acceptodds
Under review as a conference paper at ICLR 2027

Geometry-guided Exploration with Adaptive Termination for GFlowNets

11 views

Abstract

Generative Flow Networks (GFlowNets) learn to sample compositional objects in proportion to their rewards, requiring exploration that discovers diverse high-reward objects. Existing exploratory policies can repeatedly acquire experiences that remain useful for learning but are already available through replay. We propose *Geometry-guided Exploration with Adaptive Termination (GEAT)*, a simple yet effective algorithm for efficient exploration in GFlowNets. Its core mechanism, *novelty-aware termination (NAT)*, generates terminal candidates through local backtracking and reconstruction, then selects the candidate with the highest novelty relative to reference states sampled from replay. Only the selected candidate requires a task-reward evaluation. To guide this selection, we derive *geometric novelty* from the derivative of geometry-aware Shannon entropy. The resulting score incorporates similarities between state representations and ranks candidates by their first-order effect on the entropy of the reference distribution. GEAT uses this score both as an intrinsic reward for the exploratory policy and as the terminal-selection criterion. Across extensive benchmarks, we show that GEAT consistently discovers high-reward modes faster and achieves broader mode coverage while maintaining competitive reward quality and distributional accuracy, demonstrating improved sample efficiency in mode discovery across diverse tasks.

Primary area:
generative models
Keywords:
GFlowNets, Amortized Sampling, Exploration, Geometric Novelty, Mode Discovery, Experience Replay

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HDWkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors argue that, during GFlowNet training, one should acquire experiences that are dissimilar to those already available in the replay buffer, rather than sampling trajectories simply because they are useful for learning. To this end, they propose Geometry-guided Exploration with Adaptive Termination (GEAT). GEAT derives a novelty score by differentiating a geometry-aware Shannon entropy at an empirical reference distribution uniformly sampled from the replay buffer. This score is used both to select among candidate terminal states generated by backtracking and reconstruction, and as an intrinsic reward for training the exploration GFlowNet. Across a range of benchmarks, GEAT demonstrates improved exploration efficiency.

Critical strengths

The motivation is clearly stated, and experiments across multiple benchmarks support the effectiveness of the method, particularly in terms of mode discovery. The authors identify a potentially overlooked opportunity to improve current exploration practice.

Critical weaknesses

The contribution appears to be a strategic combination of existing components: LSGFN, PER, and an exploration GFlowNet (as in SA-GFN), together with a novelty score computed from the replay buffer to guide both the local search and the exploration GFlowNet. Notably, in the ablation study, the NAT-only variant performs almost identically to LSGFN. While the combination works well empirically, the connection between these design choices is not well established, which makes the novelty of the contribution unclear. If the authors can adequately address this point and Q1, and the additional results requested in Q2 and Q3 do not change the conclusions of the experiments, I would be willing to raise my rating.

Questions for authors
  1. The paper argues that exploration should acquire experiences dissimilar to those already in the replay buffer, and the experiments use PER for replay. However, GEAT decouples the novelty computation from the PER rule: the reference set is sampled from the replay buffer uniformly rather than according to the PER distribution. This raises two concerns. First, a trajectory that remains in the buffer but has low PER priority is rarely sampled for training, yet it is still treated as "covered" by the novelty score. Second, if the buffer has finite capacity, an evicted sample will be treated as "uncovered" again, even if it has already been well learned. The buffer size and eviction policy are also not reported. Could the authors clarify the relationship between the novelty score and the PER rule?
  2. All baselines are run with PER, although many of them, unlike GEAT, do not necessarily require a replay buffer. Could the authors also report the performance of the baselines without a replay buffer?
  3. GEAT relies on hand-designed, domain-specific kernels, whereas SA-GFN, the only baseline with an intrinsic novelty signal, uses RND. Could the authors report results for the baselines equipped with the same kernel on each task (e.g., SA-GFN with RND replaced by the kernel-based novelty)?
5 comments

We appreciate your time and feedback on our paper. Below are our responses to your comments.


Justification for Component Connection

We appreciate the opportunity to further clarify the rationale behind GEAT's design.

We have revised the manuscript to explain why GEAT combines geometric novelty and NAT: expanding exploration coverage requires both reaching less-explored regions and selecting novel terminal states within them. These two components address complementary decisions.

  • Geometric novelty guides NAT's local search toward less-explored regions. NAT selects among local proposals, so its effectiveness depends on where the exploratory policy reaches. By combining task rewards with geometric novelty, the exploratory policy guides NAT toward promising but less-explored regions. Without this guidance, NAT can remain near previously discovered high-reward basins, helping explain the similar performance of NAT-only and LSGFN.
  • NAT makes novelty the explicit criterion for terminal-state acquisition. Geometric novelty alone cannot guarantee novel terminal-state acquisition because task rewards also influence the exploratory policy. NAT addresses this by selecting the most novel terminal state among local proposals. Thus, geometric novelty guides where NAT searches, while NAT determines which terminal state is acquired.

The contribution lies in connecting these two stages to improve exploration coverage. We have incorporated this discussion and supporting experimental evidence into Section 5 (Experiments). An additional visualization, available here, compares the baseline, NAT-only, geometric-only, and full GEAT: (a) shows cumulative explorer terminal visits across training steps, while (b) tracks the number of high-reward basins reached during training.


Baselines without Replay

Removing replay generally reduces baseline performance, and none of the replay-free baselines outperforms GEAT in mode discovery. This suggests that our shared replay protocol does not systematically disadvantage the baselines.

To verify this, we evaluated TB, SA-GFN, Adaptive Teacher, and ACE with and without replay on HyperGrid, Bit Sequence (K=8), and QM9. Without replay, each baseline is trained only on newly collected experiences, retaining its original exploration mechanism and sampling budget. Replay-enabled configurations use a shared buffer capacity of 6,553 unique terminal states.

All values are reported as Replay ON → OFF, with mean ± sample standard deviation over three seeds.

HyperGrid (D=2, H=256)

MethodModes ↑TV ↓
TB1,127.7 ± 78.7 → 634.7 ± 2.10.5566 ± 0.0221 → 0.7597 ± 7×10⁻⁶
SA-GFN2,509.3 ± 55.7 → 2,086.7 ± 67.20.2778 ± 0.0558 → 0.7598 ± 1.6×10⁻⁴
Teacher1,301.0 ± 0.0 → 1,879.7 ± 267.40.5362 ± 0.0509 → 0.7597 ± 3×10⁻⁶
ACE1,301.0 ± 0.0 → 1,292.7 ± 284.70.5070 ± 0.0064 → 0.7598 ± 6.4×10⁻⁵

Removing replay substantially worsens TV distance for all baselines, even when mode discovery increases, as observed for Teacher.

