{"id":"c43dc210-e8d7-4e92-976c-2e838d52c655","arxiv_id":"2602.04675","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GSBoG learns topology-respecting controlled CTMC policies for graph transport that match endpoint distributions and optimize running costs via a data-driven IPF and temporal-difference training scheme.","lead":"This paper presents GSBoG, a method that learns controllable random-walk policies on graphs to transport probability mass from a source to a target distribution while minimizing intermediate costs such as congestion. It extends the generalized Schrödinger bridge framework to discrete graph state spaces and demonstrates the approach on supply-chain routing, assignment, and protein-folding tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TD loss (Eq. 28) is mis-scaled for CTMCs: the running-cost signal is O(Δt^2) while martingale noise is O(Δt), so in the continuous-time limit the TD objective cannot enforce cost-aware generator consistency.","rationale":"Good-faith reading: the paper proposes a sensible discrete analogue of DeepGSB on graphs and gives an extensive experimental suite. The derivation of Theorem 3.1 and the Hopf–Cole system is standard and largely correct. The reader's weakest assumption correctly identifies the unproved IPF upper-bound/tightness claim (Eq. 74) as a gap. The stress-test concern is independent and more specific: the running-cost mechanism itself is not continuous-time consistent. Since f cancels in the IPF losses, the TD loss is the sole channel for intermediate-cost shaping. The scaling analysis above is not a matter of missing proof but of the objective's asymptotic behavior: the martingale noise in δ\\hat Y has variance O(Δt), so a squared-residual loss is dominated by that noise, and the O(Δt^2) bias carrying f vanishes as the grid is refined. This directly threatens the second half of the central claim. A small exact-problem test can settle whether the TD loss distinguishes correct from incorrect potentials as Δt→0. Credit where due: the paper's endpoint-matching experiments, if the IPF bound is later justified, would provide support for the first half; and the empirical cost effects at fixed Δt are real observations, though not evidence for the continuous-time statement. Verdict remains CONDITIONAL rather than ACCEPT/REJECT because the flaw is concrete but potentially fixable with a correctly scaled or conditional-expectation-based TD loss and a finite-Δt analysis.","tokens_in":28879,"tokens_out":21539,"duration_ms":205091,"concrete_test":"On a small graph with an exactly computable GSB solution for a nonzero running cost (solve Eq. (15) to high precision on a 3-node chain), evaluate the expected TD loss (Eq. 28) for the exact potential and for a perturbed potential with a large generator residual, at Δt=0.1, 0.01, 0.001. If the loss difference between correct and perturbed potentials shrinks as Δt while the absolute loss remains O(1), the TD objective becomes asymptotically blind to running-cost consistency. A second verification: train GSBoG with λ_TD fixed at Δt=0.1, 0.01, 0.001 and measure whether the learned policy's expected cost E∫f(X_t)dt approaches the exact GSB optimum or collapses to the f=0 solution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim has two components: endpoint matching and intermediate cost shaping. The endpoint part rests on the unproved IPF upper bound, as the reader notes. The cost-shaping part is more fragile. The only place f_t enters training is the TD loss (Eqs. 28–29), since f_t cancels in the IPF losses (Eqs. 26–27). For a CTMC, the one-step increment δ\\hat Y_t of a potential is zero with probability 1−O(Δt) and O(1) with probability O(Δt); its conditional variance is therefore O(Δt), while the squared bias caused by an incorrect generator/PDE is (A_true\\hat Y−A_pred\\hat Y)^2 Δt^2 = O(Δt^2). Hence as Δt→0 the TD loss is dominated by the variance term, which depends on (\\hat Y(y)−\\hat Y(x))^2 and favors flat potentials; the cost-consistency signal f_t is asymptotically negligible. No Δt normalization or conditional-expectation estimator is provided. The finite-Δt experiments may still show cost effects, but the paper's continuous-time GSB formulation is not actually enforced by the stated objective.