acceptodds
Under review as a conference paper at ICLR 2027

Archive-Guided Inference Reallocation for Conditional Generative Models in Offline Multi-Objective Optimization

Abstract

Conditional generative models provide a promising approach to offline multi-objective optimization, but existing inference procedures typically fix all conditioning targets before sampling begins. This prevents them from using information revealed by partially generated solutions. We introduce Archive-Guided Reallocation (ARC), a closed-loop inference controller for pretrained conditional diffusion and flow models. ARC audits intermediate generations, records their estimated trade-offs in an evolving archive, and redirects selected trajectories toward new target objective vectors while preserving their current states and the total number of backbone evaluations. Our experiments separate two effects - reallocating the inference budget improves Pareto-frontier quality, while selecting new targets from the archive improves convergence. On Off-MOO-Bench, closed-loop reallocation improves mean full-set hypervolume by 4.49% over a compute-matched open-loop control and improves all 15 engineering tasks over its backbone. Archive-guided fronts dominate random-target fronts more often (C-metric 0.36 vs. 0.16) at comparable objective-space occupancy. We further show that the benefit depends on how much of sampling remains steerable after a trajectory is redirected, consistent with the weaker, statistically non-significant effects on a flow-matching backbone whose direct guidance is active during the final 20% of integration. ARC requires no retraining, trajectory restarts, true-objective queries during generation, or additional backbone evaluations.

Primary area:
optimization
Keywords:
Offline multi-objective optimization, Conditional diffusion models, Flow matching, Inference-time control.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_TrpcOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces an approach for reallocating conditioning targets used to guide diffusion and flow models during inference. More specifically, the authors focus on offline MOO, and study which conditioning targets should be replaced, and what direction the new conditioning targets should take. To this end, the authors propose Archive-Guided Reallocation (ARC), which acts as a closed loop controller for conditional diffusion and flow models, requiring no retraining, backbone evaluations or objective queries during the generation. The framework retains an archive of conditioning targets, current solution candidates, and auditor (pretrained MLP ensemble) estimates of their ranking at the end of the generation. To improve the archive, the proposed method uses a single objective that takes into account the top 100%, 75% and 50% solution set performance. During each sampling round, the requests to be retargeted are determined based on the leave-one-out hypervolume contribution. The retargeted conditions are then directed based on the predicted landings from the auditor, computed such that the new directions maximize the weighted sum of hypervolumes for the top 100%, 75% and 50% solution sets.

The authors showcase that the balanced hypervolume is submodular, giving a theoretical motivation to perform the retargeting with respect to the hypervolume greedily. To validate the proposed method, the authors evaluate ARC across 48 tasks from offline MOO bench, comparing against surrogate based and generative model based baselines. The results show consistent improvement over PCD, the closely related method that uses fixed sampling targets.

Critical strengths

Authors study a well motivated question of reallocating some of the conditioning targets based on the estimated final performance of the corresponding solutions. Since the real oracle evaluations are expensive, ensuring that each generated sample contributes to the final solution set is a desirable goal.

Critical weaknesses
  • In its current form, the paper is very difficult to read. The flow from one paragraph to another is missing, making it challenging to understand how each component is related to each other, and to the overall picture. Moreover, the text introduces additional mathematical notation frequently, making the text very cluttered.
  • The evaluations are questionable. The authors compare results computed from 3 seeds (baselines from offline MOO benchmark) to 5 seeds (generative methods). Some comparisons include all 48 tasks, while some include a subset of those, yet the authors give no clear indicator why such choices were done in the main text. The text also includes various p-values for different results, without mentioning which test was used to determine the p-value. At minimum, I would expect a small note that these details are described in the appendix.
Questions for authors
  • ARC combines the performance of top 100%, 75% and 50% solution sets, and assigns them coefficients (0.5, 0.3, 0.2) (Eq 1). How sensitive is the proposed method to these coefficients, and how were these coefficients chosen?
  • Table 3 showcases that contribution based source ordering decreases the performance slightly, which is quite unintuitive. Do you have any analysis on why this happens?
1 comment
AuthorsOpenReview ↗

Response to Reviewer: Evaluation and Technical Clarifications

We thank the reviewer for the careful evaluation and constructive feedback. We appreciate the recognition of ARC's motivation and address the concerns below.

W1: Presentation and clarity. We acknowledge that the presentation could be clearer, particularly in connecting the methodological components and introducing notation. ARC separates two decisions: which trajectories to redirect (source selection) and where to redirect them (archive-guided destination selection). We agree that a more structured explanation would improve readability.

W2: Evaluation protocol. We appreciate these concerns. Generative methods use five seeds, while Off-MOO-Bench baselines retain their originally reported three-seed results. The main paired ARC–PCD comparisons cover 48 tasks with matched checkpoints, noise, and sampling budgets. The all-method ranking uses 44 tasks where every method completed its seed budget, while individual ablations cover 45–48 tasks depending on experimental completion. Statistical significance is assessed using two-sided paired Wilcoxon signed-rank tests across tasks, with further details in Appendix C.2. We agree that these distinctions should be clearer in the main text.

