acceptodds
Under review as a conference paper at ICLR 2027

Threshold-Guided Composition of GFlowNets for Multi-Objective Generation

Abstract

Generating candidates that satisfy multiple property requirements is a central challenge in multi-objective generation. Optimizing an aggregate reward can improve some properties while leaving others below their required levels. We introduce GFN, which composes independently trained GFlowNets through threshold-guided sampling. Each objective model provides a flow ratio, , that estimates the expected terminal reward under its own policy. The sampler uses these estimates to penalize objective-wise threshold shortfalls, with no penalty once an estimate meets its threshold. This focuses guidance on predicted deficits while allowing the same models to serve different objective subsets and thresholds without retraining. Hypergrid diagnostics illustrate how objective-specific thresholds redirect sampling. Molecular experiments on three tasks with two to four objectives show competitive performance on the two-objective task and higher mean yields of distinct feasible molecules than preference-conditioned baselines on the three-objective task. On the four-objective task, achieves the highest mean objective-space coverage and top-candidate quality in the main comparison, although jointly feasible molecules remain rare. Ablations show that guidance improves mean unique success over unguided flow composition.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Generative Flow Networks, Multi-Objective Optimization, Threshold-Guided Composition, Flow Moment Extraction

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_5WGrOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Satisfying objective thresholds in multi-objective sampling is a difficult problem since an acceptable or feasible object is only achieved when all objectives pass their performance thresholds. In order to improve the diverse sampling of feasible objects in these multi-objective setting, this paper proposes a post-hoc sampling method whereby multiple GFlowNets (trained with different rewards) are composed in such a way to avoid sampling below-threshold actions. Threshold violations are evaluated in a clever way: by using a property of GFlowNets with squared and unsquared rewards! This is an interesting solution to the problem of thresholded sampling because it allows one to change the thresholds without re-training a model, thereby allowing for more or less greedy sampling with just two GFlowNets. The experimental results in the hypergrid environment are quite interesting, demonstrating that threshold-guided sampling with GFlowNets is the only work capable of generating samples within a threshold area. In a scientific discovery task with a variable number of objectives, the proposed method often outperforms baselines, albeit by a small margin with a smaller number of objectives.

Critical strengths

The paper motivates the problem well, making clear the desire to sample according to multiple objectives. The empirical results are especially interesting as they draw a sharp distinction between the proposed method and prior work.

Critical weaknesses

First, I feel that while the problem is well-motivated and somewhat placed into the context of prior work in the introduction, the writing in Section 2 made it unfortunately extraordinarily difficult to understand the actual substance of the work. After much reading, I finally understood the method to essentially be sampling with penalisations on probabilty mass according to the collective below-thresholdness of each action. A cleaner presentation would have substantially aided this understanding, especially if the related works section went through prior approaches and how they do a similar thing, and how your method differs. The writing alone is a major factor in my decision to reject the paper.

Second, I am not convinced that sections 2.2 and 2.3 are actually presenting anything new. Section 2.2: it is fine to show the training loss and state that you're training a neural network with two heads, but the section feels almost torturous with unnecessary definitions of terms that never re-appear. The only thing that Section 2.2 needed to say was that the backward policy was the same between the two networks. This sets up the result of Section 2.3, which is already proven in the GFlowNet foundation paper in Appendix D (or at least something very similar). This citation is never produced, nor is it ever stated that this is an already known property of squared reward GFlowNets against unsquared reward GFlowNets. In addition, the proof of this proposition need much more detail, in addition to the citation to the prior work.

EDIT: After rebuttal, I think the manuscrit has improved somewhat, but still not enough for me to recommend acceptance.

Questions for authors
  • I wonder whether the increased number of networks improves performance, and whether a larger single GFlowNet, or larger networks in MO-GFN would produce similar results, given that you seem to need substantially more parameters for a given task?
  • From Figure 2 in the main body and Figure 6 in the appendix it seems that the mu-GFN with theta=0.6 would best-recover the optimal sampling region of the mu-GFN with theta=0.4. Why wouldn’t it be optimal in general to deploy with more restrictive-than-necessary thresholds when they seem to be better at fitting into the space?
  • From Section 3.2 and 3.3, it seems that μ\muGFN excels really when there are more objectives. Please provide some intuitions for why this may be the case, and how MO-GFN, for instance, is unable to scale to the same extent.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 5WGr : Weaknesses and Q1

Thank you for the detailed feedback. We agree that the presentation made the sampling mechanism harder to understand than necessary and did not sufficiently distinguish the guidance contribution from the underlying flow-moment construction. We address these points first, followed by the three questions.

Weaknesses: Presentation and relationship to prior work

The main idea is to compose objective-specific flows and reduce the probability of actions whose estimated continuation rewards fall below requested cutoffs. We will reorganize Section 2 around this sampling rule, define its quantities before use, and move the detailed training construction and proof to the appendix. The main text will retain the shared-backward-policy condition and explain precisely which policy defines each reward estimate.

