{"id":"1eb992cf-eb0f-4fce-bff6-1f4f2d451f7a","arxiv_id":"2501.05498","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"GFlowNets cast sampling from an unnormalized discrete distribution as a flow-conservation problem, and the thesis applies them to approximate Bayesian posteriors over DAGs.","lead":"This thesis develops Generative Flow Networks, a family of models that sample from complex distributions over discrete objects like graphs by constructing them one piece at a time. It then uses them for Bayesian structure learning, approximating the full distribution over causal graph structures instead of picking one graph.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The exact-flow theory is sound, but the central approximation claim lacks error bounds: the trained GFlowNet is only guaranteed correct in the limit of perfect flow matching, and the thesis provides no finite-capacity or finite-sample guarantee.","rationale":"The reader's weakest assumption precisely identifies the load-bearing concern: the exact theory (Prop. 3.1.8 with the flow matching condition) is correct, but the application claim relies on the trained parametric flow being close to a valid flow, without any quantitative bound. I agree with this assessment. The thesis is a compilation of previously peer-reviewed work, and the theoretical core is rigorous; the main risk is the gap between exact flow-matching and approximate training. The proposed test would settle whether the approximation gap is a fundamental limitation of the parametric family or only an empirical scaling issue. Therefore the reader's CONDITIONAL verdict remains appropriate.","tokens_in":55829,"tokens_out":12772,"duration_ms":125645,"concrete_test":"For a DAG space with d≤5 nodes, enumerate all states and train a GFlowNet with the exact flow-matching loss computed exhaustively over every state (full support), with a flexible architecture and many steps until the loss plateaus. Compare the induced terminating distribution (by dynamic programming) to the exact posterior R/Z. If the KL does not approach zero, the parametric family cannot represent an exact valid flow, so the approximation gap is inherent rather than a finite-sample artifact. If it does approach zero, repeat the comparison on states not seen during training to test the coverage assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (GFlowNets can approximate the posterior over DAGs) rests on the exact result that a flow satisfying the flow matching condition (3.2.1) with boundary F(x→⊥)=R(x) induces the terminating distribution R(x)/Z (Prop. 3.1.8). The theory does not carry over to the trained parametric flow. The training objective (Section 4.2.1) minimizes squared flow-matching residuals sampled from a training distribution; its global minimum is zero only on the support of that distribution. If the training distribution does not cover the full state space, or if the parametric family does not contain any function satisfying (3.2.1) and the boundary condition, the minimizing flow is not a valid flow and the induced terminating distribution may be arbitrarily far from R. The thesis acknowledges this gap in Sections 7.4.3, 7.4.4 and 8.3.5, but provides no quantitative bound. Thus the practical claim is conditional on an unverified representation and optimization assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The thesis develops the theory of Generative Flow Networks (GFlowNets) as a class of sequential, amortized generative models for sampling from unnormalized distributions over discrete and compositional objects, and applies them to Bayesian structure learning. Part I introduces flow networks, defines Markovian flows, and proves that exact flows satisfying the flow matching condition (3.2.1), detailed balance (3.2.21), trajectory balance (3.2.31), or sub-trajectory balance (3.2.49) conditions are equivalent characterizations of a unique Markovian flow; Proposition 3.1.8 shows that the terminating distribution of such a flow is the normalized terminating edge flow. The thesis also connects GFlowNets to maximum entropy reinforcement learning and variational inference, and extends them to general state spaces. Part II constructs GFlowNets over DAGs, approximates the Bayesian posterior over structures (DAG-GFlowNet) and over structures and parameters jointly (JSP-GFN), and validates the approximations against exact posteriors computed by dynamic programming on small graphs and against MCMC baselines.","tokens_in":56046,"tokens_out":5670,"duration_ms":53771,"significance":"If the central claims hold, GFlowNets offer a principled and scalable alternative to MCMC for posterior inference over discrete structures, with clear connections to variational inference and max-entropy RL. The manuscript's strengths include the careful proofs of the flow conditions, the explicit empirical check against exact posterior enumeration for small graphs (Section 7.5), the integration of interventional and