Q1: Balanced utility coefficients. The coefficients (0.5,0.3,0.2)(0.5, 0.3, 0.2) were fixed across all tasks to prioritize full-set HV while accounting for the quality of the top 75% and 50% of solutions. Table 3 shows modest improvements from balanced utility (+0.05% at the 100th percentile and +0.19% at the 50th percentile). We acknowledge that a systematic coefficient-sensitivity analysis would strengthen this evaluation.

Q2: Contribution-based source ordering. Table 3 shows that contribution-based ordering slightly underperforms uniform source selection (−0.25%). A possible explanation is that intermediate HV contributions do not reliably predict which trajectories benefit most from redirection. ARC therefore uses contribution scores to protect promising trajectories while uniformly sampling the remaining eligible sources. We acknowledge that this explanation requires further empirical validation.

We sincerely appreciate the reviewer's suggestions and hope these clarifications address the main concerns regarding ARC's methodology and evaluation.

Reviewer_m33vOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposed a new method that wrap the generative modeling PCD method to solve the offline multi-objective optimization which aims to find optimal Pareto-set by leveraging only the offline datasets. However this paper contains too many mismatch and incoherent technical details.

Critical strengths

None

Critical weaknesses

Overall the paper looks super artificial as an LLM-generated materials, lacking of coherent and logical in their technical details. Several serious issues were listed as follows :

  1. No code is provided for reproducible. The implementation details contain too many technical mismatch and looks artificial, make the results be unreliable
  2. Omission of crucial discussion with related works : the author did not provided a clearly discussion to show their advantages over the prior works. Lacking of several sota baselines: nor discussion or empirical comparison with [1], [2]
  3. The presentation is super artificial without coherent : For example, the background comes with conditional diffusion model first (typically the classifier free guidance) but then write the flow model after, and do not technically show how PCD and ParetoFlow use these generative models specifically to solve the offline MOO. These points result in a highly disjoined narrative.
  4. Significant technical inconsistency: The author wrote as using ther proposed ARC to wrap PCD, and call the frozen PCD model as auditor GθG_\theta. However they use GθG_\theta for predict the objective values y^\hat{y} and the auditor uncertainty uu. Meanwhile, the original PCD is a generative modeling method, that try to generate the Pareto sets (a set of designs). This PCD is not a surrogate model that provides objective value estimation or uncertainty predictions. Therefore, how the author could use a generative modeling PCD as a surrogate model?
  5. The archive do not contain the design trajectories xx to return the Q candidate trajectories, how the method maintain these trajectories ?
  6. Too many terminology that was used without introduction : for example, what is the controller, the request, sources and destination to be selected in their method. I believe that these specified terminology was not used in the offline MOO literature. The predicting candidates landing is unclear, where they introduced y~(v)\tilde{y}(v) and just wrote vv as candidate without any clear definition for this mapping. What is the vv here relate to y^,c\hat{y},c or designs xx ? Why this mapping is parameterized as an affine model, why this need to be a linear relationship between vv and y~\tilde{y} ?

References :

[1] Preference-Guided Diffusion for Multi-Objective Offline Optimization

[2] Diversity-Driven Offline Multi-Objective Optimization via Nested Pareto Set Learning

Questions for authors

See Weakness

1 comment
AuthorsOpenReview ↗

Response to Reviewer: Technical Clarifications and Reproducibility

We thank the reviewer for the detailed feedback and careful evaluation of our work. We appreciate the opportunity to clarify the technical concerns and acknowledge that some aspects of the presentation could be improved.

W1: Reproducibility and implementation. We acknowledge the concern regarding code availability. The implementation details, including the controller configuration, auditor training, candidate selection, and evaluation protocol, are provided in Sections 3–4 and Appendices A–D. We agree that releasing the code would further strengthen reproducibility.

W2: Related work and baselines. We thank the reviewer for highlighting these relevant studies. Unlike methods that develop new generative models or Pareto-set learning approaches, ARC focuses on improving inference in pretrained conditional generators through adaptive target reallocation, without backbone retraining or trajectory restarts. We agree that a more explicit comparison and discussion would strengthen the paper.

W3: Presentation clarity. We appreciate this observation. PCD and ParetoFlow are introduced to demonstrate how ARC can operate with both diffusion- and flow-based generators. We agree that explaining their specific roles in offline MOO more clearly would improve the narrative.

W4: Auditor and PCD distinction. We respectfully clarify that ARC does not use the frozen PCD generator as an auditor. As stated in Section 3.1, the auditor is a separately trained ensemble of three bootstrap-resampled MLPs per objective. PCD generates designs, whereas the auditor predicts their objective values and uncertainty. These are distinct components.

W5: Trajectory maintenance. The archive stores historical requests, predicted objective values, and uncertainty estimates, rather than complete generative states. The live trajectories are maintained separately by the backbone. As described in Section 3.3 and Algorithm 1, ARC updates only selected conditioning requests while preserving the corresponding generative states, allowing all Q trajectories to continue until completion.