We also agree that reward-moment extraction has prior foundations. Appendix D of GFlowNet Foundations discusses downstream reward estimation using squared-reward and reward flows. We will explicitly credit this connection and distinguish its conditioning from continuation under an objective's forward policy. We will not present the general flow-moment principle as a new contribution.

Our contribution is the use of objective-wise continuation estimates in a root-normalized composition with capped relative-shortfall penalties, allowing requests to change at inference time. The cap removes an objective's penalty once its estimate reaches its cutoff, while an uncapped log-shortfall rule loses its cutoff dependence after action normalization. This distinction establishes a mechanism for responding to requests, not a guarantee of higher feasible yield or exact sampling from the feasible set.

Q1 Parameter-matched baselines

We trained a larger MOGFN-PC model to match the total trainable parameter count of the two MuDBPB models used for JNK3 + GSK3β. Each MuDBPB model has 1,525,053 parameters, so the deployed two-objective system contains 3,050,106 trainable parameters in total. MOGFN-PC with embedding width 416 has 3,041,066 parameters, a difference of approximately 0.30%.

Table 5WGr-1. Total-parameter comparison on JNK3 + GSK3β. All rows use five seeds and 5,000 generated molecules per seed. Training uses 10,000 iterations and batch size 16 per model. Values are means ± across-seed SD. SR_unique counts distinct feasible molecules divided by all 5,000 draws. The μGFN row is reproduced from the main manuscript table; width 416 is the additional experiment.

MethodTotal trainable parametersSR_unique ↑HV ↑Geom ↑
MOGFN-PC, width 2561,216,1060.0050 ± 0.00190.6747 ± 0.03940.5182 ± 0.0175
MOGFN-PC, width 4163,041,0660.0045 ± 0.00280.6182 ± 0.01490.5096 ± 0.0306
μGFN DBPB, two models3,050,1060.0055 ± 0.00090.6707 ± 0.00570.5274 ± 0.0057

Widening MOGFN-PC did not improve mean unique success under this training protocol. We do not interpret the small difference between μGFN and the width-416 baseline as a demonstrated statistically significant advantage, or conclude that capacity is irrelevant. Retuning or longer training of the larger baseline could change the result.

The parameter comparison also distinguishes training from inference. μGFN's backward-policy heads are trained but not evaluated during sampling, leaving 2,601,896 active inference parameters across the two models. An additional width-384 baseline matches this quantity with 2,610,538 parameters. We report it as a supplementary check rather than rely on its lower performance for the capacity argument.

Table 5WGr-2. Supplementary inference-parameter comparison. The full five-seed row retains the unstable run. Its exclusion is shown only as a sensitivity analysis.

MOGFN-PC configurationSeedsSR_unique ↑HV ↑Geom ↑
Width 384, all runs50.0018 ± 0.00120.5003 ± 0.27400.4030 ± 0.1822
Width 384, unstable seed excluded40.0022 ± 0.00080.6214 ± 0.04800.4836 ± 0.0321

Parameter matching does not imply matching every resource. For the molecular training protocol, each objective receives 160,000 score evaluations in either approach, while μGFN trains a separate network for each objective. We will distinguish oracle evaluations, total optimization steps, parameter counts, and model reuse rather than describe all resources as equal.

AuthorsOpenReview ↗

Response to Reviewer 5WGr : Q2 and Q3

Q2 Stricter guidance cutoffs

Thank you for raising this point. In this grid, the feasible set is empty at cutoff 0.6, so the corresponding panel cannot show samples that meet the requested constraints. We mentioned this in the figure caption and Section 3.1, but we should have explained its implications more clearly in the main text. We will clarify this when discussing the threshold progression.

Table 5WGr-3. Exact feasible-set sizes on the 16 × 16 grid. Both objectives use the same cutoff.

Evaluation cutoff0.30.40.50.6
Number of jointly feasible terminal states471730

This does not rule out the reviewer's suggestion that stricter guidance could improve sampling at a lower, fixed evaluation cutoff. Our molecular sensitivity study addresses that separate question.

Table 5WGr-4. Guidance sensitivity on JNK3 + GSK3β. Both evaluation cutoffs remain 0.5 and λ remains 1. The same five checkpoints are reused, with 5,000 draws per seed and requested cutoff. Entries are means ± SD.

Requested score cutoffSR_unique ↑
0.50.0055 ± 0.0009
0.60.0045 ± 0.0005
0.70.0046 ± 0.0008
0.80.0050 ± 0.0009
0.90.0052 ± 0.0013

The observed means are not monotonic. They do not establish significance for individual differences, but do not support a general prescription to use more restrictive cutoffs. Raising a request can change the active shortfall penalties and the concentration of the sampling distribution.

