness–diversity and fitness–novelty analyses further demonstrate the effectiveness of RD3 in generating diverse and novel candidates without compromising fitness. Overall, the results highlight the value of combining search locality, reward alignment, and finite-sample diversification to balance fitness, novelty, and diversity under limited oracle budgets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Biological Sequence Design, Generative Model

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_f8DtOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Reward-tilted Diverse Discrete Diffusion (RD3), combining masked discrete diffusion, reward-guided fine-tuning, and particle guidance for biological sequence optimization under limited experimental budgets. The method is evaluated on nine benchmarks spanning DNA, RNA, peptide, and protein design.

Critical strengths
  1. The proposed benchmark is comprehensive, including diverse biological sequence design tasks and MANY NON-RL baselines.
  2. The proposed method achieve good performance, particularly in maximum fitness, and provides additional evaluations of diversity and novelty.
Critical weaknesses
  1. The method mainly combines existing techniques/methods, including masked diffusion, RTB, and particle guidance. The novelty of method is limited. This is my largest concern.
  2. The evaluation for budget-efficiency is limited. Although limited experimental budgets are a central motivation, the evaluation primarily emphasizes maximum fitness, which cannot fully reflect the motivation.
  3. The fitness evaluation is unclear. The biological meaning of fitness scores vary largely across DNA, RNA, and protein tasks.
  4. The ablation metrics are not fully aligned. The main ablation figures emphasize mean fitness, diversity, and novelty, but the main benchmark results (Table 1) focus on maximum fitness.
  5. More baselines and related works, especially for which may be very COMPETATIVE ACROSS THESE BENCHMARKS should be mentioned. For example, [1-3]. More importantly, there are OVERLAPs between this method and the unmentioned related works. At the very least, the authors need to mention them even if not compare them. Otherwise, the scientific claim is soundless.

[1] Tang, S., Zhang, Y., & Chatterjee, P. (2025). Peptune: De novo generation of therapeutic peptides with multi-objective-guided discrete diffusion. ArXiv, arXiv-2412. [2] Cao, H., Shi, H., Wang, C., Pan, S., & Heng, P. A. (2026). GLID 2^ 2 e: A gradient-free lightweight fine-tune approach for discrete biological sequence design. Advances in Neural Information Processing Systems, 38, 139875-139898. [3] Tang, S., Zhu, Y., Tao, M., & Chatterjee, P. (2025). Tr2-d2: Tree search guided trajectory-aware fine-tuning for discrete diffusion. ArXiv, arXiv-2509.

Questions for authors
  1. How is the quality of each generated candidate evaluated, and how do the predicted reward correlate with oracle fitness?
  2. How are the fitness functions constructed for different biological tasks? How to ensure independent evaluation and avoid circular validation given different kinds of fitness?
  3. Why is maximum fitness the primary metric? Would #proportion of candidates passing a fitness threshold better reflect practical experimental outcomes since we also expect more positive candidates during experiments.
  4. What is the practical benefit of higher novelty or diversity?
  5. What is the essential innovation beyond combining existing methods or modules?
5 comments

Dear Reviewer f8Dt,

We appreciate your detailed review and constructive criticism. Your comments have helped us clarify the methodological novelty of RD3, the roles of proxy-based rewards and benchmark oracles, and the practical motivation for diverse exploration. We also provide additional experimental analyses to address your concerns, as detailed below.

[W1, Q5: Methodological novelty and overlap with prior methods]

We thank the reviewer for this comment. As discussed in our response to Reviewer YJaZ W5, the key methodological contribution of RD3 is KL-calibrated categorical particle guidance for masked discrete diffusion, integrated with local reconstruction and reward-aligned generation for budget-constrained exploration. Please refer to that response for a detailed discussion of our novelty and its distinction from prior methods.

[W2: Oracle-query efficiency and budget considerations]

We thank the reviewer for raising the question of oracle-query efficiency. RD3 achieved the highest normalized AUC on seven of nine tasks (Table 1). On AAV, RD3 achieved a mean maximum fitness of 0.7460 with 200 oracle queries, exceeding the best baseline result with 1,280 queries (AdaLead, 0.7160; Table 2). Please see our responses to Reviewer YJaZ W2 for optimization progress across nine tasks and Reviewer YJaZ W3 for performance under reduced query budgets on AAV.