Bit Sequence (K=8)

MethodModes ↑Top-100 ↑FCS ↓
TB23.3 ± 0.6 → 25.3 ± 2.58.459 ± 0.012 → 8.448 ± 0.0160.0249 ± 0.0016 → 0.0259 ± 0.0022
SA-GFN35.7 ± 3.5 → 34.7 ± 2.58.477 ± 0.009 → 8.476 ± 0.0100.0244 ± 0.0003 → 0.0222 ± 0.0010
Teacher36.0 ± 2.0 → 34.7 ± 3.18.488 ± 0.008 → 8.480 ± 0.0040.0232 ± 0.0002 → 0.0223 ± 0.0004
ACE35.0 ± 1.7 → 35.7 ± 4.08.485 ± 0.010 → 8.483 ± 0.0060.0239 ± 0.0007 → 0.0243 ± 0.0003

Replay has a relatively limited effect on Bit Sequence, with comparable mode discovery and Top-100 rewards across configurations.

Bio (QM9)

MethodModes ↑Top-100 ↑FCS ↓
TB601.0 ± 25.5 → 470.0 ± 55.99.239 ± 0.130 → 9.039 ± 0.1030.0894 ± 0.0138 → 0.0957 ± 0.0111
SA-GFN703.0 ± 11.5 → 625.7 ± 25.59.219 ± 0.089 → 9.024 ± 0.0410.0800 ± 0.0066 → 0.0752 ± 0.0037
Teacher726.7 ± 15.5 → 658.3 ± 17.09.120 ± 0.065 → 9.000 ± 0.0820.0821 ± 0.0051 → 0.0847 ± 0.0087
ACE720.3 ± 6.7 → 636.3 ± 27.39.149 ± 0.051 → 9.078 ± 0.1060.0780 ± 0.0027 → 0.0845 ± 0.0061

Removing replay reduces mode discovery by 9.4–21.8% and lowers Top-100 rewards across all baselines, although its effect on FCS is mixed.

All results are reported as mean ± sample standard deviation over three seeds. HyperGrid and QM9 use 2,000 training iterations; Bit Sequence uses 20,000.

Overall, replay improves baseline performance on HyperGrid and QM9, while its effect on Bit Sequence is relatively small. These results provide no evidence that our shared replay protocol systematically disadvantages the baselines.


The Separation Between Novelty Computation and PER Is Intentional

Novelty measures what has already been collected, while PER prioritizes which stored experiences need further learning. This separation prevents learning priorities from driving unnecessary data collection.

  • Low-priority samples are correctly treated as covered. Under loss-based PER, low priority may indicate a small TB residual. Regardless of training priority, stored experiences can be reused through replay without additional reward evaluations.
  • PER-weighted references would make covered regions appear novel. Previously collected, low-priority samples would be underrepresented in the reference set, appear novel, and attract repeated acquisition, which is exactly the redundancy GEAT aims to avoid.

Our additional experiments support this distinction. As illustrated here, PER-weighted reference sampling can make previously visited, low-priority states appear more novel.

To examine this effect, we compared uniform and PER-weighted references using identical replay buffers and candidate sets. We focused on previously visited, low-reward terminal states in the bottom 25% of PER priorities, averaging results over training iterations 500–2,000.

MetricUniform ReferencesPER-Weighted References
Reference Inclusion Rate15.61%0.23%
Geometric Novelty1.452.57
Selection Rate1.49%3.43%

PER-weighted references substantially increase both the estimated novelty and selection rate of previously collected, low-priority states. This indicates that incorporating learning priorities into novelty estimation can redirect exploration toward experiences assigned relatively low priority for further training.

We also compared the final performance of uniform and PER-weighted reference sampling:

Reference SamplingTV ↓Discovered Modes ↑
Uniform GEAT0.223 ± 0.0632,574 ± 8
PER-Weighted GEAT0.279 ± 0.0622,388 ± 158

Mean ± sample standard deviation over three seeds.

Uniform GEAT achieves better average final performance in these runs. Together, these results support separating novelty-based coverage estimation from PER-based learning priorities.


Effect of Buffer Eviction

GEAT maintains comparable mode coverage across buffer sizes, while removing task-policy sampling substantially reduces mode discovery. The task GFlowNet continues to sample from its learned reward-proportional distribution and replenish the shared buffer, so eviction does not necessarily leave previously explored high-reward regions uncovered.

Buffer configuration and ablation. All main experiments use a shared FIFO replay buffer with a capacity of 6,553 unique terminal states, as now reported in App. D. To assess sensitivity to buffer eviction, we evaluated GEAT on QM9 with three buffer capacities (0.25×, 1×, and 4×), with and without task-policy sampling.

Both settings use the same budget of 64 fresh samples per iteration: 32 task and 32 exploration samples when task sampling is enabled, versus 64 exploration samples otherwise. The task GFlowNet remains trained off-policy in both settings.

QM9: Buffer-Size and Task-Sampling Ablation

Buffer SizeTask SamplingModes ↑Top-100 ↑FCS ↓
1,638 (0.25×)ON734.3 ± 8.19.256 ± 0.1330.0879 ± 0.0028
OFF514.7 ± 5.09.349 ± 0.0590.1219 ± 0.0064
6,553 (1×)ON739.7 ± 3.59.208 ± 0.1130.0922 ± 0.0079
OFF530.3 ± 8.69.322 ± 0.2010.1203 ± 0.0120
26,212 (4×)ON738.0 ± 5.39.293 ± 0.2180.0929 ± 0.0056
OFF542.7 ± 7.29.311 ± 0.0510.1091 ± 0.0037

Mean ± sample standard deviation over three seeds.

Results. With task-policy sampling, GEAT maintains comparable mode coverage across all buffer sizes. In contrast, removing task-policy sampling reduces mode discovery by 26–30% and increases FCS across all buffer sizes, while Top-100 rewards remain comparable.

These results support GEAT's robustness to buffer capacity and highlight the importance of continued task-policy sampling for maintaining mode coverage.