In the main molecular comparison, requested score cutoffs equal evaluation cutoffs, with 0.5 for JNK3, GSK3β, and SEH, 0.6 for QED, and 0.7 for SA. Each is converted to reward units separately. This alignment does not apply to sensitivity experiments such as Table 5WGr-4. We will also correct the statement that λ was fixed before the experiments. The task-specific strengths were selected through exploratory evaluation, and the revised protocol will report the explored settings and sensitivity results rather than describe the selection rule as independently validated.

Finally, the illustrated grid sweep uses the linear gate. An empty feasible set does not imply that every action estimate lies below every cutoff, so we will not attribute the 0.6 panel to complete cutoff cancellation without checking that condition. The full log-shortfall sampler has now been evaluated separately, as reported in our response to PnJS.

Q3 Behavior as the number of objectives increases

The mechanism motivating μGFN is that an aggregate reward can improve while an individual requirement remains unmet. Separate shortfall penalties retain a signal for each predicted deficit, while capping stops penalizing an objective whose estimate meets its cutoff. This can help coordinate multiple requirements, but it is not a proof of an increasing advantage with objective count.

The four-objective main result still corresponds to approximately three distinct feasible molecules per 5,000 draws. Moreover, uncapped guidance has higher mean unique success on that task, 0.0011 versus 0.0006. We will retain these qualifications. The separately established reuse benefit is that the trained objective models can be recombined across subsets without further objective-model training.

Thank you for addressing these weaknesses and answering my questions. After these responses, I am a little more satisfied that your method is effective and interesting, at least experimentally. It does seem that this composition of different GFlowNets is a useful technique. That said, having read over the revised methods section (Section 2), I cannot say that the writing or understandability has improved significantly. Thus, while I will increase my score, I still cannot fully support publication.

When updating your manuscript, please highlight changes with red text or a yellow or orange highlight, this makes it a lot easier to see what you've changed. Please try to improve the writing and motivation, and also strongly consider including a background section after the introduction. There are a lot of missing steps that would have greatly improved readability.

Thank you for the follow-up and for reconsidering your score.

We need to apologize for a confusion on our side: at the time of your comment, the revised PDF had not yet been uploaded, so the Section 2 you read was the original submission. Our earlier response described the changes but did not make this clear. The revised manuscript is now available, with all changes marked in red.

We would be grateful if you could take another look at the revised version. If any specific step remains unclear, we will revise it further within the discussion period.

Reviewer_FUVKOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses the problem of multi-objective guided generation using an ensemble of GFlowNets (GFN), a special family of generative models that generate objects by carrying out a sequence of actions. The proposed method μ\muGFN trains separate GFNs that learn to sample objects proportional to the utility function Rk(x)R_k(x) of each objective separately, i.e. pk(x)∝Rk(x)p_k(x) \propto R_k(x). Sampling from pk(x)p_k(x) yields high fidelity samples for objective kk with high probability. The aim is to combine the models to sample objects that meet a minimum requirement under all objectives simultaneously, i.e. objects satisfying ck(x)≥τkc_k(x) \ge \tau_k for every selected objective.

μ\muGFN combines the flows learned by the individual GFNs through a root-normalized sum, B(s,a)=∑kFk(1)(s,a)/ZkB(s,a) = \sum_k F^{(1)}_k(s,a)/Z_k. Using a dual flow architecture that models both Rk(x)R_k(x) and Rk2(x)R_k^2(x) in parallel, using a shared backward probability to tie both together, allows to easily estimate the expected utility μ\mu of performing an action aa in a given state ss when following the action policy of the GFN. μ\muGFN uses this expected utility to down-weight the sampling probability of actions that are linked to low expected utility under individual policies, multiplying each action score by min⁡(1,(μ^k/θk)λ)\min(1, (\hat\mu_k/\theta_k)^\lambda) per objective. The idea is that this naturally guides object generation into regions which have high utility under all objectives.

Critical strengths
  • The paper addresses a highly important topic. Multi-objective generation is important for many scientific applications, like drug design, where a candidate must simultaneously meet different requirements.
  • μ\muGFN can be applied for different multi-objective generation configurations without re-training.
  • The paper introduces the dual flow architecture that allows to easily calculate the expected terminal reward when following the policy. This was interesting to me although it seems it is a known trick. However, the paper reference seems to be broken and needs to be fixed.
  • The paper honestly reports negative results, e.g. capping helps on only for two of three tasks in the experimental section.
Critical weaknesses

Conceptual Problems

1. PDF under base composition

My major concern is about the base composition rule, which is at the very core of μ\muGFN. Following Eq. 1, the GFN seems to model P0(x)=1K∑kRk(x)Zk=1K∑kpk(x)P_0(x) = \frac{1}{K} \sum_{k} \frac{R_k(x)}{Z_k} = \frac{1}{K} \sum_{k} p_k(x) under base composition. Intuitively, sampling from this distribution is equivalent to: 1. pick one of the k single objective GFN randomly and 2. run the forward policy to sample from it. But, for each of the single objective GFNs, the probability mass sits near its own optimum, i.e. we sample the extremes of the Pareto front. More specifically, the chosen flow combination results in an OR like combination of the individual PDFs.