Table 1. Normalized AUC across nine tasks. Bold indicates the highest reported value in each column.

MethodTF-Bind-8RNA-ARNA-BRNA-CAMPAAVPab1E4BGFP
RD30.9851.0261.0001.0460.9620.7691.3778.0383.590
GFN-AL- δ\delta CS0.9810.9720.9681.0210.9400.6311.3097.9793.585
AdaLead0.9930.9360.9710.9370.9440.6301.3438.0033.585
ProSpero—————0.6461.2707.9963.604

Table 2. Maximum fitness on AAV over ten rounds. (B) is the number of oracle queries per round.

MethodB = 20B = 50B = 128
RD30.7460 ± 0.01130.8003 ± 0.02320.8793 ± 0.0230
δ-CS0.6843 ± 0.00940.7096 ± 0.01270.7084 ± 0.0117
AdaLead0.6685 ± 0.01040.7024 ± 0.01340.7160 ± 0.0192
ProSpero0.6839 ± 0.04030.6912 ± 0.03280.7043 ± 0.0233
PEX0.6573 ± 0.02280.6725 ± 0.01600.6585 ± 0.0239

[W3: Biological meaning of fitness]

TF-Bind-8 uses an experimentally measured binding landscape; RNA tasks use ViennaRNA-based energy scores; AAV uses a measurement-derived additive model; and AMP, GFP, Pab1, and E4B use fixed pretrained predictors. We have clarified the oracle definitions and sources in Appendix [D.1] and Appendix [E.6], respectively.

[W4: Maximum versus mean fitness in ablation studies]

We use maximum fitness as the primary evaluation metric, as our main objective is to discover the best-performing sequence under a limited oracle budget. In contrast, mean fitness in our ablation studies evaluates the overall quality of generated candidates, allowing us to assess whether improvements in diversity come at the cost of fitness. We also provide Maximum-fitness results in Appendix [E.8].

[W5: Related work on PepTune, GLID²E, and TR2-D2]

We thank the reviewer for drawing our attention to these important and closely related studies. Monte Carlo Tree Guidance from PepTune, reward-guided fine-tuning of GLID²E, and integration of tree search with trajectory-aware learning from TR2-D2 offer valuable approaches to guided biological sequence design. While these works also investigate guided or reward-based biological sequence generation, RD3 specifically addresses diversity-aware exploration across iterative oracle-query rounds under a limited experimental budget. We have expanded the related-work discussion to better recognize all three studies.

[Q1: Correlation Between Acquisition Scores and Oracle Fitness]

Thank you for requesting this analysis. Candidate quality is estimated using an upper-confidence-bound (UCB) acquisition score, based on the proxy ensemble's predicted mean fitness and uncertainty.

We evaluated the Spearman rank correlation between acquisition scores and oracle fitness for all generated candidates (Full-pool) and the final batches selected for oracle evaluation (Selected-batch). Full-pool Spearman correlations ranged from 0.622 to 0.969, supporting the usefulness of acquisition scores for ranking the candidate pool. Correlations remained positive but were lower within selected batches (0.347–0.598), where the restricted acquisition-score range makes fine-grained ranking more challenging. We report both Pearson and Spearman correlations, along with the evaluation protocol and round-wise results, in Appendix [E.6].

TaskFull-pool SpearmanSelected-batch Spearman
AAV0.969±0.0070.969\pm0.0070.347±0.0530.347\pm0.053
RNA-A0.935±0.0060.935\pm0.0060.490±0.0130.490\pm0.013
RNA-B0.943±0.0050.943\pm0.0050.488±0.0600.488\pm0.060
RNA-C0.882±0.0090.882\pm0.0090.598±0.0210.598\pm0.021
TF-Bind-80.622±0.0100.622\pm0.0100.374±0.0200.374\pm0.020

[Q2: Candidate Evaluation Using Proxy-Based Acquisition Scores]

We apologize that the distinction between the learned proxy and the benchmark oracle was insufficiently clear. The proxy is updated using the measurements available at each round and provides acquisition scores for generation and query selection. Reported fitness is obtained from a separate, fixed benchmark oracle, rather than from the proxy used to guide the search. The benchmark oracle is accessible only during the final evaluation stage of each round and is not available during candidate generation or selection. Instead, RD3 relies on a separately trained proxy to estimate candidate fitness and guide the search, using oracle feedback only to update the proxy for subsequent rounds.