Fairness of Kernel-Based Novelty Comparison

GEAT's mode-discovery improvements cannot be attributed solely to the geometric novelty kernel or its alignment with the evaluation metric. We conducted two additional comparisons to examine these potential concerns.

1. Comparison Using the Same Geometric Novelty

GEAT discovers more high-reward modes than SA-GFN even when both use the same geometric novelty. To isolate the effect of the novelty kernel, we evaluated SA-GFN + Geo, replacing SA-GFN's RND intrinsic reward with GEAT's geometric novelty while keeping all other components unchanged.

Number of Discovered Modes ↑

BenchmarkSA-GFN (RND)SA-GFN + GeoGEAT
Bit Sequence (K=8)35.7 ± 2.535.0 ± 3.555.7 ± 2.5
Bio (AMP)300.3 ± 16.0301.3 ± 18.8442.0 ± 11.1

Mean ± sample standard deviation over three seeds.

These results indicate that GEAT's improvements cannot be explained solely by its choice of geometric novelty.

2. Evaluation Using Alternative Mode Definitions

GEAT maintains the highest average mode discovery even when the evaluation distances differ from those used in its novelty kernel.

GEAT originally uses Levenshtein-distance-based geometric novelty for Bit Sequence and ProtAlbert embedding-based L2 geometric novelty for AMP. The original mode-discovery criteria also rely on these respective distances.

To assess whether this alignment favors GEAT, we kept the novelty kernels unchanged and reevaluated mode discovery using:

  • Bit Sequence: Hamming and Jensen–Shannon (JS) distances.
  • AMP: Hamming and Levenshtein distances.

We tested three thresholds for each distance metric.

Bit Sequence: Number of Discovered Modes ↑

MethodHamming 0.26Hamming 0.275Hamming 0.29JS 0.58JS 0.60JS 0.62
TB2.3 ± 1.216.7 ± 3.832.7 ± 4.022.0 ± 8.039.0 ± 14.050.7 ± 10.2
Teacher5.0 ± 1.728.3 ± 1.548.3 ± 0.622.7 ± 7.041.3 ± 10.854.0 ± 5.2
SA-GFN3.3 ± 2.127.3 ± 0.646.3 ± 1.522.0 ± 8.040.0 ± 12.352.3 ± 8.1
ACE3.3 ± 2.127.3 ± 3.246.7 ± 2.317.7 ± 15.035.0 ± 21.751.0 ± 11.3
GEAT19.0 ± 5.652.7 ± 2.357.7 ± 1.559.3 ± 0.659.7 ± 0.660.0 ± 0.0

Mean ± sample standard deviation over three seeds. JS denotes Jensen–Shannon.

AMP: Number of Discovered Modes ↑

MethodHamming 0.75Hamming 0.80Hamming 0.85Levenshtein 0.75Levenshtein 0.80Levenshtein 0.85
TB373.3 ± 16.3364.7 ± 14.2157.3 ± 7.8336.0 ± 17.7156.0 ± 4.618.3 ± 1.2
Teacher462.7 ± 29.3450.7 ± 27.5176.3 ± 12.6409.7 ± 18.1177.7 ± 5.519.7 ± 1.2
SA-GFN605.3 ± 19.4579.0 ± 19.9198.7 ± 8.0518.0 ± 4.0199.7 ± 10.120.3 ± 0.6
ACE615.0 ± 41.6588.3 ± 35.9197.3 ± 5.7517.0 ± 24.3205.3 ± 10.622.3 ± 0.6
GEAT872.3 ± 21.5820.3 ± 22.9237.7 ± 6.7702.0 ± 22.0239.0 ± 14.424.3 ± 1.5

GEAT achieves the highest average mode discovery across all twelve alternative evaluation settings, supporting that its advantage is not merely due to alignment between the novelty kernel and evaluation metric.

Reviewer_mArLOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a GFlowNet training procedure focused on adaptive exploration of the environment. The main GFlowNet policy is trained via the task reward, and a separate exploratory policy is trained via a combination of the task reward and the novelty score. The novelty-aware termination mechanism explores the neighborhood of a generated sample through backtracking and reconstruction, then picks the candidate with the highest novelty score and generates a trajectory leading to it through the backward policy, adding it to a replay buffer. The novelty score itself is defined as a derivative of geometry-aware Shannon entropy. Experimental evaluation shows improved convergence speed and mode discovery compared to other adaptive exploration methods for GFlowNet training.

Critical strengths

The problem of exploration that the paper studies is a crucial and important challenge in GFlowNet training. The paper is very well-written and easy to follow, I perfectly understood all components of the proposed method, the experimental setup and all experimental results from the first read. The proposed training procedure looks sound to me, and, while it is a combination of many components, the use of all of them is well-motivated. Algorithm 1 provides a full and clear pseudocode of the method. The experimental evaluation is carried out on established challenging benchmarks from the prior GFlowNet literature. Some of the experimental results demonstrate impressive improvements over the baselines. I also appreciated the additional ablation studies in Appendix.

Critical weaknesses

In my eyes, the crucial weakness of this paper is the fact that it has little to no novelty. Practically all components of the proposed training methods were introduced in prior GFlowNet and reinforcement learning literature, and the paper simply demonstrates that a certain combination of them produces good results in practice. Training a separate policy for adaptive exploration was studied in a plethora of previous works (Kim et al., 2025; Malek et al., 2026; Dall’Antonia et al., 2026). The local exploration method based on backtracking and reconstruction was proposed in Kim et al. (2024). Replay buffers are also common in GFlowNet training (Vemgal et al., 2023; Shen et al., 2023), and prioritized experience replay specifically was used for GFlowNet training in [1]. The proposed novelty score is also directly based on prior RL literature (Posada et al., 2020; Guo et al., 2021). I appreciate the fact that the authors honestly attribute all of these components to prior literature. However, in my eyes, the paper brings no new knowledge from both the methodological and engeneering prospectives, since the method itself is a straightforward training pipeline that combines existing techniques.