I think this is not a desirable choice, because what multi-objective generation aims for is to put the mass in the region where we have compromizes between the objectives. The hypergrid example seems to support that μ\muGFN suffers from such a coverage problem, because mass is concentrated and is not distributed over the feasible region.

MO-GFN uses a much more suited AND-like combination ∏Rk\prod R_k, which seems to be a more natural base.

Experiments that visualize or measure coverage and fidelity are missing. Convincing experiments which show that the proposed threshold guided combination actually helps to put mass in the compromize region are missing.

2. PDF under reward based guidance

Related to the previous problem is the fact, that the distribution under threshold guidance is unknown. This takes away parts of the beauty (controllable, interpretable) of GFNs on the one side, but also falls short of the raised expectations that we want to sample under hard performance guarantees ck(x)≤τkc_k(x) \leq \tau_k.

3. Policy miss-match for μ^\hat{\mu}

μ^=EPF,k(1)[Rk(X)∣s,a]\hat{\mu} = E_{P_{F,k}^{(1)}}[R_{k}(X)|s,a], the expected terminal reward when following the policy PF,k(1)P_{F,k}^{(1)} is used to calculate the shortfall penalty. However, the μ\muGFN does not follow PF,k(1)P_{F,k}^{(1)}. Hence, it is not obvious why this choice is meaningful. An explanation would strengthen the paper.

Minor Presentation Issues

  1. line 117: notation Fk(1)(s,a)F_k^{(1)}(s,a) is defined only later in Eq. 5. A short explanation of the notation would increase readability.
  2. line 120: defines estimate μ^\hat{\mu} with same Flow as Eq.7, which defines true μ\mu. This is not consistent.
  3. line 132: dk(s)=pk(s,a)pk(a∣s)=pk(s)d_k(s) = \frac{p_k(s,a)}{p_k(a|s)} = p_k(s). I would argue pk(s)p_k(s) is much more intuitive to read.
  4. line 236: reference missing
  5. Eq. 1: Meaning of subscript P0(a∣s)P_0(a|s) only understandable after line 159. Explanation or re-ordering would improve readability.

Minor Improvements upon SOTA

The reported improvements compared to SOTA seem to be very small.

Missing Related Work Checking the literature reveals Order-Preserving (OP) GFlowNets as another competitive GFlowNet based approach for MOO (https://proceedings.iclr.cc/paper_files/paper/2024/file/ea06e6e9e80f1c3d382317fff67041ac-Paper-Conference.pdf). Citation and comparisons are missing.

Summary

  • Defining the sampling problem with hard performance guarantees (ck(x)≤τkc_k(x) \leq \tau_k), the paper raises high expactations. In the end, the proposed method μ\muGFN cannot meet these expactations. Performance guarantees are not fulfilled per se and the sampling distribution of the model is unknown. This is explicitly mentioned in the paper.
  • While certain design choices feel intuitive initially (choice of base composition), they are neither universal nor self-evident. Because they lack explicit rationale and empirical validation through experiments, we need stronger documentation and comparative data to support these decisions. Alternative choices are not adequately discussed.
  • The proposed method has at a methodological inconsistency (Policy miss-match for μ^\hat{\mu}).
Questions for authors

Base composition (Problem 1)

  1. Is my reading of Eq. 1 correct, that under exact flows and before stop masking the unguided base targets P0(x)=1K∑kpk(x)P_0(x) = \frac{1}{K}\sum_k p_k(x), an equal-weight mixture of the single-objective distributions? If so, could this be stated explicitly in Sec. 2.1?

  2. A mixture places mass near each objective’s own optimum, whereas joint feasibility requires the compromise region. Why is a sum the right base for a conjunctive requirement, rather than a product or harmonic mean?

  3. n Fig. 2 at θ=0.5\theta = 0.5, samples appear concentrated on a few cells rather than spread across the feasible region. How much of the feasible region receives non-negligible terminal mass?

  4. Could you provide additional experiments measuring fidelity and coverage of the generated distribution? Specifically, for the molecular tasks: 1) IGD+ (Pareto-compliant, unlike IGD) against a reference front. 2) Pareto count: the number of non-dominated molecules among the deduplicated samples. 3) PC-entropy or a comparable uniformity measure to show whether coverage is spread across the front or concentrated on part of it.