[Q3: Proportion of candidates exceeding a fitness threshold]

MethodAAV hit rateRNA-C hit rateTF-Bind-8 hit rate
RD30.465±0.0830.465\pm0.0830.861±0.0090.861\pm0.0090.780±0.0270.780\pm0.027
GFN-AL- δ\delta CS0.001±0.0010.001\pm0.0010.838±0.0180.838\pm0.0180.719±0.0310.719\pm0.031
AdaLead0.004±0.0060.004\pm0.0060.781±0.0810.781\pm0.0810.731±0.0580.731\pm0.058

We thank the reviewer for highlighting the importance of obtaining multiple high-fitness candidates. Maximum fitness measures the best sequence discovered within a fixed oracle budget; we complement it with the hit rate, defined as the fraction of newly queried candidates exceeding a threshold. We set the threshold to τ=b0+0.5(c−b0)\tau = b_0 + 0.5(c-b_0), where b0b_0 is the best fitness in the initial dataset and cc is the highest fitness observed across the compared methods over the runs and evaluation horizon. RD3 achieved higher mean hit rates than both baselines on the three tasks shown above. Full threshold definitions and results are provided in Appendix [D.3] and Appendix [E.3], respectively.

[Q4: Practical value of novelty and diversity]

In biological sequence design, discovering new functional sequences beyond known datasets is important because existing sequences may already have been investigated, developed, or patented. While novelty helps identify promising sequences beyond the starting collection, diversity provides alternative candidates for validation. As discussed by Jain et al. (2022), early screening is imperfect, so highly similar candidates may fail at later stages. A diverse set of high-fitness candidates may reduce this risk and offer more opportunities to satisfy additional requirements not captured by the optimization objective.

Thank you very much for your rebuttal. For your rebuttal:

  1. Thank you for your clarification. From the ML perspective, I still can not agree the novelty of this method towards this task. It may be a good adaption, but it is not a novel algorithm. Also, the motivation and the "novelty" are mis-aligned.

  2. The response looks good.

  3. I did see Appendix D1 and E(6? I didn't find E.6 actually, please check). I agree they are clear, and from open-source database. But two questions: 1) there should be different oracle/reward design towards different "fitness"; and 2) The exact processing may not follow biological rules. For example, you have no clear definition to AMP activity. I appreciate the workload you have done, but your results and data may not be precise and realiable.

  4. I didn't see E.8.

  5. Thank you. It addressed my concern.

Overall, I think this work still need a very major revision. I will keep my score and confidence.

Reviewer_YJaZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Reward-tilted Diverse Discrete Diffusion (RD3), an active learning framework for biological sequence design across DNA, RNA, peptides, and proteins. It balances fitness, diversity, and novelty by combining local partial masking and unmasking, reward-tilted fine-tuning via Relative Trajectory Balance, and a step-by-step Categorical Particle Guidance repulsion mechanism. Evaluated across 9 benchmark tasks, RD3 outperforms 12 existing baselines by achieving the highest maximum fitness on 7 tasks while maintaining strong sequence diversity.

Critical strengths
  1. Combines Multiple Smart Techniques Well: The paper effectively combines local search (partial masking), reward guidance (Relative Trajectory Balance), and diversity control (Particle Guidance) into a single unified framework.
  2. Strong empirical results presented well.
  3. Clear writing.
Critical weaknesses
  1. Missing Baselines: Key biological sequence optimization baselines such as Feedback GAN (Gupta & Zou, 2019) and IDEAS (Pandey et al., Interpretability Driven Evolutionary Approach for the Design of Biological Sequences, 2026) are missing.