I also have concerns regarding the experimental evaluation. Firstly, the fairness of the presented mode discovery comparisons. In a number of tasks, the kernel encodes the same metric under which modes are counted. In bit sequences, modes are counted by Levenstein distance, which the kernel directly uses. The same with ProtAlbert embedding distance in the AMP environment. This means that the novelty score during training gives the model information that is correlated with the evaluation metric, while baselines do not get such information. If the samples produced during novelty-aware termination are counted for mode discovery, this is an additional algorithmic source for mode discovery that the baselines do not have. A more fair comparison would count only the samples generated by the main GFlowNet policy and the exploratory policy, as well as use different distance metrics to define the kernel for the novelty score, so that there is no indirect leak of the evaluation metric. Secondly, while the number of reward evaluations is aligned across all baselines, the proposed method requires additional neural network forward passes in the local exploration phase. The authors should also include the comparison with baselines in terms of the number of forward passes and/or wall-time aligned training curves.

A more minor point, but I want to point out that all experiments are performed with TB loss, while there are other losses in GFlowNet literature that have been shown to result in faster convergence, e.g., SubTB [2]. I understand that the proposed approach is agnostic to the utilized loss, as long as it is computed on whole trajectories, but it would be nice to add a few experiments demonstrating the compatibility with other losses in practice.

References:
[1] Tiapkin et al. Generative Flow Networks as Entropy-Regularized RL. AISTATS 2024.
[2] Madan et al. Learning GFlowNets from partial episodes for improved convergence and stability. ICML 2023.

Questions for authors

See Weaknesses.

I also have an additional question regarding the experimental results. Figure 3 shows that the best methods converge to total variation distance values between 0.2 and 0.4. The first strange thing is that these values are rather large, considering that the authors evaluate the exact TV. These environments are very small, so one should expect the GFlowNet to converge to the optimum where the target distribution is sampled exactly, which should result in TV values closer to 0. Do the authors have an explanation for this?

4 comments

We appreciate you taking the time to review our paper and provide valuable feedback. Below are our responses to your questions.


New Perspective and Methodological Novelty

We appreciate this opportunity to clarify and emphasize our contributions.

GEAT builds on well-established components from prior work. We deliberately leverage these components to implement our central insight in a simple and direct manner. However, the main contribution of GEAT lies in a new perspective on exploration and a mechanism that changes how exploration data are collected, as detailed below:

  • New perspective: Distinguishing exploration from learning. Existing methods often direct exploration toward regions requiring further learning, such as those with high training loss or insufficient probability mass (Kim et al., 2025; Malek et al., 2026; Dall’Antonia et al., 2026). However, the need for further learning does not necessarily imply the need for additional data collection, since previously acquired experiences can be reused through replay. To the best of our knowledge, we are the first to explicitly incorporate this distinction into the design of a GFlowNet exploration algorithm, directing new data collection toward expanding coverage while relying on replay for further learning.
  • New mechanism: Using novelty for terminal-state acquisition. Prior intrinsic-motivation-based methods use novelty simply as an intrinsic reward combined with the extrinsic task reward through a weighted sum to train the exploratory policy (Madan et al., 2025). In contrast, GEAT uses novelty as an explicit terminal-state selection criterion. While the exploratory policy guides candidate generation toward promising regions, NAT selects the most novel candidate before evaluating its task reward. Although the local proposals build on backtracking and reconstruction (Kim et al., 2024), novelty determines which terminal state is actually acquired, preventing high extrinsic rewards from dominating the final selection.
  • Novel Criterion: Replay-conditioned geometric novelty. Guo et al. (2021) maximize geometry-aware state-visitation entropy through a learned auxiliary function and stochastic intrinsic rewards. GEAT instead directly evaluates the entropy derivative at an empirical distribution of previously collected terminal states, interpreting the resulting score as each candidate's marginal contribution to coverage. We do not claim a new entropy functional; our contribution lies in using this entropy-based signal as a replay-conditioned criterion for terminal-state acquisition.

To further illustrate these distinctions, we provide cumulative exploration-visitation visualizations here. These visualizations show that GEAT's exploration pattern differs markedly from those of all evaluated baselines. We have also revised the manuscript to better articulate our contributions.


Computational Cost

Despite its additional computational cost, GEAT achieves the highest mode-discovery rate within the same wall-clock budget across all three benchmarks, as shown in the wall-clock mode-discovery curves available here. We clarify the computational cost of NAT and provide the corresponding wall-clock comparisons below.

Wall-clock comparison. We measure modes/min using a common wall-clock budget defined by TB's total training time, while reporting each method's full training time separately.

Modes/min ↑ / Training Time (min) ↓

MethodHyperGridBit SequencesEH
TB53.72 ± 4.64 / 21.0 ± 0.70.750 ± 0.004 / 45.4 ± 0.23.52 ± 0.24 / 1100.4 ± 12.6
SA-GFN61.61 ± 1.54 / 56.4 ± 1.60.426 ± 0.143 / 127.1 ± 1.43.08 ± 0.10 / 1603.7 ± 32.3
Adaptive Teacher60.56 ± 2.12 / 47.6 ± 4.20.611 ± 0.090 / 71.8 ± 0.73.12 ± 0.08 / 1741.2 ± 4.3
ACE60.22 ± 2.98 / 59.6 ± 3.20.569 ± 0.112 / 73.4 ± 1.12.87 ± 0.23 / 1725.8 ± 38.3
GEAT80.28 ± 7.45 / 57.7 ± 2.81.256 ± 0.045 / 94.1 ± 1.612.21 ± 1.37 / 2131.5 ± 210.6

Cost analysis. For trajectories of length TT, sampling one task and one exploration trajectory requires 2T2T forward-policy calls. Assuming kk forward reconstruction steps per proposal, NAT adds nknk forward calls and, when PBP_B is learned, nknk backward calls for proposal generation plus TT backward calls to generate the training trajectory after terminal selection. Thus, the policy-sampling cost ratio with versus without NAT is approximately 1+nk2T1+\frac{nk}{2T} for fixed or deterministic PBP_B and 32+nkT\frac{3}{2}+\frac{nk}{T} for learned PBP_B.


Fairness of Mode Discovery Comparisons

GEAT's mode-discovery improvements are not solely attributable to alignment between the novelty kernel and evaluation metric.

GEAT uses Levenshtein-distance-based geometric novelty for Bit Sequence and ProtAlbert embedding-based L2 geometric novelty for AMP. Since the original mode-discovery criteria also use these respective distances, we kept the novelty kernels unchanged and reevaluated mode discovery using alternative distances and thresholds:

  • Bit Sequence: Hamming and Jensen–Shannon (JS) distances.
  • AMP: Hamming and Levenshtein distances.

