acceptodds
Withdrawn from ICLR 2027

Capacity-Aware Sampling for Frozen Generative Flow Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

GFlowNets are trained so that the terminal distribution is proportional to reward, yet scientific discovery deployments draw only a finite pool of samples from a frozen GFlowNet. We show that reward-proportional i.i.d. deployment is generally misaligned with finite-budget coverage and cannot adapt local flow to capacity as the sample pool grows. We first derive the finite-pool coverage target, which is generally not reward-proportional. To approximate this target, we propose Capacity-Aware Sampling (CaS), a local reweight-and-renormalize rule applied to the frozen policy. The penalty combines two deployment-time signals: a soft threshold on low-probability actions, and capacity-relative congestion of intermediate resources under the current sample history. CaS requires no retraining, no extra reward calls, and does not delete valid actions. Across five frozen GFlowNet benchmarks, a single default CaS setting improves finite-budget mode discovery over standard i.i.d. sampling on every task and achieves the best or near-best final-budget performance against tuned history-blind inference baselines. At , CaS increases discovered modes from to on QM9Str and from to on sEHStr.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Gflownet

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_pYXKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

After a GFlowNet is trained, it is typically deployed as a frozen model to draw a finite number of samples, with the goal of obtaining diverse, high-reward objects. The authors argue that reward-proportional i.i.d. sampling does not fully match this goal, and propose CAS (Capacity-Aware Sampling), which adjusts the frozen sampling policy by reweighting actions according to the sampling history so far.

Critical strengths

The motivation is clear and insightful, and it is the main reason I do not consider this paper a clear reject. The paper identifies a practical but largely overlooked problem in GFlowNet deployment, namely how to sample from a frozen model under a finite budget, whereas most existing work focuses on training.

Critical weaknesses

The readability of the paper is relatively low. The claims are not tied to clear intuition, and some concepts are used before they are clearly defined or described. Much of the content is presented as theorems and propositions, which is fine in itself, but the narrative connecting them is weak, and the link between the theory and the practical algorithm is thin. As a result, I found the paper hard to follow and was left doubting whether this mathematical machinery is necessary to introduce the practical algorithm.

The baselines do not appear to be well tuned (see Q3), and the frozen GFlowNet policies seem to be undertrained (see Q4). Both issues weaken the credibility of CAS's empirical advantage.

Note: The authors do not need to respond to the weaknesses above separately; answers to the questions below are sufficient.

Questions for authors
  1. The second limitation of standard frozen deployment (i.e., "the same amount of flow can congest a low-capacity state or transition") is confusing, as is Proposition 1. What is the actual cost of congestion at a low-capacity resource? In addition, the term capacity is used frequently before it is defined.
  2. Lemma 1 justifies the per-step reweighting form, but the penalty κ involves several heuristic design choices and many hyperparameters. For example, why are the soft threshold and the congestion ratio combined via a max? Moreover, with β = 1 the BPR function reduces to a ReLU; why do the authors introduce the BPR form?
  3. Nucleus sampling is reported in the main text with p = 0.8, whereas Figure 3 seems to suggest that p = 0.5 performs better on 4 of the 5 tasks. In addition, results for p < 0.5 and T < 0.5 are not reported. Could the authors explain these choices?
  4. The results of standard i.i.d. sampling (particularly on SEHSTR and QM9STR) suggest that the frozen GFlowNet policies may not be well trained: soft thresholding alone accounts for most of the improvement, and lowering the temperature alone improves performance on QM9STR dramatically. Could the authors comment on this?
  5. Proposition 2 appears to be a standard result. In fact, any support-preserving policy π can be written as the solution of Eq. 10 for some κ (e.g., κ(a) = log P_F(a|s) − log π(a)), so the KL-regularized form does not by itself justify the choice of κ. If κ is not connected to κ† in Lemma 1, what does Proposition 2 add beyond subtracting a heuristic penalty from the logits?
Reviewer_g72FOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Capacity-Aware Sampling for frozen GFlowNets. The method claims to improve sampling from frozen gflownet under finite-budget settings. The authors combinin soft thresholding with history-dependent congestion feedback to improve mode discovery.