  2. Could the authors include or discuss comparisons against these methods to better contextualize performance? Convergence Speed & Area Under the Curve (AUC): Many methods in Figure 2 eventually converge to similar maximum fitness values by round 10. To show which algorithms reach high-scoring sequences faster and better measure sample efficiency, could the authors compute and report the Area Under the Curve (AUC) for the fitness progression curves across rounds?

  3. Performance Under Reduced Query Budgets (BB): The experiments currently evaluate a query budget of B=128B = 128 (and B=1024B = 1024 for AMP). How does RD3 perform under smaller, more realistic wet-lab budgets, such as B=20B = 20 or B=50B = 50?

  4. Impact of Initial Dataset Size (∣D0∣\vert{}\mathcal{D}_0\vert{}): Pretraining the masked discrete diffusion prior relies on the initial measured dataset D0\mathcal{D}_0. What is the effect of varying the initial dataset size on the quality of the learned prior and overall optimization performance?

  5. Overall paper has limited novelty. It is a combination of existing techniques.

Questions for authors

Mentioned in Weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer YJaZ

Dear Reviewer YJaZ, We thank you for your careful evaluation and valuable feedback. In response to your concerns, we provide additional baseline comparisons, examine performance under reduced oracle-query budgets and initial dataset sizes, and clarify the methodological contribution of RD3.

[W1: IDEAS and FBGAN] We thank the reviewer for highlighting FBGAN and IDEAS. Given the time available during rebuttal, we prioritized a direct comparison with IDEAS, the state-of-the-art approach. We evaluated IDEAS under the same active learning setting as RD3. We selected AAV and RNA-C to cover long and short sequence design tasks, respectively. RD3 achieved higher mean values across all three metrics on both tasks, supporting its effectiveness in the evaluated limited-query setting. We expanded related-work discussion and included IDEAS results.

TaskMethodMax. Fitness ↑Mean Fitness ↑Diversity ↑
AAVRD30.879 ± 0.0230.827 ± 0.02914.546 ± 2.242
IDEAS0.536 ± 0.0140.491 ± 0.0045.789 ± 0.677
RNA-CRD31.115 ± 0.0021.001 ± 0.0067.678 ± 0.390
IDEAS1.026 ± 0.1100.874 ± 0.0923.696 ± 0.922

[W2: Evaluation of query efficiency using AUC] RD3 achieved the highest round-normalized AUC on seven of nine tasks, with TF-Bind-8 and GFP as exceptions. On AAV and AMP, no evaluated baseline run reached the RD3 final-fitness reference within 10 rounds. The results complement final-fitness comparison by showing optimization progress over the available budget. Full results are provided in Appendix [E.2].

MethodTF-Bind-8RNA-ARNA-BRNA-CAMPAAVPab1E4BGFP
RD30.9851.0261.0001.0460.9620.7691.3778.0383.590
GFN-AL- δ\delta CS0.9810.9720.9681.0210.9400.6311.3097.9793.585
AdaLead0.9930.9360.9710.9370.9440.6301.3438.0033.585

[W3: Performance under reduced query budgets] We evaluated RD3 and the baselines on AAV using per-round query budgets of B = 20 and 50 over 10 active learning rounds, results averaged over 3 seeds. RD3 achieves the highest maximum fitness across all evaluated query budgets. Notably, even with B = 20, RD3 outperforms all baselines evaluated with B = 128. These results demonstrate the effectiveness of RD3 under substantially reduced oracle budgets. We will include the complete results and details in Appendix [E.4].

MethodB = 20B = 50B = 128
RD30.7460 ± 0.01130.8003 ± 0.02320.8793 ± 0.0230
δ-CS0.6843 ± 0.00940.7096 ± 0.01270.7084 ± 0.0117
AdaLead0.6685 ± 0.01040.7024 ± 0.01340.7160 ± 0.0192
PEX0.6573 ± 0.02280.6725 ± 0.01600.6585 ± 0.0239

[W4: Sensitivity to initial dataset size] We evaluated RD3 with uniformly subsampled initial datasets on AAV and RNA-C. On AAV, retaining 50% and 25% of the initial data yielded maximum fitness of 0.827 and 0.791, respectively, compared with 0.879 using all data. Both remained above the compared baselines’ full-data results (0.659–0.716). On RNA-C, maximum fitness remained at 1.002 and 0.992, compared with 1.115 using the full dataset, corresponding to reductions of 10.1% and 11.0%, respectively. These results show that RD3 retains substantial performance with less data for prior training, proxy learning, and seed selection. Full details are provided in Appendix [E.5].