QM9 and sEH already use evaluation criteria distinct from GEAT's novelty kernel.

Bit Sequence: Number of Discovered Modes ↑

MethodHamming 0.26Hamming 0.275Hamming 0.29JS 0.58JS 0.60JS 0.62
TB2.3 ± 1.216.7 ± 3.832.7 ± 4.022.0 ± 8.039.0 ± 14.050.7 ± 10.2
Teacher5.0 ± 1.728.3 ± 1.548.3 ± 0.622.7 ± 7.041.3 ± 10.854.0 ± 5.2
SA-GFN3.3 ± 2.127.3 ± 0.646.3 ± 1.522.0 ± 8.040.0 ± 12.352.3 ± 8.1
ACE3.3 ± 2.127.3 ± 3.246.7 ± 2.317.7 ± 15.035.0 ± 21.751.0 ± 11.3
GEAT19.0 ± 5.652.7 ± 2.357.7 ± 1.559.3 ± 0.659.7 ± 0.660.0 ± 0.0

Mean ± sample standard deviation over three seeds. JS denotes Jensen–Shannon.

AMP: Number of Discovered Modes ↑

MethodHamming 0.75Hamming 0.80Hamming 0.85Levenshtein 0.75Levenshtein 0.80Levenshtein 0.85
TB373.3 ± 16.3364.7 ± 14.2157.3 ± 7.8336.0 ± 17.7156.0 ± 4.618.3 ± 1.2
Teacher462.7 ± 29.3450.7 ± 27.5176.3 ± 12.6409.7 ± 18.1177.7 ± 5.519.7 ± 1.2
SA-GFN605.3 ± 19.4579.0 ± 19.9198.7 ± 8.0518.0 ± 4.0199.7 ± 10.120.3 ± 0.6
ACE615.0 ± 41.6588.3 ± 35.9197.3 ± 5.7517.0 ± 24.3205.3 ± 10.622.3 ± 0.6
GEAT872.3 ± 21.5820.3 ± 22.9237.7 ± 6.7702.0 ± 22.0239.0 ± 14.424.3 ± 1.5

GEAT achieves the highest average mode discovery across all 12 alternative evaluation settings, supporting that its advantage is not solely driven by alignment between the novelty kernel and evaluation metric.

Mode-counting protocol. Only terminal states selected by NAT and those sampled by the task GFlowNet contribute to mode discovery. Unselected local proposals are excluded.


Compatibility with Other Losses

GEAT achieves higher mode discovery and lower TV distance under both TB and SubTB, demonstrating that its effectiveness is not limited to TB loss. We evaluated Subtrajectory Balance (SubTB) on HyperGrid (2D, H=256), using λ=0.9\lambda=0.9, a fixed backward policy, and 2,000 training iterations.

MethodTraining LossNumber of Modes ↑TV Distance ↓
GFlowNetTB1,158 ± 360.5494 ± 0.0165
GFlowNetSubTB2,235 ± 640.2441 ± 0.0136
GEATTB2,576 ± 240.2248 ± 0.0865
GEATSubTB2,498 ± 500.2122 ± 0.0090

Mean ± sample standard deviation over three seeds.


High Total Variation Distance in HyperGrid

The relatively high TV distances in Figure 3 are due to the limited number of training iterations. To verify this, we extended the training duration and observed that the TV distances approach zero with continued training.

The extended-training results, available here, show that GEAT achieves the fastest convergence toward zero TV distance among the compared methods.

I thank the authors for their detailed responses and apologize for the late reply. The presented computational cost comparisons are fair in my opinion, and new experiments further support mode discovery claims. However, I retain my opinion regarding the limited novelty of the proposed approach, and decided to keep my score. I will leave the final decision on whether the novelty of the contributions are sufficient for a conference of this level up to AC discretion.

What prior work left unasked, and what GEAT contributes as a whole

We sincerely appreciate your careful consideration of our responses and your constructive engagement throughout the discussion. We are particularly grateful for your recognition that our additional experiments further substantiate GEAT's substantial improvements in high-reward mode discovery, as well as for your assessment that our computational cost comparisons are fair.

Before concluding, we would like to offer one final clarification regarding what we believe to be the central contribution of GEAT to GFlowNet exploration. Looking back at our previous responses, we realize that we focused primarily on explaining the novelty of individual components, without sufficiently articulating what new research question motivates their integration and what GEAT contributes as a whole. We hope the following perspective helps clarify this distinction.

What this paper is about. The paper concerns a connection that is easy to overlook in GFlowNet training: the interaction between the exploration policy and the replay buffer. That replay improves sample efficiency is established [1–3], and it is standard in practice. What has not been asked is how the exploration policy should be designed once replay is taken as given. Existing methods either omit the buffer from the algorithm (ACE [4]) or include one while training the explorer on the task model’s loss, the same signal prioritized replay follows (Adaptive Teacher [3]). In both cases the explorer is designed in isolation from replay.

What changes when replay is taken as given. (1) Decoupling Learning from Experience Acquisition. Accounting for a replay buffer allows us to separate two design choices for learning and collecting new trajectories. This enables much more aggressive, targeted exploration without destabilizing off-policy training. (2) Buffer-Driven, Training-Free Novelty Estimation. Rather than relying on auxiliary neural networks (e.g., RND [5]) to track state visitation, the replay buffer itself provides a direct historical record of explored regions, enabling lightweight, non-parametric novelty computation via Geometric Entropy.

Why a new algorithm. Our contribution is not simply demonstrating that replay improves sample efficiency, but identifying how exploration should be designed to complement replay. This motivates several concrete design choices, which we systematically evaluate: the novelty reference distribution (uniform vs. PER-weighted; Table 3), the novelty criterion (count, RND, geometric; Fig. 2), where novelty acts (reward, selection, or both; Fig. 7), the terminal selection rule (max, softmax, uniform; Fig. 7), and comparisons with reward-based and MH-based local search [6] (Fig. 11). These experiments support the effectiveness of GEAT's coupled design, which delivers substantial improvements in high-reward mode discovery. Our primary contribution is thus not merely combining existing techniques, but proposing and empirically validating a new exploration–replay design principle that leads to more efficient experience acquisition.

Once again, we sincerely thank you for your careful review, constructive suggestions, and continued engagement with our work. Your feedback has helped us substantially strengthen both the experimental evaluation and the presentation of our contributions. We are grateful for the opportunity to clarify our perspective, and we hope this final explanation helps convey the overall contribution of our work.


