acceptodds
Under review as a conference paper at ICLR 2027

MoFlow: Multi-Objective Agentic Workflow Generation

Abstract

We study the generation of agentic workflows that jointly optimize multiple objectives, such as accuracy, cost, latency, robustness, and consistency. Existing methods for workflow generation typically optimize accuracy alone or a weighted sum of objectives, so each trained generator commits to one fixed trade-off and must be retrained from scratch when preferences change. To alleviate this, we propose MoFlow, which generates workflows optimized across varied preferences. Specifically, MoFlow formulates workflow generation as a multi-objective Markov decision process and solves it by leveraging Convex-Hull Monte Carlo Tree Search with optimistic set-valued backups, where every node stores a set of reachable trade-offs rather than one weighted score. A single search thus approximately covers the Pareto front, from which MoFlow can return a workflow for any preference by lookup without retraining. We evaluate MoFlow against six strong baselines on six benchmarks spanning mathematics, code, and question answering. Since the baselines are single-scalar optimizers by design, an apples-to-apples comparison is difficult. We instead adopt an evaluation setup that favors the baselines, in that they are rerun for each testing preference, which MoFlow never sees. Even under this stringent setup, MoFlow achieves the highest average hypervolume.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agentic workflow, LLM, Multiple objectives

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gfnLOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MoFlow, a multi-objective framework for automatic agentic workflow generation. It formulates workflow construction as a multi-objective MDP and combines Convex-Hull MCTS, contextual zooming, and GNN-based value estimation to optimize multiple objectives within a single search. Experiments on six benchmarks demonstrate improvements over existing workflow generation methods in terms of hypervolume.

Critical strengths
  1. The paper addresses a practical and important problem by jointly optimizing accuracy, cost, latency, robustness, and consistency in agentic workflow generation.
  2. The experimental evaluation is comprehensive, covering six benchmarks and multiple LLM backbones. The results demonstrate consistent improvements over existing workflow generation baselines.
Critical weaknesses
  1. Insufficient comparison with multi-objective optimization methods. Although the paper compares against several workflow generation baselines, it lacks comparisons with established multi-objective optimization algorithms, making it difficult to assess the effectiveness of the proposed approach specifically for Pareto optimization.
Questions for authors

How sensitive is MoFlow’s performance to the number and distribution of preference vectors sampled during search, particularly when the number of objectives increases?

3 comments

Thank you for the positive and constructive review. We are glad you found the problem important and the evaluation comprehensive. We respond to the two points below.

W1. Comparison with established multi-objective optimization (MOO) algorithms

Our comparison protocol follows published workflow generation papers such as AFlow (ICLR 2025) and MaAS (ICML 2025), and all six baselines are strong workflow optimizers. As noted in the paper (Sec. 4, lines 394–395), the protocol deliberately favors the baselines. They are rerun separately for each of the 11 test preferences, so each run targets the exact weights it is scored at, whereas MoFlow serves all 11 preferences from a single search that never sees them, and it still attains the highest mean hypervolume. Our ablation study (Table 4, lines 456–467) further compares MoFlow with three established multi-objective MCTS algorithms, Pareto-UCB (Chen & Liu, 2019), Hypervolume-UCB (Perez et al., 2015) and Chebyshev (Xu & Li, 2017), by swapping in their action-selection rules, and MoFlow again attains the highest hypervolume with the fewest tokens. Taken together, we consider the comparison already comprehensive, as you also noted in the strengths.

Most generic MOO algorithms are not readily applicable to workflow generation, but if there is any specific work you think we should compare ours against, please let us know and we would be happy to perform comparison.

Q1. Sensitivity to the number and distribution of sampled preferences

MoFlow needs only a few dozen sampled preferences, even with all five objectives. Across the six online searches, hypervolume reaches 90% of its final value after 15 to 40 preferences, and Figure 4 shows one such curve. As for the distribution, we sample preferences uniformly by design. Users' preferences are unknown at search time, so the search should not favor any region of the preference space. The sampling distribution only steers which actions the search explores, while the set-valued backups and serving by lookup never use it. We therefore treat a sensitivity analysis over other sampling distributions as orthogonal to validating MoFlow's design, and we would be glad to pursue it in future work.