Guidance signal (Problem 3)

  1. μ^k\hat\mu_k is the expected terminal reward under PF,k(1)P_{F,k}^{(1)}, but the sampler follows neither PF,k(1)P_{F,k}^{(1)} nor any single objective’s policy. Can you give an argument, or empirical evidence, for why this estimate remains informative under the composed sampler?
8 comments
AuthorsOpenReview ↗

Response to Reviewer FUVK: Conceptual Problem 1(Q1,Q2,Q3)

Thank you for identifying the missing interpretations and comparisons. We address the conceptual weaknesses together with the corresponding questions below, and separately discuss the magnitude of the reported improvements and the missing related work.

Q1 and Conceptual Problem 1 Terminal distribution of the base composition

Yes. Under exact flow conservation, equal objective weights, and before additional action masking, the normalized sum of flows has terminal distribution

P0(x)=1K∑k=1KRk(x)ZkP_0(x)=\frac{1}{K}\sum_{k=1}^K\frac{R_k(x)}{Z_k}

This is the equal-weight mixture of the individual terminal distributions. We will state it explicitly in Section 2.1. The base does not encode conjunction by itself; the shortfall penalty is the component responding to objective-wise requirements. Learned flow errors and additional stop masking can invalidate the exact terminal-mixture identity, and we will place these conditions beside the statement.

Q2 and Conceptual Problem 1 Sum versus product or harmonic composition

A sum is not uniquely suited to joint feasibility. We evaluated alternative local pooling rules using the same grid objective models. The harmonic base gives a better operating point in the tested comparison.

Table FUVK-1. Alternative local bases at cutoff 0.5. These are mean exact FR values over ten DB checkpoints, using the log-shortfall penalty and T=1. A sum with λ=0 is the root-normalized unguided base. For all three guided rows, expected coverage in 1,000 draws rounds to 1.00.

Local baseGuidance strengthFR ↑
Sumλ=00.0150
Sumλ=10.0434
Harmonicλ=00.0223
Harmonicλ=10.0586
Geometricλ=10.0508

Guidance improves the sum base from 0.0150 to 0.0434 and the harmonic base from 0.0223 to 0.0586. Thus the base operator and shortfall guidance are separate design choices, and we do not claim that the sum is optimal.

We also distinguish local edge-flow pooling from pooling terminal distributions. The exact normalized product of the two terminal distributions has feasible mass 0.0263 at cutoff 0.5. A simple product of local flow values does not generally sample from that terminal product. This does not rule out product-based composition methods; it limits what can be inferred from directly multiplying edge scores. We will retain the sum for its simple exact-mixture interpretation and report the harmonic alternative.

Q3 and Conceptual Problem 1 Coverage of the feasible region

At cutoff 0.5, the grid has three jointly feasible terminal states. We compute coverage from the exact terminal probabilities rather than estimate it from heatmap intensity. Specifically, C₁₀₀₀ is the expected fraction of these states observed in 1,000 draws.

Table FUVK-2. Coverage and feasible mass at cutoff 0.5. Results use ten DB checkpoints. FR is mean ± SD across checkpoints and coverage is rounded.

SamplerFR ↑Expected feasible coverage ↑
Original linear gate, λ=10, T=0.50.2625 ± 0.00600.71
Full log-shortfall, λ=1, T=10.0434 ± 0.00021.00
Full log-shortfall, λ=10, T=0.50.7159 ± 0.00240.94

These results distinguish concentration, feasible mass, and coverage. A rounded coverage of 1.00 is not a guarantee for every finite sample or evidence of a uniform distribution. The full sampler is also evaluated at cutoffs 0.3 and 0.4, where the feasible sets contain 47 and 17 states, respectively. We will report these alongside the small three-state case. We will also disclose that the grid μGFN models use exhaustive full-grid training, whereas the Scalarized and ordering baselines use sampled trajectories.

AuthorsOpenReview ↗

Response to Reviewer FUVK: Conceptual Problem 1(Q4), Conceptual Problem 2

Q4 and Conceptual Problem 1 Molecular fidelity and coverage metrics

We computed IGD+, nondominated counts after SMILES deduplication, and objective-space occupancy entropy. The reference front for IGD+ is the nondominated set of pooled objective vectors from the evaluated methods and five seeds per method. It is an empirical reference, not the unknown true molecular Pareto front. For maximization, the modified distance is the norm of the positive part of z − a, minimized over generated vectors a and averaged over reference points z.

Table FUVK-3. JNK3 + GSK3β distributional diagnostics. Every row uses five seeds and 5,000 draws per seed in the supplementary sample collection. Values are means ± across-seed SD. Raw Pareto counts use each seed's deduplicated set. Matched counts average 20 subsamples of 262 distinct molecules per seed. H_feas is normalized occupancy entropy within the feasible objective-space box. This analysis is reported separately from the main-table yield results.

