Pith. sign in

REVIEW 5 major objections 5 minor 34 references

Secrets of GFlowNets' Learning Behavior: A Theoretical Study

T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A theoretical study claims GFlowNet learning is governed by provable bounds on convergence, sample complexity, regularization, and robustness.

desk verdict The paper asks the right questions and states plausible conjectures, but the proofs do not support the advertised theorems, and the max-entropy claim for FM is false for the natural parameterization. read the letter →

arxiv 2505.02035 v1 pith:DYOBIGY6 submitted 2025-05-04 cs.LG stat.ML

Tianshu Yu This is my paper · ORCID
classification cs.LGstat.ML MSC 68Q3268T07
keywords GFlowNetsflowmatchingdetailedbalancetrajectorysamplecomplexityconvergencerateimplicitregularizationrobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the learning behavior of Generative Flow Networks can be characterized by provable bounds along four axes: convergence speed, sample complexity, implicit regularization, and robustness to reward noise. It shows that the Flow Matching objective converges at rate $O(1/\sqrt{T})$ while Detailed Balance converges at the slower $O(1/T^{1/3})$ because ratio-based gradients have higher variance. It bounds the number of trajectories needed for $\epsilon$-accurate flow estimation as $O(|S|L \log(|S|/\delta)/\epsilon^2)$, identifies FM with maximum-entropy implicit regularization and DB with KL-regularization, and shows that robustness to reward noise degrades as $R_{\min}^{-4}$. These results give formal grounding to design choices that were previously empirical.

What carries the argument

The central object is the flow-consistency linear system $Af = b$ from Proposition 1: each edge carries a nonnegative flow $f$, $A$ is the DAG's incidence matrix, and $b$ encodes source and terminal reward constraints. Theorem 1 shows that Flow Matching, Detailed Balance, and Trajectory Balance are different regularized versions of solving this underdetermined system. The convergence proofs hinge on the exponential parameterization $F_\theta(s \to s') = \exp(W_\theta(s,s'))$ with Lipschitz $W_\theta$ and bounded gradients; the sample-complexity argument depends on Assumption 3, which asserts that each length-$L$ trajectory provides $\Omega(L)$ independent constraints, allowing the edge dimension $|E|$ to be replaced by $|S|L$ and then reduced to $|S|$ by cancelling $L$.

What would settle it

Check a DAG whose state space is a single long path from $s_0$ to a terminal state. Every trajectory follows the same path, so each trajectory contributes at most one independent flow-conservation constraint regardless of its length $L$; there $\mathrm{rank}(A_\tau) = O(1)$, not $\Omega(L)$. Under that graph the cancellation producing $O(|S|L \log(|S|/\delta)/\epsilon^2)$ fails, and the claimed sample complexity would be contradicted.

Watch

Extended reading notes

Core claim

On the paper's own terms, GFlowNet training solves the underdetermined linear system $Af = b$ from Proposition 1, and the three popular objectives are regularized versions of this system. Under exponential parameterization $F_\theta(s \to s') = \exp(W_\theta(s,s'))$ and with a fixed backward policy, FM reaches a gradient-norm bound of $O(1/\sqrt{T})$ with a $\eta_t = \eta_0/\sqrt{t}$ schedule, while DB reaches $O(1/T^{1/3})$ with a more conservative $\eta_t = \eta_0/t^{2/3}$ schedule. For sample complexity, the number of trajectories required to reach $\epsilon$-accuracy with high probability is $O(|S|L \log(|S|/\delta)/\epsilon^2)$, reflecting state-space size, trajectory length, and target accuracy. The paper further claims that FM implicitly maximizes the entropy of the flow distribution, that DB implicitly minimizes the KL divergence between forward and backward joint transition distributions, and that reward noise degrades the TB objective, the terminal-state distribution, and sample complexity with the dependence scaling as $R_{\min}^{-4}$.

Load-bearing premise

The load-bearing premise is Assumption 3 (Information Content Assumption): every trajectory of length $L$ supplies $\Omega(L)$ independent constraints on the flow system, which is what lets the sample complexity bound cancel the trajectory length and depend on $|S|$ instead of $|E|$.