TaskInitial datasetSizeMaximum fitness
AAVFull15,3070.879±0.0230.879\pm0.023
AAV50%7,6550.827±0.0240.827\pm0.024
AAV25%3,8280.791±0.0420.791\pm0.042
RNA-CFull5,0001.115±0.0021.115\pm0.002
RNA-C50%2,5001.002±0.1091.002\pm0.109
RNA-C25%1,2500.992±0.1710.992\pm0.171

[W5: Methodological novelty of RD3] We respectfully disagree with the assessment that RD3 lacks methodological novelty. Please note that Categorical Particle Guidance in our methodology is a novel approach. Inspired by the idea of interacting particles in continuous-diffusion PG, Categorical PG formulates these interactions directly through categorical token probabilities to discourage particles from generating highly similar sequences. Target-KL calibration adjusts the guidance coefficient to maintain a target KL divergence between guided and unguided token distributions. These mechanisms promote diverse exploration during candidate generation. The choice of sampling algorithm is orthogonal to our contribution. We adopt RTB as one way to shift the generative distribution toward high-reward regions, enabling more efficient exploration of promising candidates. RD3 is a new method for exploring biological sequence space through successive rounds of generation and oracle evaluation, seeking diverse, high-fitness sequences under a limited query budget.

Reviewer_JzA7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces RD3 (Reward-tilted Diverse Discrete Diffusion) to tackle biological sequence design. RD3 introduces three exploration techniques: (1) local exploration with masked discrete diffusion; (2) reward-tilted posterior learning; (3) categorical particle guidance. Under an active learning loop, RD3 achieves better maximum fitness while balancing diversity and novelty. Ablation further shows the value of reward tilting and particle guidance in RD3.

Critical strengths
  1. RD3 constitute a nice technical fit for the application problem studied. The adaptation of masked diffusion to biological sequence design is well rationalized and the additional use of reward tilting and particle guidance effectively improves the max fitness and diversity results.
  2. Clear presentation with sound in-silicon experimental results. The presentation is well-structured and easy to follow. The main results, trade-off analysis, ablation and sensitivity analysis demonstrates the empirical advantage of RD3.
Critical weaknesses
  1. Insufficient contextualization: Although the paper cites EvoDiff and DPLM, the technical distinctions with these discrete diffusion frameworks remain insufficiently discussed. A more detailed comparison with these approaches would help clarify RD3’s technical contribution.
  2. Limited mechanistic analysis: The ablation results quantify the effects of individual components, but provide limited insight into how they influence the exploration process. More analysis / case studies would help explain when and why reward tilting and particle guidance are effective.
Questions for authors

See Weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer JzA7

Dear Reviewer JzA7,

We sincerely thank you for your thoughtful review and constructive suggestions. We appreciate the opportunity to clarify the relationship between RD3 and prior diffusion-based methods and to provide additional analyses of the roles of reward tilting and particle guidance. We address your comments in detail below.

[W1: Distinction from EvoDiff and DPLM]

We apologize that the original manuscript did not clearly position RD3 relative to EvoDiff and DPLM. EvoDiff focuses on modeling protein sequence distributions and controllable generation. DPLM extends diffusion-based protein generation to property-conditioned sampling through classifier guidance. In contrast, RD3 focuses on diverse and effective exploration over the sequence space through iterative generation and oracle evaluation under a limited query budget.

[W2: Mechanistic roles of RTB and particle guidance]

We thank the reviewer for this valuable suggestion. We conducted controlled analyses using final-round checkpoints on AAV (3 independent runs). More detailed analyses are provided in Appendix [E.8.1].

(1) Reward tilting. We compared the diffusion prior and RTB posterior under identical partially masked inputs, with PG disabled. RTB consistently shifts generation toward higher oracle fitness, improving mean fitness by 20.83%. Corresponding shift in the oracle-fitness distribution is illustrated in Appendix [E.8.1]. This suggests that reward tilting shifts the generative distribution toward high-fitness regions, where categorical particle guidance can then diversify candidate sequences during unmasking.