Dear Reviewer gfnL,

Thank you again for your valuable comments. We understand this is a busy time, and we would greatly appreciate it if you could let us know whether our rebuttal has addressed your concerns. If there are any unsolved concerns, we will be very willing to continue the discussion.

Best Regards,

Authors

Thank you for your rebuttal and clarifications. Taking into account your response and my overall assessment of the paper, I believe my initial rating remains appropriate and will therefore keep it unchanged.

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

this paper basically does multi-objective optimization, but applied to workflow generation. So it has two parts, first is recognizing the problem of workflow generation as a multi-objective optimization, and the second is a clever method for doing so effectively, contextual zooming trees (CZT).

Critical strengths

It works, and the algorithm looks sensible.

Critical weaknesses

I think this paper very much needs a cost-controlled comparison for MoFlow vs the baselines.

On the x-axis would be dollar amount spent in tokens, and the y-axis should be hypervolume of the multi-objective optimization problem. So we can ask the question, if I spend $100, what is the best hypervolume achievable by some generous interpreation of the baseline methods, and what is the hypervolume MoFlow can achieve.

Questions for authors

do the plot I requested.

2 comments

Thank you for the concrete suggestion. Section 5 of the paper (lines 419–426) already compares the search costs of online MoFlow, offline MoFlow and the baselines, and Table 1 compares the search and inference complexity of MoFlow and each baseline. As you requested, we have prepared a figure that plots held-out hypervolume against search cost. You can view it in this anonymized repository (https://anonymous.4open.science/r/MoFlow-anonymous-7CE3/rebuttal_cost_vs_hv.pdf). FlowSteer and SkillFlow are not plotted because they need to train a workflow generator model. This training requires gradient updates, a cost that is not counted in tokens, so it does not fit the figure.

Observation: A single online MoFlow search that serves all 11 preferences beats every baseline by at least 7.2 HV while costing less than ADAS (the strongest baseline in Table 2) and EvoMAS. Offline MoFlow, which reuses the execution logs of the online search, reaches a similar HV for about 2% of that search cost. We will add this plot to the next revision.

Dear Reviewer PgLb,

Thank you again for your valuable comments. We understand this is a busy time, and we would greatly appreciate it if you could let us know whether our rebuttal has addressed your concerns. If there are any unsolved concerns, we will be very willing to continue the discussion.

Best Regards,

Authors

Reviewer_6e2HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary

This paper studies workflow generation when users care about several objectives and may change their preferences after search. MoFlow constructs workflows through graph edits and maintains sets of objective vectors in a shared search tree. Preference-conditioned action selection directs exploration, while a GNN estimates the value of partial workflows. An offline variant reuses execution data collected by online searches. Experiments compare six adapted workflow generators across six benchmarks and report the highest average held-out hypervolume for both MoFlow variants.

Strengths

S1. Searching once and reusing the results for different preferences is a useful formulation. The paper clearly distinguishes the cost of constructing the workflow collection from selecting a workflow for a later preference.

S2. The evaluation covers mathematics, coding, and question answering, uses disjoint build and held-out sets, and adapts the baselines to the same five objectives. The accuracy-cost plots in Appendix A.1 help interpret the aggregate hypervolume results.

S3. The paper reports the substantial execution cost of online search and explains several approximations and limitations. The construction traces also make the generated workflows easier to inspect than a final workflow alone.

Weaknesses

W1. The numerical scales used for preference optimization are not specified. Section 3.1 defines the scalarized reward as w-transpose-v, while Section 4 measures cost in tokens, latency in calls, and the other objectives on a zero-to-one scale. For the mixed preferences in the evaluation, even a modest token difference can overwhelm the entire range of accuracy. I could not find the transformations used before scalarization in Appendix D or Table 6. This matters for both MoFlow's action selection and the adapted baselines, rather than being only an implementation detail. The authors should state the exact objective transformations, their fixed bounds, and the scale of the reported hypervolumes, then show that the mixed preferences produce meaningful trade-offs under those definitions.