Editorial extensions

If this is right

  • FM's faster convergence rate suggests that when a fixed backward policy is acceptable, Flow Matching is the safer objective for tight iteration budgets.
  • The sample complexity bound identifies $|S|$, $L$, and $\epsilon$ as the key levers: shortening trajectories or shrinking the state space reduces required samples, and the dependence on accuracy is quadratic.
  • The implicit max-entropy regularization of FM implies trained samplers are biased toward diversity, which is beneficial for exploration-heavy generative tasks.
  • The $R_{\min}^{-4}$ robustness scaling implies that raising the minimum reward, for example by thresholding, produces a disproportionately large improvement in noise resilience.
  • Order-dependence of training, formalized in Proposition 2, explains why GFlowNet optimization is path-dependent even when the same set of trajectories is reused in a different order.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct consequence the author does not spell out: because the DB rate is bottlenecked by variance from small backward probabilities $P_B(s|s')$, developing low-variance estimators of the ratio term is the most direct practical route to closing the gap with FM; the paper points to existing gradient-estimation work as a starting place.
  • Theorem 5's clean $|S|$-dependence only holds when Assumption 3 is satisfied; for graphs with bottlenecks or long chains the effective complexity may be closer to a coverage problem for rare states, suggesting per-state visitation-aware sampling as a testable improvement.
  • The max-entropy characterization of FM suggests a concrete diagnostic: train FM on a reward landscape with a known unique maximum and measure whether the sampled terminal-state distribution spreads beyond the reward ratio, as entropy regularization predicts.
  • If the $R_{\min}^{-4}$ dependence holds in practice, reward normalization plus a small positive floor should be a near-universal preprocessing step for GFlowNet training under noisy reward estimators.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes a unified theoretical analysis of GFlowNet training under four headings: convergence rates of the Flow Matching (FM) and Detailed Balance (DB) objectives, sample complexity, implicit regularization, and robustness to reward noise. It formulates the FM, DB, and Trajectory Balance (TB) objectives as regularized versions of an underdetermined linear system (Proposition 1), states convergence rates O(1/sqrt(T)) for FM and O(1/T^{1/3}) for DB, derives sample-complexity bounds of the form O(|S|L log(|S|/delta)/epsilon^2), claims that FM implicitly maximizes flow entropy while DB induces KL regularization, and bounds the effect of reward noise by terms scaling as R_min^{-4}. Proofs are collected in Appendices A through D.

Significance. The scope is ambitious and, if the results were correct, the paper would be a useful reference for comparing GFlowNet training objectives. The linear-system viewpoint in Proposition 1 is a clean way to expose the non-identifiability of flow solutions, and the importance-sampling discrepancy in Theorem 4 is a sensible quantity for reasoning about off-policy training. However, the verification in the appendices does not support the advertised claims. The implicit-regularization theorem is a non sequitur, the convergence proofs contain a schedule mismatch and a dropped logarithmic factor, the sample-complexity proof introduces the trajectory-length factor by a heuristic step under strong assumptions, and one robustness proof omits a zero-mean condition. No code or machine-checked proofs accompany the paper. The paper therefore does not currently establish its central conclusions.