MetricDiffusion PriorRTB PosteriorImprovement
Mean Oracle Fitness ↑0.3692 ± 0.00210.4461 ± 0.0093+20.83 ± 2.60%

(2) Particle guidance. We conducted an ablation study on particle guidance using the same RTB posterior and identical masked inputs. PG increases expected token diversity at each denoising step by modifying categorical sampling probabilities.(Appendix [E.8.1]) This leads to progressively greater diversity among generated sequences, with a 14.59% increase in final mean pairwise Hamming distance.

MetricPG OFFPG ONImprovement
Mean Pairwise Hamming Distance ↑14.686 ± 0.19516.829 ± 0.273+14.59 ± 0.34%

Together, these analyses provide mechanistic evidence for the complementary roles of RTB in reward-aligned generation and PG in denoising-time diversification.

Thanks the authors for the response. The response clarifies the distinction in problem setting and the role of categorical particle guidance, and I will maintain my positive score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their thoughtful feedback on RD3. Their questions helped us clarify what categorical particle guidance contributes to biological sequence design and evaluate how effectively RD3 uses a limited oracle budget. We summarize the additional experiments and manuscript revisions below.

Additional Experiments

  • Mechanistic analysis. In a controlled AAV comparison with particle guidance disabled, the RTB posterior increased mean oracle fitness by 20.83% relative to the diffusion prior. In a separate ablation, enabling KL-calibrated categorical particle guidance increased candidate-pool diversity by approximately 35% before query selection.
  • Oracle-query efficiency. RD3 achieved the highest normalized fitness-curve AUC on seven of nine tasks. On AAV, RD3 with 200 total queries achieved a higher mean maximum fitness than the best evaluated baseline with 1,280 queries (0.746 vs. 0.716).
  • Sensitivity to initial data. We evaluated RD3 with smaller initial datasets on AAV and RNA-C. On AAV, retaining 25% of the initial data reduced mean maximum fitness from 0.879 to 0.791, still above the best full-data baseline of 0.716.
  • Acquisition scores and oracle fitness. Across five tasks, Spearman correlations ranged from 0.622 to 0.969 in the full candidate pools and from 0.347 to 0.598 within the selected query batches.
  • Yield and diversity of high-fitness candidates. We evaluated the fraction of queries meeting shared retrospective fitness thresholds. At the midpoint threshold on AAV, RD3 achieved a hit rate of 46.5%, compared with 0.4% for AdaLead and 0.1% for GFN-AL-δCS. Mean pairwise sequence distance among hits was 16.17 for RD3 and 5.01 for AdaLead.
  • Additional baseline. Under the same oracle budget, RD3 achieved higher mean values than IDEAS on AAV and RNA-C for all three reported metrics: maximum fitness, top-128 mean fitness, and diversity.

Paper Revisions

  • Methodological contribution. We clarify how RD3 extends the idea of interacting particles from continuous diffusion to masked sequence generation. Categorical PG uses interactions among candidates to update token probabilities during unmasking. Target-KL calibration controls how far these probabilities shift from the unguided distributions. RTB supplies the reward-aligned sampler used in our implementation.
  • Evaluation protocol. We distinguish the learned search proxy from the fixed benchmark oracle and explain how maximum fitness, hit rate, novelty, and diversity capture different aspects of sequence discovery.
  • Related work. We expand the discussion of EvoDiff, DPLM, PepTune, GLID²E, TR2-D2, FBGAN, and IDEAS to acknowledge their contributions and position RD3’s focus on exploration across successive rounds of generation and oracle evaluation.

We appreciate the reviewers’ suggestions, which strengthened our evaluation and helped us state RD3’s contribution more precisely.

Area_Chair_iX5rOpenReview ↗

Reminder: Please Respond to the Authors’ Rebuttal

Dear Reviewers,

Thank you for your time and effort in reviewing this submission. Please take a moment to read the authors’ responses and follow up with any remaining questions or concerns. If the responses have changed your assessment, please update your review and score accordingly.

Your engagement is greatly appreciated and will help us reach a well-informed decision.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.