W2. The inference procedure is not aligned with the value represented by the optimistic backup. Equation 1 retains the best vectors across realizations, whereas inference samples a realization according to its empirical frequency (lines 295-298). An action can therefore receive the value of a rare good realization even when its expected return is worse than that of another action. Appendix C.2 explicitly acknowledges this optimism and explains its computational motivation; the remaining question is whether the returned policy serves the stated preference well. Measuring executed returns does not establish that connection. A focused comparison against directly selecting the best stored, completed workflow for each preference would help. The paper should also clarify whether Table 2 evaluates sampled tree traversals or deterministic selection from the executed front.

W3. The value model is central to the method but insufficiently specified and evaluated. Appendix D and Table 6 give its architecture, training schedule, and uncertainty threshold, but do not explain how completed-workflow observations define training targets for partial workflows, how uncertainty is calculated, or how the threshold is calibrated. This is especially consequential for offline MoFlow, whose strongest mean result relies entirely on the pretrained evaluator during search. The authors should define the prediction target, loss, and uncertainty estimator, and report prediction and uncertainty quality using the available execution data. A small ablation of the evaluator or uncertainty gate would also help separate its contribution from that of the search rule.

Critical strengths

S1. Searching once and reusing the results for different preferences is a useful formulation. The paper clearly distinguishes the cost of constructing the workflow collection from selecting a workflow for a later preference.

Critical weaknesses

W1. The numerical scales used for preference optimization are not specified. Section 3.1 defines the scalarized reward as w-transpose-v, while Section 4 measures cost in tokens, latency in calls, and the other objectives on a zero-to-one scale. For the mixed preferences in the evaluation, even a modest token difference can overwhelm the entire range of accuracy. I could not find the transformations used before scalarization in Appendix D or Table 6. This matters for both MoFlow's action selection and the adapted baselines, rather than being only an implementation detail. The authors should state the exact objective transformations, their fixed bounds, and the scale of the reported hypervolumes, then show that the mixed preferences produce meaningful trade-offs under those definitions.

W2. The inference procedure is not aligned with the value represented by the optimistic backup. Equation 1 retains the best vectors across realizations, whereas inference samples a realization according to its empirical frequency (lines 295-298). An action can therefore receive the value of a rare good realization even when its expected return is worse than that of another action. Appendix C.2 explicitly acknowledges this optimism and explains its computational motivation; the remaining question is whether the returned policy serves the stated preference well. Measuring executed returns does not establish that connection. A focused comparison against directly selecting the best stored, completed workflow for each preference would help. The paper should also clarify whether Table 2 evaluates sampled tree traversals or deterministic selection from the executed front.

Questions for authors

Q1. What objective normalization is applied before preference scalarization, and is it identical across methods and fixed independently of held-out results?

Q2. Can a previously materialized realization be selected directly at inference? If so, why is frequency-based sampling retained, and how does it affect preference-conditioned utility?

Q3. Table 6 lists seed 0, and Table 2 reports variation across benchmarks rather than uncertainty within a benchmark. Could the authors provide paired uncertainty estimates on the existing held-out problems, particularly for the close comparisons, and clarify which sources of search randomness were repeated?

6 comments

Thank you for the careful and detailed review. We appreciate your positive comments on the formulation and the evaluation, and we respond to each point below.

W1/Q1. Objective normalization and hypervolume scale

All methods use the same normalization and reference point, and MoFlow keeps the highest mean HV under every reference point and search bound we tested.

  1. Normalization and bounds. Before scalarization, each objective vector is mapped to N(v)=[acc, 1−tokens/C, 1−calls/L, rob, cons]N(v)=[\mathrm{acc},\,1-\mathrm{tokens}/C,\,1-\mathrm{calls}/L,\,\mathrm{rob},\,\mathrm{cons}], so every objective is best at 1 and the cost and latency terms reach 0 at CC tokens and LL calls per problem. MoFlow's search reward is clip(w⊤N(v),0,1)\mathrm{clip}(w^\top N(v),0,1). We set the upper bounds (C,L)(C,L) from preliminary runs. They are (20k, 8) on AIME, MBPP and HotpotQA, (70k, 10) on MATH and GPQA, and (500k, 40) on SWE-Bench. On AIME, Gemini 3 Flash uses (150k, 8) and DeepSeek uses (40k, 8). The bounds differ because cost varies by orders of magnitude across tasks and models. For example, repository-level workflows on SWE-Bench reach 305k tokens and 25 calls, and Gemini 3 Flash's long reasoning traces make even its cheapest AIME workflow cost 18k tokens (vs. 1.5k with Haiku 4.5). Every workflow that MoFlow generates for Tables 2 and 3 stays within its bound (at most 90% of CC and 78% of LL), so no reported trade-off is truncated by the bounds.