References:

[1] Vemgal et al., An Empirical Study of the Effectiveness of Using a Replay Buffer on Mode Discovery in GFlowNets, ICML Workshop, 2023.

[2] Shen et al., Towards Understanding and Improving GFlowNet Training, ICML, 2023.

[3] Kim et al., Adaptive Teachers for Amortized Samplers, ICLR, 2025.

[4] Dall'Antonia et al., Avoid What You Know: Divergent Trajectory Balance for GFlowNets, ICML, 2026.

[5] Burda et al., Exploration by Random Network Distillation, ICLR, 2019.

[6] Kim et al., Local Search GFlowNets, ICLR, 2024.

Reviewer_5oPnOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an exploration method for GFlowNets called Geometry-guided Exploration with Adaptive Termination (GEAT). The authors aim to improve exploration by training two GFlowNets: a task GFlowNet that seeks to sample terminal states proportionally to their rewards, and an exploration GFlowNet that takes into account both the task reward and a geometric novelty bonus.

The training algorithm samples trajectories from both the task and exploration GFlowNets. Starting from the terminal states sampled by the exploration GFlowNet, it generates additional candidates through backtracking and reconstruction, selecting the candidate with the highest novelty score. This procedure is called Novelty-Aware Termination (NAT), while the geometric novelty bonus is derived from the gradient of geometry-aware Shannon entropy.

The authors evaluate their method on HyperGrid, Bit Sequence, QM9, sEH, and antimicrobial peptide (AMP) benchmarks, comparing it against TB, SA-GFN, Adaptive Teacher, and ACE. The results show substantial improvements in high-reward mode discovery while maintaining competitive measures of distributional consistency for the task GFlowNet.

Critical strengths

1. Interesting formulation of coverage using geometry-aware Shannon entropy.

The authors use geometry-aware Shannon entropy to evaluate whether new terminal states represent regions that have not yet been sufficiently explored. More specifically, they derive a geometric novelty score from its gradient, allowing the method to account for similarities between different terminal states. This seems to be an interesting mechanism for improving exploration.

2. Distinguishing learning from acquiring new experiences.

The paper makes an interesting distinction between regions where the GFlowNet still needs to learn and regions where new experiences need to be acquired. Since GFlowNets support off-policy training, previously collected trajectories can be reused for learning, while exploration can focus on acquiring new experiences in less-explored regions. I believe this is an interesting motivation for the proposed method.

Critical weaknesses

1. Issues with the baseline implementations.

My main concern is with some implementation choices in the sEH environment, where the authors inconsistently add a constant of 1e-8 to the reward. Although this value may seem negligible, many terminal objects in this environment have rewards considerably smaller than 1e-8.

In Adaptive Teacher, while the Student is trained using the original reward RR, the Teacher computes its discrepancy using R+10−8R + 10^{-8}. This directly affects how the discrepancy is calculated. Even when the Student satisfies the original Trajectory Balance condition, the Teacher may still receive a nonzero discrepancy signal, directly affecting how it explores the state space.

A similar issue occurs in ACE, which relies on the criterion R^(x)≥αR(x)\hat{R}(x) \geq \alpha R(x) to identify over-allocated and under-allocated states. Again, the inconsistency arises because the constant is added to the reward in the threshold, but not to the estimated reward R^(x)\hat{R}(x). This can substantially change how states are classified as over-allocated or under-allocated and, consequently, affect its exploration behavior.

2. Benefits versus computational cost.

GEAT performs additional backtracking, reconstruction, and geometric novelty evaluations at each iteration. For each terminal state sampled by the exploration GFlowNet, the algorithm generates nn additional candidates, each involving kk backward steps followed by reconstruction. Therefore, in practice, GEAT performs considerably more trajectory-generation work than the other methods.

Although this additional computational cost may be relatively small in environments where reward evaluation dominates the total cost, I believe this trade-off should be discussed more explicitly. The paper would benefit from comparisons based on wall-clock training time, in addition to the number of acquired trajectories or training iterations.

3. Additional ablations.

I would also like to see more values of ϵ\epsilon, corresponding to the random-action probability in the epsilon-greedy policy, being tested. Evaluating only one value per environment does not seem sufficient to establish the robustness of the results, especially when comparing different exploration methods. I also believe the paper would benefit from experiments using learned backward policies instead of fixed uniform ones. The choice of backward policy can substantially affect the training and exploration behavior of GFlowNets.

4. Limitations of FCS as an evaluation metric.

I have concerns about the interpretation of the FCS values reported in the experiments. The evaluation uses terminal states sampled from each method's own task GFlowNet, which means that it only assesses whether the probabilities are consistent with the rewards within the sampled subsets.

This does not necessarily indicate that the learned distribution correctly covers the target distribution. For example, if the task GFlowNet collapses to a single terminal state, all evaluation samples could correspond to that same state. After removing duplicates and normalizing the probabilities and rewards within each group, the FCS would be zero, which is the best possible value, despite the complete mode collapse.

Therefore, I believe FCS has a substantial limitation as a measure of global distributional accuracy, and the reported results should be interpreted with this limitation in mind.

Questions for authors

1. Reward implementation in sEH.

Given that the constant 10−810^{-8} is applied inconsistently in the implementations of ACE and Adaptive Teacher, could the authors provide an ablation showing the results after correcting these inconsistencies?

2. Computational cost.

Given the additional trajectory-generation work performed by GEAT, it would be interesting to see a comparison based on wall-clock training time between GEAT and the other methods.

3. Additional ablations.

How sensitive are the results to different values of ϵ\epsilon in the epsilon-greedy policy? It would also be interesting to see an ablation of both GEAT and the other methods using learned backward policies instead of fixed uniform ones.

5 comments

We appreciate your time to review our paper and your feedback. We address your questions below.


Correction of Reward Inconsistency in sEH Baselines

The 10−810^{-8} reward offset was used as a safeguard against numerical instability and is unrelated to GEAT's proposed components. That said, we have removed this offset consistently and rerun the sEH experiments.

MethodModesTop-100FCS
Adaptive Teacher4869.7 ± 119.76.6320 ± 0.01790.6615 ± 0.0241
ACE4757.3 ± 85.46.6320 ± 0.01790.6631 ± 0.0333
GEAT18711.0 ± 688.26.8018 ± 0.09990.5973 ± 0.0425

The correction does not change our experimental conclusions. With the reward offset consistently removed, GEAT retains the highest average number of discovered modes and Top-100 reward, as well as the lowest average FCS, among the reevaluated methods.