major comments (5)
  1. [Section 3.1, Theorem 2; Appendix A] The stated learning rate does not match the one used in the proof: Theorem 2 says eta_t = eta_0 sqrt(t), but the proof's summations assume eta_t = eta_0 / sqrt(t). Moreover, the bound derived in Eqs. (20)-(22) contains (1 + ln T)/sqrt(T), and Eq. (23) silently replaces this by C/sqrt(T) with a constant C that does not correspond to the expression actually bounded. At best the argument establishes O(log T / sqrt(T)) for a decreasing schedule. Theorem 3 also assumes, through K(theta) = sup 1/P_B(s|s')^2 and 'K_max finite', a lower bound on the backward probabilities that is not stated in the theorem or in its 'similar conditions'. The convergence comparison in Section 3.1 is therefore not established as written.
  2. [Section 3.2, Theorem 4; Appendix B] The proof contains two contradictory importance-sampling derivations. The gradient estimator uses weights P_target/P_sample, but the Hoeffding 'general bound' defines X_i = P_sample/P_target * g and then bounds |X_i| <= G D; since D = max P_target/P_sample, the inverse ratio is not uniformly bounded by D. That derivation gives N = O(D^2 log(1/delta)/epsilon^2), while the Bernstein derivation gives N = O(D log(1/delta)/epsilon^2), and the theorem states the latter. As written, the proof does not provide a coherent derivation of the claimed bound.
  3. [Section 3.3, Theorem 6; Appendix C] The proof maximizes entropy over all flows satisfying the flow constraints and observes that the max-entropy flow has exponential form; it never shows that the FM objective or its optimization dynamics selects that flow. By Proposition 1, L_FM vanishes on the entire feasible affine set {f >= 0 : Af = b}; with per-edge exponential parameters every feasible positive flow is representable, so the max-entropy solution is one of many global minimizers and is not singled out by the loss. For example, in a DAG with two parallel paths from s0 to a common terminal, the 90/10 and 50/50 splits of flow between the two paths are both zero-loss and stationary under FM. The claim that FM 'implicitly maximizes' flow entropy is therefore unsupported and is false for the natural per-edge exponential parameterization. The stationarity calculation also has a sign error: differentiating the Lagrangian yields F(s -> s') proportional to exp(phi_s - phi_{s'}), not exp(phi_{s'} - phi_s).
  4. [Section 3.2, Theorem 5; Appendix B] The claimed sample complexity N = O(|S|L log(|S|/delta)/epsilon^2) is not a consequence of the displayed proof. Equation (84) gives N_accurate = O(|S| log(1/delta)/epsilon^2), which has no dependence on L; Eq. (85) takes a maximum of the coverage and accuracy terms, and the factor L appears only through an informal statement that per-step error must be O(epsilon/sqrt(L)) because of error accumulation. That step is not derived from the preceding bounds. In addition, the proof relies on Assumption 1 (min_{t,s} pi_t(s) >= c/|S|) and Assumption 3, which are described as 'quite mild' but are not justified; uniform state visitation during early training is a strong condition, and the rank assumption is not connected to any verifiable graph structure. The advertised |S|L dependence is therefore a heuristic rather than an established theorem.
  5. [Section 3.4, Theorem 7; Appendix D] Theorem 7 does not assume that the reward noise has zero mean, but the proof's Taylor expansion drops the first-order term g'(R(s_T)) epsilon(s_T). The justification that g'(R) is approximately zero relies on F_theta being 'well-trained' on the true rewards, which is an additional condition not stated in the theorem. For a model that is not exactly at the optimum and noise with nonzero mean, the omitted term is of order 1/R_min and can dominate the claimed O(R_min^{-4}) bound. The statement needs either a zero-mean noise assumption or a quantitative closeness-to-optimality assumption.
minor comments (5)
  1. [Throughout] There are numerous typos, including 'have seeked' (should be 'have sought'), 'parituclar' (should be 'particular'), 'discrepency' (should be 'discrepancy'), 'Rearraning' (should be 'Rearranging'), 'trajetories' (should be 'trajectories'), and 'positiv semidefinite' (should be 'positive semidefinite').
  2. [Appendix A, proof of Theorem 2] The justification 'since the minimum operation is convex' is incorrect: the pointwise minimum of random variables is concave. The displayed inequality E[min_t X_t] <= min_t E[X_t] is nevertheless the correct one for concave functions, so the step can be repaired by fixing the justification.
  3. [Appendix B, proof of Theorem 4] Equation (56) defines X_i with the ratio P_sample/P_target, whereas the estimator in Eq. (48) uses P_target/P_sample; the notation should be made consistent throughout.
  4. [Appendix D, proof of Theorem 8] The text says 'E[epsilon(s_T)] = sigma^2'; this should be E[epsilon(s_T)^2] = sigma^2 with E[epsilon(s_T)] = 0. The theorem statement assumes zero-mean noise, but the proof should not conflate the mean with the second moment.
  5. [Section 4, Limitations] The limitations section acknowledges that the implicit-regularization results 'particularly depend on specific neural network architectures,' which is in tension with the unqualified wording of Theorem 6; the theorem statement and Section 3.3 should carry the qualifications acknowledged in Section 4.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 6's max-entropy claim reduces to the exponential-parameterization ansatz by construction; the FM loss itself never selects the max-entropy feasible flow.