  2. HV scale. HV is computed from each method's raw held-out vectors against the reference point r0r_0 at (300k, 20), with the default (C,L)(C,L) of (20k, 8), so the maximum HV is 1×15×2.5×1×1=37.51\times15\times2.5\times1\times1=37.5.

  3. The ranking holds under different HV and search settings. Changing (C,L)(C,L) in the HV computation rescales all methods by the same factor, so only r0r_0 can change the ranking. We therefore re-ranked the eight methods in Table 2 by mean HV under three other choices of r0r_0.

r0r_0 (tokens, calls)MoFlow onlineMoFlow offlineBest baseline (ADAS)
(300k, 20), Table 22 (24.02)1 (24.56)3 (21.66)
(200k, 20)2 (15.79)1 (16.25)3 (14.20)
(80k, 15), w/o SWE-Bench2 (5.23)1 (5.33)3 (4.60)
(80k, 10), w/o SWE-Bench2 (3.36)1 (3.42)3 (2.96)

Observation. The two MoFlow variants rank first and second under every r0r_0. Running MoFlow's search with (20k, 8) instead of (70k, 10) on MATH and GPQA also leaves them in first and second place.

  1. We give two examples to show that MoFlow produces meaningful trade-offs under mixed preferences. First, as the weight shifts from accuracy to cost, MoFlow trades some accuracy for large token savings. On held-out AIME, offline MoFlow returns a workflow with 0.98 accuracy at 13.3k tokens for the accuracy-heavy preference and one with 0.85 accuracy at 5.0k tokens for the cost-heavy preference. Second, the cost weight matters even when accuracy dominates. On MBPP, the accuracy-heavy preference keeps the 0.99 accuracy of the accuracy-only workflow while using 1.5k tokens instead of 17.9k.
AuthorsOpenReview ↗

Rebuttal (Cont.)

W2/Q2. Inference rule versus the optimistic backup

  1. Does Table 2 use sampled traversals or deterministic selection? Sampled traversals. At each decision node the policy takes the ww-optimal action, and at each chance node it samples a realization according to its visit frequency pp.

  2. Can a realization be selected directly? Yes. Every realization seen during search is stored, so at each chance node inference can instead select the realization that gave the action its value in Eq. 1.

  3. Why was sampling retained? Because it is the natural algorithmic design for the MOMDP in Sec. 3.1 and preserves the stochasticity of the realizer. As in CHMCTS, the policy chooses actions while the realizer's output is a random transition, and pp estimates the probability of each realization. Sampling according to pp therefore executes the policy defined in Sec. 3.1, whose expected return is vπv^\pi, and keeps the diversity of realizations the search found. Direct selection instead performs like a greedy algorithm that hurts exploration during inference.

  4. How does sampling affect preference-conditioned utility? Per your suggestion, we compared sampling against directly selecting the best stored, completed workflow for each preference, using the returns measured when these workflows were executed during search. The table below reports the expected utility w⊤N(v)w^\top N(v) of the sampled policy and of the best stored workflow.

RunSampled policyBest stored workflow
SWE-Bench, online0.810.85
MATH, online0.910.91
AIME, Haiku 4.50.830.87
AIME, DeepSeek0.870.89
AIME, Gemini 3 Flash0.880.88
All0.850.88

Observation. Sampling serves each preference nearly as well as direct selection on the search set. For 30 of the 39 preferences the sampled policy is within 0.05 of the best stored workflow, including 12 with no loss at all, and only one preference loses more than 0.10. The gap is widest on SWE-Bench, where the alternative realizations use far more tokens.

Thank you again for the great suggestion. We will add direct selection as an alternative inference rule and include a comprehensive analysis in the appendix (e.g. including that on the held-out set) in the next revision.

AuthorsOpenReview ↗

Rebuttal (Cont.)