FCS as an Evaluation Metric

Use of FCS follows existing evaluation practice rather than constituting a methodological contribution of GEAT. Importantly, our evaluation does not rely solely on FCS. We also report high-reward mode discovery across all environments and exact TV distance on HyperGrid. Together, these metrics assess both distributional consistency and exploration coverage.

We additionally report Spearman correlation and Pearson correlation here.

Spearman Correlation ↑

BenchmarkTBTeacherSA-GFNACEGEAT
Bit K=40.767 ± 0.0310.777 ± 0.0060.777 ± 0.0110.768 ± 0.0050.777 ± 0.003
Bit K=60.708 ± 0.0180.728 ± 0.0020.719 ± 0.0250.725 ± 0.0100.728 ± 0.011
Bit K=80.711 ± 0.0160.734 ± 0.0210.737 ± 0.0250.736 ± 0.0160.738 ± 0.017
QM90.926 ± 0.0030.930 ± 0.0160.929 ± 0.0060.932 ± 0.0020.933 ± 0.010
sEH0.609 ± 0.0090.696 ± 0.0120.688 ± 0.0090.714 ± 0.0050.715 ± 0.006
AMP0.870 ± 0.0020.871 ± 0.0010.872 ± 0.0020.872 ± 0.0030.873 ± 0.002

Pearson Correlation ↑

BenchmarkTBTeacherSA-GFNACEGEAT
Bit K=40.713 ± 0.0400.714 ± 0.0100.708 ± 0.0250.651 ± 0.1150.724 ± 0.012
Bit K=60.639 ± 0.0250.649 ± 0.0290.633 ± 0.0250.622 ± 0.0490.663 ± 0.018
Bit K=80.669 ± 0.0230.693 ± 0.0540.676 ± 0.0410.673 ± 0.0210.699 ± 0.068
QM90.772 ± 0.0080.844 ± 0.0130.805 ± 0.0120.783 ± 0.0080.892 ± 0.004
sEH0.622 ± 0.0170.731 ± 0.0100.730 ± 0.0190.739 ± 0.0020.749 ± 0.006
AMP0.838 ± 0.0010.843 ± 0.0010.840 ± 0.0020.840 ± 0.0030.841 ± 0.001

GEAT achieves the highest Spearman correlation across all six settings, while ranking first in Pearson correlation in five settings and second in the remaining one. These results provide additional evidence of consistency between rewards and learned sampling probabilities.


Computational Cost

Despite its additional computational cost, GEAT achieves the highest mode-discovery rate within the same wall-clock budget across all three benchmarks, as shown in the wall-clock mode-discovery curves available here. We clarify the computational cost of NAT and provide the corresponding wall-clock comparisons below.

Wall-clock comparison. We measure modes/min using a common wall-clock budget defined by TB's total training time, while reporting each method's full training time separately.

Modes/min ↑ / Training Time (min) ↓

MethodHyperGridBit SequencesEH
TB53.72 ± 4.64 / 21.0 ± 0.70.750 ± 0.004 / 45.4 ± 0.23.52 ± 0.24 / 1100.4 ± 12.6
SA-GFN61.61 ± 1.54 / 56.4 ± 1.60.426 ± 0.143 / 127.1 ± 1.43.08 ± 0.10 / 1603.7 ± 32.3
Adaptive Teacher60.56 ± 2.12 / 47.6 ± 4.20.611 ± 0.090 / 71.8 ± 0.73.12 ± 0.08 / 1741.2 ± 4.3
ACE60.22 ± 2.98 / 59.6 ± 3.20.569 ± 0.112 / 73.4 ± 1.12.87 ± 0.23 / 1725.8 ± 38.3
GEAT80.28 ± 7.45 / 57.7 ± 2.81.256 ± 0.045 / 94.1 ± 1.612.21 ± 1.37 / 2131.5 ± 210.6

Cost analysis. For trajectories of length TT, sampling one task and one exploration trajectory requires 2T2T forward-policy calls. Assuming kk forward reconstruction steps per proposal, NAT adds nknk forward calls and, when PBP_B is learned, nknk backward calls for proposal generation plus TT backward calls to generate the training trajectory after terminal selection. Thus, the policy-sampling cost ratio with versus without NAT is approximately 1+nk2T1+\frac{nk}{2T} for fixed or deterministic PBP_B and 32+nkT\frac{3}{2}+\frac{nk}{T} for learned PBP_B.


Additional Ablations

1. Epsilon-Greedy Sensitivity

We evaluated the sensitivity to the random-action probability ε\varepsilon on HyperGrid, Bit Sequence, and QM9. Across all tested values of ε\varepsilon, GEAT achieves the lowest average TV distance on HyperGrid (2D) and the highest average mode discovery on Bit Sequence (K=8) and Bio (QM9).

BenchmarkεTBSA-GFNTeacherACEGEAT
HyperGrid (TV↓)0*0.549 ± 0.0170.251 ± 0.0280.464 ± 0.0820.508 ± 0.0060.225 ± 0.087
0.010.708 ± 0.0880.353 ± 0.0540.493 ± 0.0800.533 ± 0.0950.293 ± 0.024
0.10.75954 ± 0.000130.75960 ± 0.000030.75962 ± 0.000160.75967 ± 0.000040.75919 ± 0.00013
Bit Sequence (Modes↑)0.0005*23.3 ± 0.635.7 ± 2.534.7 ± 3.134.7 ± 2.355.7 ± 2.5
0.0124.0 ± 1.035.0 ± 3.035.7 ± 1.534.7 ± 2.356.7 ± 1.5
0.125.0 ± 2.035.7 ± 2.336.7 ± 3.134.7 ± 4.552.7 ± 1.5
Bio (Modes↑)0637.0 ± 5.7712.7 ± 34.4745.3 ± 12.9739.3 ± 5.5756.3 ± 15.4
0.01635.0 ± 18.4716.0 ± 7.5726.7 ± 15.7723.3 ± 25.4745.3 ± 14.4
0.1*595.7 ± 19.7697.0 ± 5.6722.3 ± 0.6715.7 ± 16.3736.0 ± 9.5

Original ε settings used in the paper. Teacher denotes Adaptive Teacher.