observational data (Section 7.7), and the candid limitation sections (7.4.3, 7.4.4, 8.3.5). The main caveat is that the exact-flow theory does not by itself justify the trained parametric approximation; the degree to which this gap is quantified or appropriately labeled determines whether the application claims are fully supported.","major_comments":[{"comment":"The central practical claim is that a trained GFlowNet samples approximately from the target reward distribution. Proposition 3.1.8 and Theorem 3.2.1 establish this only for exact Markovian flows satisfying the flow matching condition (3.2.1) and the boundary reward F(x→⊥)=R(x). The training losses in Section 4.2.1 minimize sampled residuals; their global optimum is zero at best on the support of the training distribution, and nothing in the presentation ensures that the parametric family contains a valid flow or that the training distribution covers the state space. The limitation sections 7.4.3, 7.4.4, and 8.3.5 acknowledge this gap, but provide no quantitative bound. Please either supply an error bound (for example, total variation distance in terms of the expected flow-matching residual and a coverage or exploration term) or explicitly reframe the practical claims as heuristic approximations.","section":"§4.2.1; Prop. 3.1.8; Thm. 3.2.1"},{"comment":"The chapter introduction promises 'novel convergence guarantees of the approximate distribution towards the target' and Section 4.2.3 is titled 'Convergence guarantees'. The precise assumptions of these guarantees should be stated clearly. If they require exact minimization of the flow-matching loss over a hypothesis class that contains a valid flow, or access to the full training distribution, then they do not cover the finite-capacity, finite-sample training used in Chapters 7 and 8, and the application chapters remain conditional on an unquantified approximation error.","section":"§4.2.3"},{"comment":"The exact-posterior comparison in Section 7.5 is limited to small graphs, while the medium-scale experiments (d=20) rely on metrics whose limitations are themselves acknowledged in Section 7.6.3. The thesis should either provide a quantitative calibration of those metrics (for example, by comparing expected SHD and AUROC against exact posteriors on graphs small enough for enumeration) or temper the scaling claims accordingly, so that the reader can distinguish validated accuracy from heuristic evaluation.","section":"§7.5, §7.6.3"}],"minor_comments":[{"comment":"The French text reads 'lesréseaux à flots génératifs'; there is a missing space between 'les' and 'réseaux'.","section":"Résumé, p. iii"},{"comment":"The caption uses 'quicunx'; the correct spelling is 'quincunx'.","section":"Figure 2.2 caption"},{"comment":"The glossary entry for 'P⊗T F' has inconsistent sub- and superscript rendering; please ensure it matches the notation used in Appendix A.","section":"Glossary, p. xiii"},{"comment":"Since Sections 7.4.3, 7.4.4, and 8.3.5 each discuss limitations in different places, a consolidated limitations paragraph in the Conclusion that relates these caveats to the theoretical guarantees of Part I would help the reader assess the scope of the contributions.","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"This is a well-written monograph that compiles the author's published work and adds some novel unifying results. My main reservation is the gap between exact-flow theory and parametric training; the author is unusually candid about this in the limitation sections, which I appreciate. I believe the manuscript can be revised to make the status of the practical claims precise, either by adding quantitative approximation guarantees or by explicitly labeling the application results as heuristic; the theoretical core is sound and worth publishing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. This is a PhD thesis, and it reads like one: most of the technical content comes from the author's prior peer-reviewed papers (Bengio et al. 2023, Malkin et al. 2022, Deleu et al. 2022/2023), reorganized into a coherent narrative with unified notation. What's actually new is modest—the MaxEnt RL equivalence in Chapter 5, the Markov chain perspective in Chapter 6, and a few counterexamples—but the synthesis is genuinely useful. The flow-matching theorems in Chapter 3 are carefully proved, and the empirical sections in Chapter 7 compare against exact posterior enumeration on small DAGs, which is the right way to validate an approximate sampler.