MethodIGD+ ↓Pareto count, rawPareto count, matchedH_feas
Raw Flow Sum0.1383 ± 0.06714.6 ± 1.13.4 ± 0.20.327 ± 0.105
Root normalization only0.0896 ± 0.05483.4 ± 1.32.9 ± 0.50.496 ± 0.088
Uncapped, λ=10.0482 ± 0.02763.0 ± 1.73.4 ± 0.50.809 ± 0.025
Capped, λ=10.0459 ± 0.01003.0 ± 0.73.5 ± 0.40.797 ± 0.016
MOGFN-PC0.0365 ± 0.04104.0 ± 2.33.7 ± 0.60.772 ± 0.025

The feasible box is partitioned into ten fixed intervals per axis, and entropy is normalized by the logarithm of the number of feasible cells occupied in the pooled reference collection. Here the pool occupies 24 feasible cells. This is a complementary occupancy measure, not PC-entropy, a chemical-diversity measure, or proof of uniform feasible coverage. Matched and raw Pareto counts need not be ordered, because removing a dominating point can reveal additional nondominated points in a subset.

Capped guidance improves mean IGD+ and feasible occupancy entropy relative to the two unguided composition arms. MOGFN-PC has lower mean IGD+ than capped guidance, with substantial seed variability, and uncapped guidance has slightly higher mean feasible entropy. We will report these outcomes without claiming uniform superiority.

The same qualification applies across tasks. On SEH + QED + SA, MOGFN-PC has mean IGD+ 0.0159, compared with 0.0210 for capped guidance at the main λ=8 setting, and feasible entropy 0.833 versus 0.824. On the four-objective task, uncapped λ=1 has mean IGD+ 0.0319 versus 0.0351 for capped guidance. Only 11 distinct feasible objective vectors occur in the four-objective pooled collection, making its feasible-entropy estimates particularly fragile.

We will distinguish all entropy metrics from feasible yield. In particular, the feasible fraction among distinct molecules divides the number of distinct feasible molecules by the number of distinct molecules, whereas SR_unique divides by all generated draws. These quantities cannot be interchanged.

Conceptual Problem 2 Guided distribution and feasibility guarantees

We agree that the paper needs to distinguish requested constraints from guarantees about generated samples. The method provides soft guidance toward those constraints. It does not establish exact sampling from a feasible-conditioned distribution or guarantee that each generated candidate is feasible.

The guided transition probabilities define a terminal distribution, which we compute exactly by dynamic programming on the enumerable grid. What is missing is a general identity expressing that distribution as a specified terminal reward distribution or as conditioning on joint feasibility. We will make this distinction explicit and remove wording that suggests hard performance guarantees. The empirical feasible mass and coverage results below assess the sampler's behavior without replacing that missing guarantee.

AuthorsOpenReview ↗

Response to Reviewer FUVK: Conceptual Problem 3(Q5) and Minor improvements over SOTA

Q5 and Conceptual Problem 3 Informativeness under the composed policy

The reviewer is correct about the policy mismatch. Under the ideal flow conditions, μₖ is the expected continuation reward under objective k's original forward policy. We use its learned estimate as a local reference-policy score, not as the value function of the deployed sampler. The latter's terminal distribution is induced by its transition probabilities, but we do not have a general identity equating it to a known feasible-conditioned target distribution.

We evaluated the discrepancy directly on Hypergrid. Dynamic programming computes the guided policy's own continuation rewards, allowing them to be compared with the objective-wise estimates on visited state-action pairs.

Table FUVK-4. Policy-mismatch diagnostics for the full log-shortfall sampler. DB checkpoints, ten seeds, λ=1 and T=1. Absolute error and threshold disagreement use visitation weighting. Action-order agreement compares within-state action orderings. Error bars are checkpoint SD where reported.

Shared cutoffAbsolute differenceThreshold-classification disagreementAction-order agreement
0.40.0228 ± 0.00030.037 ± 0.0010.986
0.50.0203 ± 0.00020.036 ± 0.0020.985

These diagnostics provide evidence that the reference-policy scores retain useful ordering and threshold information at the tested operating points. They do not eliminate the discrepancy. Stronger guidance can increase it substantially; at λ=10, T=0.5 and cutoff 0.5, the threshold disagreement is approximately 0.498. Moreover, the penalty depends on magnitudes, not only ranks or threshold crossings. We will present these as operating-point-specific grid diagnostics and not generalize them into a molecular calibration guarantee.

Minor improvements over SOTA

We agree that the current results do not show a uniform or large improvement over the strongest baseline. In the main molecular table, mean SR_unique is 0.0055 versus 0.0050 for MOGFN-PC on JNK3 + GSK3β, and 0.0655 versus 0.0594 on SEH + QED + SA. MOGFN-PC has the higher mean HV on both tasks. On the four-objective task, the means are 0.0006 versus 0.0001, but μGFN's absolute yield is still only about three distinct feasible molecules per 5,000 draws. These descriptive comparisons do not establish statistically significant superiority.

We will frame the contribution around request-dependent sampling and reuse of objective models, supported by ablations of normalization and guidance, rather than broad SOTA dominance. The stronger harmonic-base result, the uncapped four-objective result, and the threshold-ordering baseline also delimit where the current design is preferable. We agree that reuse and controllability must be evaluated alongside yield and coverage, not used to dismiss a stronger baseline.