Critical strengths

The distinction between reward-proportional training and finite-budget inference is valuable. The method requires no retraining or additional reward queries during generation, and the improvements on QM9Str are convincing. The ablations and experiments across training methods are useful.

Critical weaknesses

The empirical evaluation is limited to short, structured construction tasks. Results on a more complex generation tasks, such as molecule generation (SMILES), will better establish practical usefulness and scalability. Those scientific tasks also provide additional metrics to evaluate the quality, for example, using molecule fingerprint similarity to assess if meets mode collapse.

It is unclear which aspects of CAS specifically benefit from GFlowNet training or structure, since policy probabilities and visitation counts are also available in other RL methods. Moreover, the main comparisons are against static decoding methods. Comparing with a representative method that coordinates samples, such as stochastic beam search, under comparable computational budgets will better establish CAS’s added value.

Moreover, the sample coverage is a vital topic in GFlownet. From my side, the authors would better support the method if they can try to invlove it into training process, like improving the diverisity of buffer or samples in rollout state. Previous work like SubTB (Madan et al., ICML 2023), PRT/SSR (Shen et al., ICML 2023) [1], and RapTB (Wang et al., ICML 2026).

[1]: Towards Understanding and Improving GFlowNet Training

Questions for authors

See weakness part. I'm willing to raise my score if the authors can solve my concerns.

Reviewer_AW4eOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This submission studies the inference-time deployment of frozen Generative Flow Networks under a finite sampling budget. While standard GFlowNets are trained to make the terminal marginal distribution to be proportional to a non-negative reward, while drawing samples in an independent and i.i.d. way from this distribution, this work identifies this is suboptimal for discovering diverse modes under a fixed budget. More specifically, this paper identifies that high-probability regions of the terminal space are saturated diminishing the chance of discovering new modes through further sampling; and that static i.i.d. sampling cannot account for the actual realised path congestion relative to the local capacity of intermediate states and transitions. To resolve these issues without retraining the GFlowNet, this work proposes Capacity-Aware Sampling, a history-aware, sequential inference rule inspired by volume-to-capacity congestion models and penalty functions from transportation network literature, that penalises the frozen forward policy based on: i) the ration of actual visited counts of a state (or transition) to the expected flow predicted; and ii) a mechanism that penalises low-probability tail action, to prevent wasting budget on highly unlikely paths.

Critical strengths
  • The proposed formulation of post-training GFlowNet inference as a finite-pool resource allocation problem, is insightful and highly practical, directly addressing a major bottleneck in scientific discovery pipelines, with minimal computation overhead due to the sole reliance of forward-policy logits that does not incur any retraining and expensive reward evaluation costs.
  • The conducted evaluation is rigorous and showcases the effectiveness of the proposed approach in the examined usecases, against reasonable globally-tuned baselines. Additionally, the conducted analysis of failure cases (related to permutation-invariant tasks that collapse trajectory histories) offer valuable insights that can drive future work.
Critical weaknesses

In the current version of the manuscript, some unexplained disconnects seem to exist between the theoretical contribution and empirical evidence of this work. More specifically:

  • Although effective overall, the conducted ablation reveals that the examined history-unaware Soft Thresholding ablated baseline accounts for the vast majority of the performance gains; while combining history tracking with this baseline only yielding a moderate synergistic boost. This contrasts the manuscript's narrative that relies heavily on "capacity-aware history feedback".
  • Additionally, although the theoretical derivations show that optimal updates require a backward-looking continuation value, CAS replaces this with a forward-looking historical-congestion ratio. Further analysis and ablation of the underlying variants of these two alternatives should be conducted to clarify the conceptual gap.

Furthermore, some computational/efficiency aspects of the proposed approach are not adequately quantified. CAS's sequential state updates and transition count after each individual trajectory appears to break the ability of vanilla GFlowNets to parallelly sample and evaluate multiple trajectories via batching. The impact of this sequential dependency in wall-clock throughput is not examined adequately.

Questions for authors

Please consider replying to the concerns raised in the weaknesses section above.

1 comment

Given the absence of any response from the authors, I am decreasing my score from 2 to 1.

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

Related papers

Discussion (0)

Sign in to comment.