\n\nThe soft spots are real but not disqualifying. The central approximation claim—that a trained GFlowNet samples from the target posterior—has no quantitative error bound. The exact flow-matching conditions guarantee the result only for exact flows; a trained parametric flow minimizes a sampled objective over a training distribution, and if that distribution doesn't cover the state space, or the function class can't represent any valid flow, the induced terminating distribution could be arbitrarily far from the target. The thesis explicitly acknowledges this in Sections 7.4.3, 7.4.4, and 8.3.5, but doesn't bound the gap. That's a known limitation of amortized inference in general, so I wouldn't call it fatal, but it does mean the practical claims are conditional on representation and optimization success. Second, there's no code or data included, which makes the empirical comparisons harder to reproduce. Third, as a research preprint, the new material isn't clearly delineated from the published work; the Prologue mentions novel results, but I had to dig to identify them.\n\nI'd send this to serious refereeing. It deserves a close review as a monograph, especially for someone who wants the GFlowNet theory laid out cleanly. But the verdict should be about the synthesis and pedagogical value, not about new results. If I were editing, I'd ask the author to either provide code and data artifacts or explicitly list what's new beyond the cited papers.","headline":"A rigorous thesis-length synthesis of GFlowNets; the theory holds up, but the practical claims still depend on the usual amortized-inference gap and the genuinely new content is thin.","tokens_in":56599,"tokens_out":2230,"would_cite":true,"duration_ms":22129,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The thesis establishes that sampling from an intractable discrete distribution reduces to finding edge flows that satisfy a conservation law, and that this is enough to approximate Bayesian posteriors over causal graph structures.","keywords":["generative flow networks","flow matching","Bayesian structure learning","Bayesian networks","directed acyclic graphs","variational inference","maximum entropy reinforcement learning","sequential decision making"],"falsifier":"On a small, fully enumerable state space (for example, all DAGs on four nodes), compute the exact terminating-state distribution $P_F^\\top(x)$ by dynamic programming, train a GFlowNet with a flow matching loss, and test whether the learned distribution approaches $R(x)/Z$ as training time and capacity grow; a persistent gap would falsify the claim that flow matching suffices for sampling.","tokens_in":55594,"feed_emoji":"🌊","tokens_out":8958,"duration_ms":77332,"temperature":0.7,"pith_summary":"Generative Flow Networks (GFlowNets) are a class of probabilistic models that sample discrete, compositional objects by treating generation as a sequence of decisions. The thesis establishes that if a flow on a directed acyclic graph satisfies the flow matching condition — the sum of incoming edge flows equals the sum of outgoing edge flows at every non-initial state — and if the flow through each terminating edge is set to the target weight $R(x)$, then the induced forward policy samples $x$ with probability proportional to $R(x)$. This turns the problem of sampling from an intractable distribution, such as a Bayesian posterior over DAGs, into a scalable optimization problem over edge flows. The second half of the thesis applies this idea to Bayesian structure learning, approximating posteriors over Bayesian network structures and their parameters from observational and experimental data.","feed_headline":"A flow-matching law turns unnormalized rewards into exact samplers","feed_subtitle":"GFlowNets build graphs edge by edge so that sampled DAGs follow the Bayesian posterior, not path counts.","key_machinery":"The load-bearing object is a Markovian flow network on a pointed DAG: a non-negative edge flow $F(s\\to s')$ with a single source $s_0$ and a single sink $\\bot$. The flow matching condition enforces conservation at every non-initial state; the boundary condition $F(x\\to\\bot)=R(x)$ fixes the terminating flows. The identity that carries the argument is the induced forward policy $P_F^\\star(s'|s)=F(s\\to s')/F(s)$, which, by Proposition 3.1.8, samples terminating states exactly with probability $F(x\\to\\bot)/Z^\\star$. Detailed balance, trajectory balance, and sub-trajectory balance are alternative re-expressions of the same conservation law at transition, full-trajectory, and partial-trajectory granularity, each usable as a training loss.","core_discovery":"The central claim is that a non-negative reward function $R$ on a finite set of compositional objects can be realized as the terminating flow of a Markovian flow network. Setting $F(x\\to\\bot)=R(x)$ on each terminating edge and choosing edge flows satisfying the flow matching condition $\\sum_{s\\in\\mathrm{Pa}(s')}F(s\\to s')=\\sum_{s''\\in\\mathrm{Ch}(s')}F(s'\\to s'')$ at every non-initial state yields a forward policy $P_F(s'|s)=F(s\\to s')/F(s)$ whose terminating-state distribution is exactly $R(x)/\\sum_x R(x)$ (Proposition 3.1.8). The thesis further claims that this conservation law can be turned into a family of learning objectives — flow matching, detailed balance, trajectory balance, and sub-trajectory balance — and that training a neural-network policy against these losses yields an