  1. self definitional [Section 3.3, Theorem 6 and Appendix C, Eqs. (102)-(105)]
    "To find the maximum entropy flow distribution satisfying the flow consistency constraints, we set the gradient... F (s → s′) = C · exp(W (s, s′)). In GFlowNet training with the FM objective, we parameterize the flow function as: F (s → s′) = exp(Wθ(s, s′)). This parameterization exactly matches the form of the maximum entropy solution derived above... Therefore, when GFlowNets are trained with the FM objective using this exponential parameterization, they implicitly seek the maximum entropy flow distribution that satisfies the flow constraints."

    The proof never uses the FM loss. By Proposition 1, L_FM vanishes on every nonnegative flow satisfying Af=b, so the objective has no term selecting the maximum-entropy member. The proof maximizes entropy as a separate constrained problem (Eqs. 102-104), obtains the exponential form, and then concludes FM 'implicitly' does the same solely because Fθ=exp(Wθ) has the same form. The claimed prediction is therefore equivalent to the exponential-parameterization ansatz: the max-entropy property is imported by construction, not derived from FM's learning dynamics. No argument shows FM-converged parameters reach the max-entropy feasible flow; Proposition 2's order-dependent weights even suggest the selected flow depends on sampling order.

full rationale

Most of the paper is self-contained and not circular. The FM and DB convergence-rate proofs (Theorems 2-3) are standard smooth-SGD arguments with bounded-gradient/variance assumptions; the sample-complexity theorem (Theorem 5) is a statistical-learning bound, albeit one that leans on Assumption 3 (rank(Aτ)=Ω(L)) to cancel L and on Assumption 4 to re-introduce it — strong assumptions, but not circular. The robustness theorems (7-9) are Taylor-expansion and importance-sampling bounds conditional on small noise. There is no load-bearing self-citation: the paper cites the GFlowNet foundations and related works by other groups, not a uniqueness theorem of the author's. The central circular step is Theorem 6: its conclusion is made true by construction because the exponential parameterization is the same functional form as the maximum-entropy solution, while the FM objective itself is indifferent among all feasible flows. The paper's own Limitations section concedes that the implicit-regularization results 'particularly depend on specific neural network architectures,' which supports reading Theorem 6 as an attribute of the ansatz rather than of FM learning dynamics.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

The central claims rest on a set of assumptions that are asserted to be 'quite mild in practice' but are not verified. Assumption 3 is the weakest; it is load-bearing and can fail in simple graphs. No free parameters are fitted to data, and no new entities are introduced.

assumptions (7)
  • domain assumption Assumption 1 (State Visitation Distribution Assumption): min_{s∈S,t} π_t(s) ≥ c/|S| for some c>0.
    Used in Theorem 5 to bound coverage; strong for a learned policy that concentrates on high-reward states.
  • domain assumption Assumption 2 (Error Independence Assumption): Cov(ε_i, ε_j) ≤ ρ^{|i-j|} σ^2.
    Used for error accumulation along trajectories in Theorem 5.
  • domain assumption Assumption 3 (Information Content Assumption): rank(Aτ)=Ω(L) for each trajectory τ.
    Load-bearing for Theorem 5; false for path-like DAGs, and not justified in the paper.
  • domain assumption Assumption 4 (Error Propagation Assumption): ‖F̂τ-F*τ‖₂ = O(√L ε_edge).
    Used to turn per-step errors into the L factor in Theorem 5's complexity.
  • domain assumption Bounded flows and transition probabilities: m ≤ F_θ(s), F*(s) ≤ M, and P_F(s'|s), P_B(s|s') bounded away from 0.
    Used in Proposition 3 and Theorem 3 to control variance; may fail with learned neural policies.
  • domain assumption Exponential parameterization Fθ(s→s') = exp(Wθ(s,s')) with Wθ l-Lipschitz and bounded gradients.
    Central to Theorems 2, 3, and 6; the results may not extend to other parameterizations.
  • domain assumption Reward lower bound R_min > 0 and zero-mean independent reward noise with variance σ².
    Used in Theorems 7-9; the bounds diverge as R_min→0.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Secrets of GFlowNets' Learning Behavior: A Theoretical Study." pith.science (2026). https://pith.science/paper/DYOBIGY6