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GSBoG, a data-driven framework for generalized Schrödinger bridge transport on fixed directed graphs. It formulates the problem as a controlled continuous-time Markov chain (CTMC) with state- and distribution-dependent running costs, derives a dual representation and a Hopf–Cole-type system (Theorem 3.1, Proposition 3.4), and then proposes an alternating iterative proportional fitting (IPF) and temporal-difference (TD) training scheme for learning the log-potentials. Experiments on a 9559-node supply-chain network, assignment problems, and a discretized molecular-dynamics folding task report endpoint marginal matching and improved intermediate-path costs. The claimed contribution is a scalable, executable stochastic routing policy that simultaneously matches prescribed endpoint marginals and shapes intermediate trajectories through general running costs.","tokens_in":29256,"tokens_out":15168,"duration_ms":148194,"significance":"If the training objective is sound, GSBoG would be a useful contribution: it extends generalized Schrödinger bridges to graph-structured state spaces, avoids dense time-expanded solvers, and provides an executable policy rather than a static coupling. The formal CTMC analogue, the sparse large-scale supply-chain comparison, and the assignment/molecular-dynamics demonstrations are valuable and go beyond existing graph-OT methods. However, two load-bearing components are not currently established: the IPF objective is justified by an unproven 'upper bound, tight at the optimum' assertion, and the TD objective that injects the running cost has an unfavorable scaling in the continuous-time CTMC limit. These issues must be resolved before the central claim is supported.","major_comments":[{"comment":"The TD loss is mis-scaled for CTMCs. For a one-step increment δY over [t,t+Δt], the process jumps with probability O(Δt), so δY is zero with probability 1−O(Δt) and O(1) otherwise; its conditional variance is O(Δt). The generator-consistency signal, (A^u Y)Δt, contributes squared terms of order (AΔt)^2 = O(Δt^2). Thus, as Δt→0, L_TD is dominated by the variance term Σ_y (Y(y)−Y(x))^2 u(y,x)Δt, which is minimized by flat potentials and contains the running cost f_t only at subleading order. Since f_t cancels in the IPF losses (Eqs. 26–27) and enters training only through this TD term, the continuous-time cost-shaping claim is not enforced by the stated objective. The authors should either normalize the loss by Δt, use a conditional-expectation estimator, or provide a rigorous finite-Δt analysis showing the cost signal survives for the discretizations used.","section":"§3.3, Eqs. (28)–(29)"},{"comment":"The assertion that replacing the true potentials by parameterized surrogates in the Dynkin identity yields an upper bound on the negative log-likelihood, tight at the optimum, is stated without proof. The Dynkin identity is an exact identity for the true pair (Y,Ŷ); no argument is given for the inequality for arbitrary parameterized (Y,Ŷ). This upper-bound property is what legitimizes minimizing the IPF losses for endpoint matching, so without a proof (or a counterexample-free justification) the convergence of the alternating IPF/TD scheme to the GSB optimum is not established. This is a load-bearing gap for every endpoint-matching experiment.","section":"Appendix A.4, Eq. (74)"},{"comment":"All reported results are from single runs without error bars, confidence intervals, or multiple seeds. Given that the training objective is stochastic and involves random initialization and rollout sampling, the quantitative comparisons (e.g., Table 2 congestion metrics, Table 4 fold rates) cannot be assessed for statistical significance. Please report means and standard deviations over at least several seeds, and state the number of seeds used.","section":"§4, Tables 2–4"}],"minor_comments":[{"comment":"The derivation of the backward generator contains sign/notational inconsistencies (e.g., Eq. (70) appears to use φ_s(x) where the context requires the backward potential, and the ∂Y/∂s signs are inconsistent across Eqs. (67)–(69)). Please correct these so that the proof of Proposition 3.6 is verifiable.","section":"Appendix A.4, Eqs. (67)–(70)"},{"comment":"The objective in Eq. (13) omits the explicit time integral/dt in the expectation notation. As written, 'E_t [...]' is ambiguous and should be written as E_{p^u}[∫_0^1 Σ ... dt].","section":"§3.1, Eq. (13)"},{"comment":"The simulation step says 'simulate CTMC under u_θ' but does not specify whether exact event-based simulation or an Euler-type discretization is used. The finite-Δt interpretation matters for the TD loss scaling; please state the simulation scheme in the main text or point precisely to the appendix.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and useful problem, and the experiments are broad. However, the theoretical underpinning of the training objective needs significant repair: the IPF upper bound is unproven, and the TD loss scaling appears to make the running-cost signal asymptotically negligible in the continuous-time limit. These are fixable within the manuscript's scope, so I am not recommending rejection. The single-run experiment reporting should also be strengthened. I would send the revised version back to the same referees."