2. Learned Backward Policy

GEAT maintains the lowest average TV distance under both fixed and learned backward policies. As specified in Appendix C.3, only HyperGrid uses a fixed uniform backward policy. We therefore additionally evaluated all five methods on HyperGrid with learned backward policies.

MethodFixed PB (TV)Learned PB (TV)
TB0.549 ± 0.0170.541 ± 0.037
SA-GFN0.251 ± 0.0280.498 ± 0.032
Adaptive Teacher0.464 ± 0.0820.564 ± 0.069
ACE0.508 ± 0.0060.510 ± 0.122
GEAT0.225 ± 0.0870.343 ± 0.135

All results are reported as mean ± sample standard deviation over three seeds.

Summary and Request for Reconsideration

We thank the reviewer again for the detailed review. We are especially grateful that the reviewer identified the reward-offset inconsistency in the sEH baselines; correcting it has strengthened the rigor of our comparison. We were also encouraged that the reviewer recognized the two ideas at the core of GEAT: the geometry-aware entropy formulation of coverage, and the distinction between regions that require further learning and regions that require new experience acquisition.

We summarize how each concern has been addressed:

  • Baseline implementation: With the offset removed consistently, GEAT still discovers roughly 3.8–3.9× more modes than Adaptive Teacher and ACE on sEH (18,711 vs. 4,870 / 4,757), while also achieving the highest Top-100 reward and the lowest FCS.
  • Computational cost: Under an equal wall-clock budget, GEAT achieves the highest mode-discovery rate on all three benchmarks (e.g., 12.21 vs. at most 3.52 modes/min on sEH). We also provide an explicit analysis of NAT's sampling overhead.
  • Robustness (ε and backward policy): GEAT remains the best-performing method across all tested ε values and under learned backward policies.
  • FCS: We agree that FCS alone cannot detect mode collapse. However, the collapse scenario is incompatible with GEAT discovering the largest number of high-reward modes in every environment. Together with exact TV distance on HyperGrid and the additional Spearman/Pearson correlations, the evidence consistently indicates that GEAT learns a distribution that is both broad and reward-consistent.

These results and discussions have been incorporated into the revised manuscript.

The reviewer characterized the soundness concerns as minor or fixable. We believe each has now been resolved through additional experiments, and none of them changes the conclusions of the paper. We would be grateful if the reviewer would consider revisiting the rating in light of these results, and we would be happy to address any remaining questions during the discussion period.

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

We sincerely appreciate your careful evaluation and constructive feedback. As the discussion period draws to a close, we summarize below the key improvements and core contributions made in response to your comments.

[Summary of Key Improvements]

We thank the reviewers for their constructive feedback, which helped us strengthen our empirical evaluation and broaden our analysis. The main improvements are summarized below:

  • Fairness of Mode Discovery Comparisons: We demonstrated that GEAT’s mode-discovery advantage is not solely attributable to its novelty kernel or its alignment with the evaluation metric. Its strong performance persists across different evaluation distances, supporting the effectiveness of the algorithm itself.
  • Computational Cost and Efficiency: Our computational analysis demonstrates that GEAT’s additional cost is outweighed by substantial gains in mode discovery, resulting in the highest mode-discovery rate under matched wall-clock budgets across all three benchmarks.
  • Robustness: We expanded our experiments to examine sensitivity to ε-greedy exploration and buffer eviction strategies, compatibility with learned backward policies and alternative training losses, and performance against baselines without replay. GEAT maintained strong performance across these settings, demonstrating its robustness to a broad range of experimental conditions.
  • Evaluation Rigor: We corrected reward inconsistencies in the sEH baselines and confirmed that our main conclusions remain unchanged. We also supplemented FCS with Spearman and Pearson correlations, providing additional evidence of consistency between rewards and GEAT’s learned sampling probabilities.

[Core Contributions]

The key methodological contribution of GEAT lies in the data-acquisition decision itself, and in the substantial empirical improvements that follow from it. GEAT changes what new experiences are collected, not merely how existing data are reused for learning. We believe meaningful methodological innovation need not arise from technical complexity; it can also come from identifying an overlooked design problem and addressing it through a simple, principled, and empirically effective solution.

We summarize our core contributions as follows:

  1. A Novel Principle for Efficient GFlowNet Exploration.
    • Before: Existing exploration methods often target regions where the task GFlowNet requires further learning, even when the corresponding experiences are already available through replay.
    • After: We are the first to exploit the synergy between exploration and replay in GFlowNets by explicitly separating new data acquisition from further learning. Exploration focuses on expanding terminal-state coverage, while replay supports further learning from previously collected experiences.
  2. Faster High-Reward Mode Discovery (NAT).
    • Before: LSGFN uses reward-based filtering to select local-search proposals, while SA-GFN uses novelty only as an intrinsic reward to train the exploratory policy. Neither explicitly prioritizes novelty in final terminal-state acquisition, potentially wasting exploration resources on already-explored high-reward regions.
    • After: GEAT introduces NAT, which prioritizes novelty in final terminal-state acquisition while maintaining reward-guided candidate generation. By selecting the most novel terminal state among local alternatives, NAT promotes faster discovery and broader coverage of high-reward modes.
  3. Broader Exploration in Representation Space (Geometric Novelty).
    • Before: Prior intrinsic-motivation-based GFlowNet methods use prediction-error-based intrinsic rewards (e.g., RND) without explicitly accounting for geometric similarity between distinct terminal states. Although Guo et al. (2021) introduced geometry-aware entropy exploration in RL, their approach relies on a learned auxiliary function.
    • After: We adapt geometry-aware Shannon entropy to GFlowNet exploration by directly computing its derivative at the empirical replay distribution. This is the first GFlowNet exploration method to derive a novelty criterion in this way, explicitly accounting for similarity between terminal-state representations. This promotes broader coverage and greater geometric diversity in exploration.
  4. Empirical Demonstration of Exploration Efficiency and Coverage.
    • Across synthetic, sequence generation, molecular, and peptide design benchmarks, GEAT achieves substantially faster high-reward mode discovery and broader mode coverage—a central measure of GFlowNet exploration—while maintaining competitive reward quality and distributional accuracy.

We believe our clarifications and additional experiments address the reviewers' main concerns. We deeply appreciate the valuable feedback on presentation and look forward to incorporating all suggestions in the final revision.

Thank you,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.