Reviewer_PnJSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes µGFN, a multi-objective method that combines pretrained GFlowNets to generate objects that belong to a feasible region defined by thresholds for each reward. For each reward, the method learns two flows, with terminal rewards RkR_k and Rk2R_k^2, sharing the same backward policy. At the optimum, the ratio between these flows is equal to the expected terminal reward under the policy learned by the first flow. The authors construct a base sampling policy by combining the normalized flows learned for each reward and modify it using a log-shortfall penalty to discourage actions whose estimated expected rewards fall below the desired thresholds, which can be chosen at inference time. The method is evaluated on Hypergrid, a classical GFlowNet benchmark, and on three molecular generation tasks. The authors report improvements over unguided flow composition and competitive results compared to preference-conditioned multi-objective GFlowNets.

Critical strengths
  1. Flow ratio as an expected reward

The relationship between the flow learned for a reward and the flow learned for its squared reward is an interesting idea, and its interpretation as a signal to guide the policy could be very useful for multi-objective GFN.

  1. Inference-time thresholds.

The ability to define new feasible regions by changing the thresholds at inference time, without having to retrain each GFN, is especially useful when we want to test different requirements for different rewards.

Critical weaknesses
  1. Flow-ratio interpretation.

When the paper first shows the relationship between the ratio of the flows and the expected reward under the policy learned by the first flow, the interpretation is very clear. However, in practice, the algorithm does not use this policy. Instead, it uses a penalized version of the base policy. As a result, the ratio no longer represents the expected reward under the policy actually being used, and there is no guarantee that this guidance will correctly direct samples toward the feasible region. Although the authors acknowledge this issue, I still consider it a significant weakness of the method.

  1. Coverage of the feasible region.

The thresholds define a feasible region, but it is not clear whether the samples would cover this entire region or provide a minimum level of diversity. Figure 1 itself seems to show a bias, with samples concentrated in only part of the feasible region, roughly half of it. This is particularly concerning since the experiment is conducted in a relatively simple environment.

  1. Learning two different flows for each GFN.

Although the method provides a simple way to combine flows at inference time, the idea of having already learned an additional flow for the squared reward seems somewhat unusual, and not something someone would naturally have trained. For the method to be practical, one would need to have learned this additional flow for Rk2R_k^2. Therefore, it is not simply a matter of combining GFNs that have already been trained for the rewards of interest. Although I do not consider this a major limitation of the paper, it does make the proposed method somewhat less practical.

Questions for authors
  1. Direct feasibility baseline.

Have you compared your method against a GFlowNet trained with a joint feasibility indicator reward, or some smooth approximation of it? Of course, a pure indicator would introduce zero-reward regions, which can be problematic for standard log-based GFlowNet objectives, but something like R(x)=ε+∏k1{ck(x)≥τk}R(x) = \varepsilon + \prod_k \mathbf{1}\{c_k(x) \ge \tau_k\}. Have you tried anything along these lines? I understand this might not be an entirely fair comparison, since your method operates at inference time, while the alternative would need to be trained from scratch.

  1. Trajectory Balance.

Nowadays, Trajectory Balance is one of the most widely used objectives for training GFlowNets, and several papers have shown advantages over other training losses. Do you see a way to extend your method to GFlowNets trained using Trajectory Balance?

  1. Consistency of the experiments.

I noticed that in the Hypergrid experiments, you use a linear guidance rule instead of the main log-shortfall sampler. Is there a particular reason for this? Could you also evaluate the main sampler on Hypergrid?

4 comments
AuthorsOpenReview ↗

Response to Reviewer PnJS : Q1, Q2

Thank you for the concrete suggestions. We have evaluated the direct-feasibility baseline and the full log-shortfall rule on Hypergrid. We also clarify the requirements for a TB extension and the interpretation of the guidance signal.

Q1 Direct feasibility reward

We trained a single GFlowNet with flow matching and the reward

Rfeas(x)=0.01+1{cJNK3(x)≥0.5, cGSK3β(x)≥0.5}.R_{\mathrm{feas}}(x)=0.01+\mathbf{1}\{c_{\mathrm{JNK3}}(x)\geq0.5,\ c_{\mathrm{GSK3\beta}}(x)\geq0.5\}.

Training used 10,000 iterations, batch size 16, exploration probability 0.05, and five seeds. The final evaluation generated 5,000 molecules per seed. No jointly feasible molecule was observed across these 25,000 samples.

Table PnJS-1. Direct-feasibility experiment on JNK3 + GSK3β. FeasFM is the additional experiment. The other rows reproduce the five-seed main manuscript results for reference. All rows use 5,000 draws per seed; entries are means ± SD. HV uses the origin and deduplicated molecules. Geom is the mean of the top 100 molecule-wise geometric means after deduplication.