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi —\n\nThe paper is a legitimate first step: a graph-native generalized Schrödinger bridge built from CTMC generators, with the Hopf–Cole transform and IPF losses derived in a self-contained way. The derivations in Theorem 3.1 and Props 3.4–3.6 are mostly sound (modulo typos in the appendix), and the experimental framing — supply-chain routing, assignment, and chignolin folding — shows the authors are aiming at real problems. That said, I would not trust the cost-shaping component in the continuous-time limit.\n\nThe TD loss (Eq. 28) is the only place the running cost f_t enters, and it is mis-scaled for a jump process. For a CTMC, the one-step increment δ\\hat Y_t is zero with probability 1−O(Δt) and O(1) with probability O(Δt); its variance is O(Δt). The generator prediction A^u \\hat Y Δt is O(Δt), so the squared residual has a variance term of order Δt and a squared-bias term of order Δt^2. As Δt→0, the loss is dominated by the martingale variance, which pushes the potential differences toward zero regardless of f_t. The running-cost signal is asymptotically negligible. No normalization or conditional-expectation trick is provided. This undercuts the claim that the continuous-time formulation enforces cost-aware dynamics. The finite-Δt experiments may still show an effect (and they do — Figure 3b, Table 3), but the paper presents this as a continuous-time GSB, not as a finite-step heuristic.\n\nThe reader's earlier concern about Eq. 74 is also real: the IPF objective is stated as an upper bound on the negative log-likelihood, tight at the optimum, but no proof is given. Endpoint matching in every experiment leans on this. That is a gap, though not necessarily a fatal one — it may be provable with more care.\n\nEmpirically, the paper would be stronger with error bars, code/data, and a successful run of the GrSB baseline on the large graph. The supply-chain numbers are impressive but single-run.\n\nNet: the core construction is novel and the derivations are mostly careful, but the TD scaling problem is a real flaw in the theoretical claim. I would not reject outright; the paper deserves a referee, but the authors need to address the TD scaling (or reframe) and prove the upper-bound claim. I'd want to see code before citing it.","headline":"A novel CTMC-based generalized Schrödinger bridge on graphs with mostly sound derivations, but the TD objective that carries the running cost is mis-scaled for jump processes, so the continuous-time cost-shaping claim is on shaky ground.","tokens_in":29690,"tokens_out":3680,"would_cite":false,"duration_ms":40469,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J27","49L20","90B06"],"pacs":[],"model":"deepseek-v4-flash","headline":"The authors introduce GSBoG, which learns continuous-time Markov chain policies on arbitrary directed graphs that match prescribed source and target distributions while minimizing state-dependent running costs during the journey.","keywords":["Schrödinger bridge","optimal transport on graphs","continuous-time Markov chain","iterative proportional fitting","temporal-difference learning","stochastic optimal control","molecular kinetics","supply-chain transport"],"falsifier":"Solve the GSB exactly on a small graph (e.g., 5–10 nodes) by numerically integrating the coupled Hopf-Cole system (Eq. 18) or by running Sinkhorn/iterative scaling until convergence with the same reference rates and running cost, then train GSBoG on the same instance and compare terminal marginal TV. If the learned terminal TV does not approach the numerical discretization error (e.g., exceeds 1% with a large training budget and small λTD), the claimed endpoint-matching property is falsified. An even sharper test: set λTD=0 with f≠0; the paper's own ablation predicts terminal matching stays ac","tokens_in":28802,"feed_emoji":"🚚","tokens_out":6614,"duration_ms":68356,"temperature":0.7,"pith_summary":"This paper introduces GSBoG, a method for transporting probability mass between two prescribed distributions on a fixed directed graph by learning a continuous-time Markov chain policy. The central claim is that one can simultaneously satisfy endpoint distribution matching and optimize the intermediate path under general state-dependent running costs, by alternating an iterative-proportional-fitting objective (which enforces the endpoints) with a temporal-difference objective (which injects the running cost). If correct, this gives an executable stochastic routing policy rather than a static flow, so operators can steer mass through a network while avoiding congestion or high-energy states. The paper demonstrates the framework on a 9,559-node supply chain graph, on assignment problems, and on steering the folding of a small protein, reporting near-perfect endpoint matching with reduced intermediate crowding.","feed_headline":"Graph transport gets cost-aware Schrödinger bridge policies","feed_subtitle":"A CTMC-based framework matches endpoint marginals on sparse 9,500-node networks while cutting congestion.","key_machinery":"The central object is the Hopf-Cole potential pair (φ,φ̂) on the nodes, which makes the optimal transition rate depend only on local