@misc{pith2026250502035,
  author       = {Pith},
  title        = {Pith review of: Secrets of GFlowNets' Learning Behavior: A Theoretical Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DYOBIGY6}},
  note         = {Machine review of arXiv:2505.02035}
}
read the original abstract

Generative Flow Networks (GFlowNets) have emerged as a powerful paradigm for generating composite structures, demonstrating considerable promise across diverse applications. While substantial progress has been made in exploring their modeling validity and connections to other generative frameworks, the theoretical understanding of their learning behavior remains largely uncharted. In this work, we present a rigorous theoretical investigation of GFlowNets' learning behavior, focusing on four fundamental dimensions: convergence, sample complexity, implicit regularization, and robustness. By analyzing these aspects, we seek to elucidate the intricate mechanisms underlying GFlowNet's learning dynamics, shedding light on its strengths and limitations. Our findings contribute to a deeper understanding of the factors influencing GFlowNet performance and provide insights into principled guidelines for their effective design and deployment. This study not only bridges a critical gap in the theoretical landscape of GFlowNets but also lays the foundation for their evolution as a reliable and interpretable framework for generative modeling. Through this, we aspire to advance the theoretical frontiers of GFlowNets and catalyze their broader adoption in the AI community.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 31 canonical work pages

  1. [1]

    Rademacher and ga ussian complexities: Risk bounds and structural results

    Peter L Bartlett and Shahar Mendelson. Rademacher and ga ussian complexities: Risk bounds and structural results. Journal of Machine Learning Research , 3(Nov):463–482, 2002

    work page 2002

  2. [2]

    Gflownet foundations

    Y oshua Bengio, Salem Lahlou, Tristan Deleu, Edward J Hu, Mo Tiwari, and Emmanuel Ben- gio. Gflownet foundations. Journal of Machine Learning Research , 24(210):1–55, 2023

    work page 2023

  3. [3]

    On a modification of chebyshev’s inequ ality and of the error formula of laplace

    Sergei Bernstein. On a modification of chebyshev’s inequ ality and of the error formula of laplace. Ann. Sci. Inst. Sav. Ukraine, Sect. Math , 1(4):38–49, 1924

    work page 1924

  4. [4]

    Fas t and accurate stochastic gradient estimation

    Beidi Chen, Yingchen Xu, and Anshumali Shrivastava. Fas t and accurate stochastic gradient estimation. Advances in Neural Information Processing Systems , 32, 2019

    work page 2019

  5. [5]

    Generative flow network s: a markov chain perspective

    Tristan Deleu and Y oshua Bengio. Generative flow network s: a markov chain perspective. arXiv preprint arXiv:2307.01422 , 2023

    arXiv 2023

  6. [6]

    Flows in networks

    Lester Randolph Ford and Delbert Ray Fulkerson. Flows in networks . Princeton university press, 2015

    work page 2015

  7. [7]

    Generative adve rsarial nets

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Y oshua Bengio. Generative adve rsarial nets. Advances in neural information processing systems, 27, 2014

    work page 2014

  8. [8]

    A distribution-free theory of nonparametric regression

    László Györfi, Michael Kohler, Adam Krzyzak, and Harro Wa lk. A distribution-free theory of nonparametric regression. Springer Science & Business Media, 2006

    work page 2006

Show all 34 references
  1. [9]

    Probability inequalities for sums o f bounded random variables

    Wassily Hoeffding. Probability inequalities for sums o f bounded random variables. The col- lected works of W assily Hoeffding, pages 409–426, 1994

    1994

  2. [10]

    Biological sequence design with gflownets

    Moksh Jain, Emmanuel Bengio, Alex Hernandez-Garcia, J arrid Rector-Brooks, Bonaven- ture FP Dossou, Chanakya Ajit Ekbote, Jie Fu, Tianyu Zhang, M ichael Kilgour, Dinghuai Zhang, et al. Biological sequence design with gflownets. In International Conference on Machine Learning, p...