W6: Terminology and request-to-landing mapping. We appreciate the suggestion for clearer definitions. A request represents a desired objective vector, a source is a trajectory selected for redirection, and a destination is its newly assigned request. The request-to-landing model estimates the objective outcomes likely to result from a given request, using previously observed request–outcome pairs. The affine mapping is a regularized approximation, not an assumption that the underlying relationship is globally linear. Further details are provided in Section 3.3 and Appendix A.2.

We sincerely appreciate the reviewer's comments and suggestions. We hope these clarifications address the main technical concerns and help communicate the methodology and contributions of ARC more clearly.

Reviewer_dZFAOpenReview ↗
Rating
2(−1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ARC, an inference-time controller for conditional generative models in offline multi-objective optimization. ARC maintains an archive of predicted outcomes and periodically reallocates conditioning targets while preserving ongoing trajectories. It combines surrogate auditing, elite protection, and archive-guided destination selection. Experiments cover 48 tasks, primarily using PCD, with additional ablations and a ParetoFlow extension.

Critical strengths
  1. I like the idea of adapting conditioning during generation without retraining the backbone or restarting trajectories. This provides a practical way to improve an existing generator, and the matched-checkpoint experiments show benefits under the same number of backbone evaluations.

  2. I found the empirical analysis fairly thorough. The random-destination, doubled-sampling, and oracle-auditor controls address several plausible alternative explanations. Archive-guided selection also improves filtered-set hypervolume and dominance metrics over random destinations, which gives some support for the feedback mechanism.

Critical weaknesses
  1. I am not fully convinced that the complete controller is necessary. Random destinations already recover much of the full-set HV improvement, while elite protection, balanced utility, and destination perturbations provide small individual gains. The filtered-set results support archive guidance, but I still miss a simple baseline that periodically switches targets using a predetermined schedule or the offline front, without online feedback. This would help establish what the adaptive controller adds beyond target switching itself.

  2. I find the computational comparison incomplete. ARC introduces an auditor ensemble, repeated archive updates, local fitting, and HV calculations, but the experiments match backbone calls rather than elapsed time. Without measured overhead or a time-matched comparison, I cannot tell how attractive the improvement is in practice.

Questions for authors
  1. Could you compare ARC with a simple target-switching schedule that uses random or offline-front targets without online feedback, matching the switching frequency and backbone budget? This would materially affect my assessment of the controller’s contribution.

  2. What are the additional auditor-training cost and per-run inference overhead? Under the same total inference time, how does ARC compare with spending the extra time on backbone sampling and final selection?

1 comment
AuthorsOpenReview ↗

Response to Reviewer: Controller Necessity and Computational Overhead

We thank the reviewer for the careful reading, constructive feedback, and two specific suggestions. We appreciate the positive assessment of the method and empirical analysis. We address both concerns below.

W1/Q1: Is the complete controller necessary?

We agree that target reallocation accounts for much of the full-set hypervolume improvement. As shown in Table 3, archive-guided destinations provide a modest additional gain of +0.78% over random destinations at the 100th percentile (p=0.32p = 0.32). However, their contribution is more substantial when considering the quality and convergence of the returned set.

In Table 2, random destinations improve full-set HV by +3.48% over PCD but provide only +0.41% at the 75th percentile and decrease HV by 5.46% at the 50th percentile. In contrast, ARC improves all three percentiles. Table 4 further supports the value of archive guidance, showing better IGD (0.030 vs. 0.052) and stronger set coverage (0.36 vs. 0.16).

These findings suggest that target switching primarily drives frontier improvement, while archive guidance helps ensure that redirected trajectories produce better-converged solutions across the returned set.

We also acknowledge the reviewer's distinction regarding a fully feedback-free switching baseline. Our random-destination control matches the switching schedule and backbone budget but retains auditor-based elite protection. Therefore, it does not completely isolate predetermined target switching without online feedback. We agree that the suggested fixed-schedule and offline-front baselines would provide a valuable additional comparison.

W2/Q2: Computational overhead and wall-clock comparison

We agree that the computational overhead deserves clearer reporting. As stated in Appendix C.1, our comparisons match backbone network evaluations rather than wall-clock runtime. ARC introduces additional costs from auditor training, intermediate auditing, archive maintenance, realization fitting, and hypervolume calculations. The current paper does not quantify these costs, and we acknowledge this limitation.

Nevertheless, Table 8 provides a related control: PCD with twice the sampling budget, followed by auditor-based selection of the final 256 designs, achieves a mean full-set HV change of −0.04% relative to PCD, compared with +4.23% for ARC. This suggests that additional sampling and final selection alone do not reproduce ARC's improvement. We recognize, however, that this experiment is not a strict wall-clock-matched comparison.

We appreciate both suggestions, which would help strengthen the empirical distinction between simple target switching and adaptive archive guidance, as well as clarify ARC's practical computational trade-offs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.