amortized sampler that approximates the target distribution, with convergence guarantees for exact flows. In structure learning, this yields DAG-GFlowNet for the marginal posterior over graphs and JSP-GFN for the joint posterior over graph and parameters, and the Markov-chain view extends the construction to continuous state spaces.","pith_inferences":["A reader could test the practical limit directly: on small enumerable state spaces, compute $P_F^\\top(x)$ exactly by dynamic programming and measure how the flow-matching residual shrinks with capacity; this would separate training error from representational bias.","The same conservation-law sampler suggests an amortized proposal for MCMC or a reusable sampler for other discrete combinatorial objects, a connection the thesis only touches on.","Because the exact theorems are for exact flows, the practical quality of a GFlowNet may be predicted by the degree to which learned edge flows satisfy conservation at intermediate states; monitoring this residual is an implicit diagnostic."],"forward_implications":["A GFlowNet trained with a flow matching loss can sample DAGs from the Bayesian posterior $P(G|D)$ rather than from a path-count-biased distribution, because the conservation law cancels trajectory multiplicities.","Off-policy training with a behavior policy can be used to improve exploration without changing the target distribution, and the convergence guarantees hold for the exact flow matching objectives.","GFlowNets and maximum-entropy reinforcement learning are equivalent after a reward correction, so the known multi-path bias of MaxEnt RL disappears when the reward is corrected as in Chapter 5.","The Markov-chain perspective extends GFlowNets to continuous and general state spaces, giving generalized detailed balance and trajectory balance conditions.","A single GFlowNet can approximate the joint posterior over the structure and the parameters of a Bayesian network, with DAG-GFlowNet as a special case."],"supporting_citations":[{"why":"Introduces the generative-flow-network objective and the terminating-flow boundary condition that the thesis builds on.","marker":"Bengio et al. (2021)"},{"why":"Supplies the flow matching and detailed balance characterization of Markovian flows used in Chapters 3 and 4.","marker":"Bengio et al. (2023)"},{"why":"Introduces trajectory balance and the backward-transition probability view used for estimating terminating-state probabilities.","marker":"Malkin et al. (2022)"},{"why":"Extends GFlowNets to general state spaces, grounding the thesis's Chapter 6 treatment.","marker":"Lahlou et al. (2023)"},{"why":"Presents joint Bayesian inference of structure and parameters with a single GFlowNet, the basis of Chapter 8.","marker":"Deleu et al. (2023)"},{"why":"Introduces sub-trajectory balance conditions used as training objectives and in joint inference.","marker":"Madan et al. (2023)"},{"why":"Establishes the maximum-entropy RL framework that the thesis connects to GFlowNets in Chapter 5.","marker":"Haarnoja et al. (2017)"},{"why":"Provides the importance-sampling estimator for terminating-state probabilities used in the empirical evaluation.","marker":"Zhang et al. (2022b)"}],"fun_headline_variants":["Flow networks sample DAGs from the Bayesian posterior exactly","GFlowNets turn unnormalized rewards into principled graph samplers","Beyond one causal model: GFlowNets estimate full posterior over DAGs","Flow-matching law enables exact sampling of compositional objects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The practical claim assumes that a parametric flow trained with flow-matching losses can be driven close enough to exact conservation over a huge discrete space; the theorems are exact only for exactly consistent flows and do not bound the error from finite-sample, function-approximation training.","fun_headline_variants_meta":{"raw":{"variants":["Flow networks sample DAGs from the Bayesian posterior exactly","GFlowNets turn unnormalized rewards into principled graph samplers","Beyond one causal model: GFlowNets estimate full posterior over DAGs","Flow-matching law enables exact sampling of compositional objects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1692,"prompt_tokens":1007,"completion_tokens":685,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":609}},"tokens_in":623,"tokens_out":685,"duration_ms":6729,"temperature":1.0,"reasoning_tokens":609,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:13:14.268763+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a small, fully enumerable state space (for example, all DAGs on four nodes), compute the exact terminating-state distribution $P_F^\\top(x)$ by dynamic programming, train a GFlowNet with a flow matching loss, and test whether the learned distribution approaches $R(x)/Z$ as training time and capacity grow; a persistent gap would falsify the claim that flow matching suffices for sampling.","supporting_citations":[],"review_version":1}