    2022

  3. [11]

    Auto-encoding varia tional bayes, 2022

    Diederik P Kingma and Max Welling. Auto-encoding varia tional bayes, 2022

    2022

  4. [12]

    Rgfn: Synthesizable molecular generation using gflownets

    Michał Koziarski, Andrei Rekesh, Dmytro Shevchuk, Alm er van der Sloot, Piotr Gai ´nski, Y oshua Bengio, Chenghao Liu, Mike Tyers, and Robert Batey. Rgfn: Synthesizable molecular generation using gflownets. Advances in Neural Information Processing Systems , 37:46908– 46955, 2024

    2024

  5. [13]

    On generalization for generative flow networks

    Anas Krichel, Nikolay Malkin, Salem Lahlou, and Y oshua Bengio. On generalization for generative flow networks. arXiv preprint arXiv:2407.03105 , 2024

    2024 arXiv

  6. [14]

    A theory of con- tinuous generative flow networks

    Salem Lahlou, Tristan Deleu, Pablo Lemos, Dinghuai Zha ng, Alexandra V olokhova, Alex Hernández-Garcıa, Léna Néhale Ezzine, Y oshua Bengio, and Nikolay Malkin. A theory of con- tinuous generative flow networks. In International Conference on Machine Learning , pages 18269–18300...

    2023

  7. [15]

    Learning gflownets from partial episodes for improved convergence and stabili ty

    Kanika Madan, Jarrid Rector-Brooks, Maksym Korablyov , Emmanuel Bengio, Moksh Jain, Andrei Cristian Nica, Tom Bosc, Y oshua Bengio, and Nikolay M alkin. Learning gflownets from partial episodes for improved convergence and stabili ty. In International Conference on Machine Lear...

    2023

  8. [16]

    G flownets for causal discovery: an overview

    Dragos Cristian Manta, Edward J Hu, and Y oshua Bengio. G flownets for causal discovery: an overview. In ICML 2023 W orkshop on Structured Probabilistic Inference & Generative Modeling, 2023

    2023

  9. [17]

    Pac-bayes un-expected bernstein inequality

    Zakaria Mhammedi, Peter Grünwald, and Benjamin Guedj. Pac-bayes un-expected bernstein inequality. Advances in Neural Information Processing Systems , 32, 2019

    2019

  10. [18]

    Monte carlo gradient estimation in machine learning

    Shakir Mohamed, Mihaela Rosca, Michael Figurnov, and A ndriy Mnih. Monte carlo gradient estimation in machine learning. Journal of Machine Learning Research , 21(132):1–62, 2020

    2020

  11. [19]

    Belief prop agation, robust reconstruction and optimal recovery of block models

    Elchanan Mossel, Joe Neeman, and Allan Sly. Belief prop agation, robust reconstruction and optimal recovery of block models. In Conference on Learning Theory, pages 356–370. PMLR, 2014

    2014

  12. [20]

    Causal inference in gene regulatory networks with gflownet: Toward s scalability in large systems

    Trang Nguyen, Alexander Tong, Kanika Madan, Y oshua Ben gio, and Dianbo Liu. Causal inference in gene regulatory networks with gflownet: Toward s scalability in large systems. arXiv preprint arXiv:2310.03579 , 2023. 10

    2023 arXiv

  13. [21]

    Hierarchical gflownet for crystal struct ure generation

    Tri Minh Nguyen, Sherif Abdulkader Tawfik, Truyen Tran, Sunil Gupta, Santu Rana, and Svetha V enkatesh. Hierarchical gflownet for crystal struct ure generation. In AI for Accelerated Materials Design-NeurIPS 2023 W orkshop, 2023