MethodSR_unique ↑HV ↑Geom ↑
FeasFM, positive indicator reward0.0000 ± 0.00000.3038 ± 0.03280.2670 ± 0.0052
Random, main-table reference0.00000.2075 ± 0.06840.2315 ± 0.0045
μGFN DBPB, main-table reference0.0055 ± 0.00090.6707 ± 0.00570.5274 ± 0.0057

FeasFM improves the auxiliary metrics over Random, so we do not describe it as learning nothing. The result is limited to the tested reward and training protocol. We have not evaluated a smooth feasibility surrogate and do not conclude that direct feasibility training generally fails. Indeed, the additional threshold-ordering grid baseline is highly effective, as reported in our response to FUVK.

Q2 Extension to Trajectory Balance

A TB extension is possible in principle. The issue is the direct readout from two raw logit heads, rather than TB's ability to learn reward-proportional terminal distributions. Trajectory Balance provides a trajectory-level training objective; applying it to normalized policies does not, by itself, fix the absolute scale of the underlying state-dependent logits.

For a fixed state, adding a constant to every action logit of a head leaves its normalized policy unchanged. If the two heads receive shifts c₁(s) and c₂(s), their exponentiated logit difference is multiplied by exp(c₂(s) − c₁(s)). This positive factor preserves action rankings within that state and objective but can change whether the readout exceeds an absolute reward cutoff.

Table PnJS-2. What is preserved by the raw-logit scale freedom.

QuantityEffect of head-specific state-dependent shifts
Each normalized forward policyUnchanged
Ranking of the ratio readout within one state and objectiveUnchanged
Ratio readout in absolute reward unitsRescaled
Comparison with a fixed reward cutoffCan change
Ranking after combining differently scaled objectivesNot generally preserved

This ranking invariance does not establish that the learned rankings are accurate. It explains why relative ordering and absolute cutoff guidance require different information. A TB-based implementation could instead evaluate continuation rewards under its policy or recover the corresponding unnormalized flows. Exact policy evaluation is available on a small enumerable grid; larger state spaces require an appropriate estimation or flow-recovery procedure. We will describe this as a limitation of the current one-forward-pass readout, not an impossibility result for TB.

AuthorsOpenReview ↗

Response to Reviewer PnJS : Q3

Q3 The full log-shortfall sampler on Hypergrid

We agree that the linear grid diagnostic does not directly evaluate the main sampling rule. We have now evaluated the root-normalized base plus logarithmic capped-shortfall rule on the same grid checkpoints. Forward dynamic programming computes the terminal distribution without Monte Carlo sampling.

Let F be the feasible terminal set and p(x) the terminal probability. We report feasible mass FR = Σₓ∈F p(x) and expected coverage in N draws

CN=1∣F∣∑x∈F[1−(1−p(x))N].C_N=\frac{1}{|F|}\sum_{x\in F}\left[1-(1-p(x))^N\right].

Table PnJS-3. Full and linear samplers on DB grid checkpoints. Each cell is FR ± checkpoint SD followed by C₁₀₀₀. Results use ten checkpoints and exact DP; coverage is rounded. Both guidance and evaluation use the indicated shared cutoff.

Sampling rule and settingsCutoff 0.3Cutoff 0.4Cutoff 0.5
Full log-shortfall, λ=1, T=10.3054 ± 0.0018 / 1.000.1679 ± 0.0011 / 1.000.0434 ± 0.0002 / 1.00
Full log-shortfall, λ=10, T=0.50.9902 ± 0.0002 / 0.930.9783 ± 0.0003 / 0.940.7159 ± 0.0024 / 0.94
Linear gate, λ=1, T=10.1009 ± 0.0002 / 0.510.0349 ± 0.0001 / 0.600.0058 ± 0.0000 / 0.71
Linear gate, λ=10, T=0.50.6374 ± 0.0002 / 0.790.5335 ± 0.0003 / 0.880.2625 ± 0.0060 / 0.71

The full sampler provides different yield–coverage operating points. We show both settings rather than compare only each rule's original setting, since linear and logarithmic penalties have different scales. The full sampler at λ=1, T=1 also gives FR 0.2407 ± 0.0015, 0.1017 ± 0.0011, and 0.0220 ± 0.0005 with FM checkpoints, showing that results depend on the learned estimates as well as the sampling rule.

I thank the authors for the detailed responses and additional experiments. Therefore, I have decided to increase my rating.

Thank you for the follow-up and for reconsidering your assessment. We’re glad we could address your concerns, and we appreciate your thoughtful feedback and the time you took to review our work.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your reviews and contributions. As we are now halfway through the discussion period, I would like to encourage you to carefully read the authors' responses and engage in further discussion where appropriate. Please also indicate whether the rebuttal has addressed your concerns or changed your assessment.

Thank you for your time and effort!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.