ratios φ(y)/φ(x) along edges. The argument runs on two discrete identities: the generator expansion of the log-potentials (Eqs. 21–22) gives the gIPF endpoint-matching losses (Eqs. 26–27), and Dynkin's formula supplies the temporal-difference residual (Eqs. 28–29) that re-injects the running cost f_t(x,p_t) that the IPF losses cancel out. What carries the argument is that both objectives are computed from sampled trajectories, so no dense time-expanded solver over the whole graph is needed.","core_discovery":"On a directed graph G=(X,E), the paper formulates the generalized Schrödinger bridge as a control problem over CTMC transition rates, and proves (Theorem 3.1) that the optimal controlled rate takes the form u*_t(y,x)=r_t(y,x)exp(-V_t(y)+V_t(x)) for a time-dependent potential V. Applying a Hopf-Cole transform φ=e^{-V}, p=φφ̂, the path-space objective decomposes into a pair of coupled equations whose solution is characterized by local edge ratios φ(y)/φ(x). The paper then derives discrete analogues of the continuous gIPF losses — using the generator/Dynkin expansion along CTMC trajectories — plus a temporal-difference loss that makes the running cost visible, and trains the log-potentials by a","pith_inferences":["If the upper-bound/tightness assertion (Eq. 74) fails for a given parameterization or discretization, the method may degrade to approximate endpoint matching; testing on small graphs with exactly solved GSB would reveal how far the learned marginals drift.","The same alternating IPF/TD scheme could be applied to time-varying topologies, hard capacity constraints (rather than soft congestion penalties), or multi-agent settings where the running cost depends on the full occupancy field — these are natural extensions not explored in the paper.","The Dynkin-generator derivation suggests the method is not limited to graphs: any Markov process with a computable generator (e.g., continuous-space jump processes) could inherit the same likelihood/TD training, opening a route to non-diffusion Schrödinger bridges."],"forward_implications":["GSBoG provides executable routing policies on fixed graphs, not just static couplings, so operators can simulate and deploy the learned stochastic dynamics directly.","Because the IPF and TD objectives are evaluated along sampled trajectories, the method scales to large, sparse graphs (the paper tests 9,559 nodes) where global solvers run out of memory.","Incorporating a congestion- or free-energy-dependent running cost shapes intermediate trajectories, which the paper shows suppresses peak occupancy and lowers energy barriers while preserving endpoint matching.","The assignment-task experiments indicate that pairwise costs can be encoded as node costs on auxiliary intermediate nodes, so the same framework solves cost-aware assignment transport."],"fun_headline_variants":["Cost-aware Schrödinger bridge for sparse graph transport","Generalized Schrödinger bridge learns scalable graph policies","Schrödinger bridge method optimizes graph transport costs","Graph transport via cost-aware Schrödinger bridge","Scalable Schrödinger bridge for cost-aware graph routing"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The training objective rests on the assertion that the Dynkin-based expression in Eq. (74) is an upper bound on the negative log-likelihood, tight at the optimum, and that alternating IPF/TD minimization converges to the generalized Schrödinger bridge optimum in the parameterized finite-Δt setting; neither the tightness nor the convergence is proven.","fun_headline_variants_meta":{"raw":{"variants":["Cost-aware Schrödinger bridge for sparse graph transport","Generalized Schrödinger bridge learns scalable graph policies","Schrödinger bridge method optimizes graph transport costs","Graph transport via cost-aware Schrödinger bridge","Scalable Schrödinger bridge for cost-aware graph routing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00038,"raw_usage":{"total_tokens":1844,"prompt_tokens":720,"completion_tokens":1124,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":1052}},"tokens_in":464,"tokens_out":1124,"duration_ms":11706,"temperature":1.0,"reasoning_tokens":1052,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T04:29:30.034831+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the GSB exactly on a small graph (e.g., 5–10 nodes) by numerically integrating the coupled Hopf-Cole system (Eq. 18) or by running Sinkhorn/iterative scaling until convergence with the same reference rates and running cost, then train GSBoG on the same instance and compare terminal marginal TV. If the learned terminal TV does not approach the numerical discretization error (e.g., exceeds 1% with a large training budget and small λTD), the claimed endpoint-matching property is falsified. An even sharper test: set λTD=0 with f≠0; the paper's own ablation predicts terminal matching stays ac","supporting_citations":[],"review_version":1}