    2023

  14. [22]

    Better training of gflownets with local credit and incomplete trajectories

    Ling Pan, Nikolay Malkin, Dinghuai Zhang, and Y oshua Be ngio. Better training of gflownets with local credit and incomplete trajectories. In International Conference on Machine Learn- ing, pages 26878–26890. PMLR, 2023

    2023

  15. [23]

    Towards understanding and improvi ng gflownet training

    Max W Shen, Emmanuel Bengio, Ehsan Hajiramezanali, And reas Loukas, Kyunghyun Cho, and Tommaso Biancalani. Towards understanding and improvi ng gflownet training. In Inter- national conference on machine learning , pages 30956–30975. PMLR, 2023

    2023

  16. [24]

    Gradient es- timation with discrete stein operators

    Jiaxin Shi, Y uhao Zhou, Jessica Hwang, Michalis Titsia s, and Lester Mackey. Gradient es- timation with discrete stein operators. Advances in neural information processing systems , 35:25829–25841, 2022

    2022

  17. [25]

    When do gflownets learn the right distribution? In The Thirteenth International Conference on Learning Represen tations, 2025

    Tiago Silva, Rodrigo Barreto Alves, Eliezer de Souza da Silva, Amauri H Souza, Vikas Garg, Samuel Kaski, and Diego Mesquita. When do gflownets learn the right distribution? In The Thirteenth International Conference on Learning Represen tations, 2025

    2025

  18. [26]

    Analyz ing gflownets: Stability, ex- pressiveness, and assessment

    Tiago Silva, Eliezer de Souza da Silva, Rodrigo Barreto Alves, Luiz Max Carvalho, Amauri H Souza, Samuel Kaski, Vikas Garg, and Diego Mesquita. Analyz ing gflownets: Stability, ex- pressiveness, and assessment. In ICML 2024 W orkshop on Structured Probabilistic Inference & Gene...

    2024

  19. [27]

    On divergence measures for training gflownets

    Tiago Silva, Eliezer de Souza da Silva, and Diego Mesqui ta. On divergence measures for training gflownets. Advances in Neural Information Processing Systems , 37:75883–75913, 2024

    2024

  20. [28]

    Score-based generative modeling through stocha stic differential equations

    Y ang Song, Jascha Sohl-Dickstein, Diederik P Kingma, A bhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stocha stic differential equations. arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  21. [29]

    Let the flows tell: Solving graph combinatorial problem s with gflownets

    Dinghuai Zhang, Hanjun Dai, Nikolay Malkin, Aaron C Cou rville, Y oshua Bengio, and Ling Pan. Let the flows tell: Solving graph combinatorial problem s with gflownets. Advances in neural information processing systems, 36:11952–11969, 2023

    2023

  22. [30]

    Generative flow networks for discrete probab ilistic modeling

    Dinghuai Zhang, Nikolay Malkin, Zhen Liu, Alexandra V o lokhova, Aaron Courville, and Y oshua Bengio. Generative flow networks for discrete probab ilistic modeling. In International Conference on Machine Learning , pages 26412–26428. PMLR, 2022

    2022

  23. [31]

    Sample- efficient multi-objective molecular optimization with gflo wnets

    Yiheng Zhu, Jialu Wu, Chaowen Hu, Jiahuan Y an, Tingjun H ou, Jian Wu, et al. Sample- efficient multi-objective molecular optimization with gflo wnets. Advances in Neural Informa- tion Processing Systems, 36:79667–79684, 2023. 11 A Proofs of Convergence Proof of Theorem 2. For S...

    2023

  24. [32]

    Error accumulation: Errors in flow estimates can compound along a trajectory

  25. [33]

    Credit assignment: It becomes harder to attribute reward to specific transitio ns

  26. [34]

    To model error accumulation, we can use results from Markov c hain analysis

    Exploration efficiency : Longer trajectories may explore fewer unique states per un it of computation. To model error accumulation, we can use results from Markov c hain analysis. According to As- sumption 2, for a trajectory of length L, the error in the flow estimate can grow ...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.