Q3. Within-benchmark uncertainty

We want to highlight that every entry in Table 2 is already an average over repeated runs of each workflow, and these repeated runs measure within-benchmark noise directly.

Each served workflow is run repeatedly on the same held-out problems in three ways. The first two are described in Sec. 4 (lines 357–362 and footnote 3).

  • Same query. 3 runs per problem. Accuracy, cost and latency are averaged over these runs, and the variance of accuracy across them gives consistency.
  • Paraphrased queries. One run each on the original query and 2 paraphrases. The variance of accuracy across these runs gives robustness.
  • Repeated evaluations of the same workflow. When the same workflow was served to several preferences, it was evaluated separately for each preference on the same problems. Since the preference does not affect execution, the differences between these evaluations are pure execution noise. The table below reports the resulting SD of each MoFlow entry in Table 2.

Table. SD of each MoFlow entry in Table 2 (HV) under re-evaluation.

AIMEMATHMBPPSWE-BenchGPQAHotpotQA
MoFlow online1.090.340.150.590.930.33
MoFlow offline0.980.410.240.711.060.46

These repeated runs capture execution randomness, while the search seed (0, Table 6) was not varied. We agree that additional search seeds would strengthen the comparison, but repeating Table 2 with one more seed would cost nearly USD 10,000 in API credits (the full study cost about USD 14,700, as reported in Table 6). A rerun of this size also goes beyond what the ICLR 2027 Reviewer Guidelines (https://iclr.cc/Conferences/2027/ReviewerGuidelines) describe for new experiments, which "should be limited in scope and serve to more thoroughly validate existing results from the submission." Table 2 and the table above already provide this kind of validation.

W3. Value model

MoFlow's advantage over the baselines does not rest on the value model. As described in Figure 2, every rollout runs until a terminal action completes the workflow, so partial workflows are valued only through CHVI backups from completed ones. Online MoFlow executed each completed workflow and never used fθf_\theta's predictions, yet it attains a higher mean held-out HV than every baseline (Table 2). Offline MoFlow scores each completed workflow with fθf_\theta, which is what allows it to search without executing any workflow, and the analysis below shows that it does not need a highly accurate evaluator. We also provide further details as you requested.

  1. Target and loss. Each training example is one execution draw of a completed workflow. The targets are its accuracy, robustness and consistency, and the loss is MSE. Early stopping uses a 70/30 train/validation split.
  2. Uncertainty and threshold. The uncertainty is u=max⁡(4σˉMC, 1/(1+n/16))u=\max(4\bar\sigma_{\mathrm{MC}},\,1/(1+n/16)), where σˉMC\bar\sigma_{\mathrm{MC}} is the mean standard deviation of the three predicted objectives over 8 MC-dropout passes and nn is the number of training examples. The 0.15 threshold is a default that we did not calibrate. The gate never fired in any reported run, so neither the threshold nor the uncertainty estimate affects any reported number.
  3. Quality. fθf_\theta picks workflows close to the best one, and offline search does not need a highly accurate evaluator. In offline search, fθf_\theta scores only completed workflows, so what matters is whether it picks the right workflow for each preference. The table below tests this on the validation workflows and also serves as the ablation you suggested. fθf_\theta closes 72% of the utility gap between a random pick and the best workflow, and its pick loses only 0.02 utility on average relative to the best workflow, compared with 0.30 for a random pick. Learning from execution logs also helps. fθf_\theta beats hand-written operator rules (72% vs. 55%), winning 35 of the 66 task–preference pairs. Even ranking by cost and latency alone, with no quality estimate, closes 70% of the gap, because these two objectives carry most of the selection signal. The value model therefore refines offline MoFlow's choices rather than driving them.

Table. Each quality estimate, combined with analytic cost and latency, picks its top-rated workflow for each of 11 preferences on 6 tasks. The pick is then scored by its utility w⊤N(v)w^\top N(v) measured from real executions, where best is the top validation workflow and random is the average one. Gap closed = (pick − random)/(best − random), averaged over the 66 task–preference pairs.

Quality estimateGap closedPicks the best workflow
fθf_\theta (learned)72%53%
Hand-written operator rules55%35%
Cost and latency only70%62%
Random pick0%17%

We hope this addresses your concerns, and we are happy to discuss further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.