Pith. sign in

REVIEW 5 major objections 7 minor 13 references

RL agent places Lightning Network liquidity better than heuristics

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-10 02:37 UTC pith:WCIG6IPI

load-bearing objection Real deployment, but the balance-sampling assumption is load-bearing and untested. the 5 major comments →

arxiv 2607.08703 v1 pith:WCIG6IPI submitted 2026-07-09 cs.LG

MPFlow: Learning Budgeted Max-Flow Optimization on the Lightning Network with Deep Graph Reinforcement Learning

Harrison Rush , Vincent Davis , Simone Antonelli , Vikash Singh , Jesse Shrader , Emanuele Rossi This is my paper
classification cs.LG
keywords networklearninglightningmax-flowoptimizationpolicyacrossagent
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that the question of where to place liquidity on the Bitcoin Lightning Network—specifically, which peer nodes to open payment channels with—can be answered better by a reinforcement learning agent than by the graph centrality heuristics currently used in practice. The authors frame the problem as selecting k edge additions to a directed capacity graph that maximize s-t max-flow, a classical measure of how much payment volume can be routed between a source and its targets. Their agent, MPFlow, uses a message-passing neural network with element-wise max aggregation (chosen to detect bottleneck and min-cut structure), trained with PPO under a hub-exclusion curriculum that removes the top-50 highest-degree nodes from training subgraphs, forcing the policy to learn capacity-aware placement rather than simply attaching to well-connected hubs. On 1000 paired evaluations against identical network realizations, MPFlow achieves a statistically significant 8.6% improvement over Betweenness centrality—the strongest heuristic and the prevailing industry baseline—and its advantage widens as hubs are pruned, suggesting it exploits structural bottlenecks rather than chasing hubs. The agent has been deployed in production, executing 4640 channel-open decisions allocating 267.3 BTC across 30 managed nodes.

Core claim

The central mechanism is the combination of max-aggregation message passing with a hub-exclusion training curriculum, applied to the max-flow objective on real Lightning Network topology. Max aggregation preserves the most constraining neighbor feature in each receptive field, matched to an objective whose value is governed by min-cut structure; the hub-exclusion curriculum prevents the policy from collapsing onto the degenerate strategy of attaching to the few largest hubs, which centrality heuristics already implement. Together, these yield a policy that allocates liquidity far lower in the degree hierarchy (median chosen peer rank 312, versus 9 for the GCN baseline and 43 for Betweenness)

What carries the argument

MPNN with max aggregation + PPO + hub-exclusion curriculum on max-flow reward

Load-bearing premise

Every max-flow computation in the paper depends on per-direction channel balances, which are not publicly observable on the Lightning Network. The authors sample these balances uniformly at random, subject to summing to the known total capacity. If real balances are systematically non-uniform—as strategic node behavior, fee optimization, and payment flow directionality would suggest—then the optimal channel placements and the relative ranking of policies could change, since

What would settle it

If real Lightning Network balances are systematically skewed, the uniform sampler could misidentify bottlenecks

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 7 minor

Summary. The paper addresses liquidity placement on the Bitcoin Lightning Network (LN): given a fixed budget of k channel openings, which peers should a node connect to in order to maximize s–t max-flow? The authors formulate this as a budgeted combinatorial optimization problem on graphs and solve it with a message-passing neural network (MPNN) using max aggregation, trained with PPO under a hub-exclusion curriculum (top-50 degree nodes removed during training). The agent is evaluated on real LN snapshots using a paired protocol (fixing source, sampled balances, and PRNG seed across methods). The main result is a statistically significant paired uplift of +8.59% ± 6.20 over Betweenness centrality on a 5k-node subgraph across 1000 paired episodes, with additional cross-snapshot generalization results and a production deployment of 4640 channel-open decisions.

Significance. The problem formulation is well-motivated and practically relevant: framing LN liquidity placement as budgeted max-flow optimization on a real financial network is a clean contribution, and the hub-exclusion curriculum is a thoughtful training strategy that forces capacity-aware rather than hub-seeking behavior. The paired evaluation protocol is methodologically sound, and the paper is commendably honest about the confounds in its GCN comparison (§6.4) and the limitations of the uniform balance sampler (§7). The production deployment, even without reported metrics, demonstrates that the approach is lightweight enough for real-world use. The cross-snapshot generalization study (§6.3) provides evidence of temporal robustness. The code and trained checkpoints are stated to be available, which supports reproducibility.

major comments (5)
  1. §5.1, App. F: The uniform balance sampling (α_uv ~ Uniform(0,1)) is the single most load-bearing assumption in the paper. Every reward computation, evaluation metric, and comparative result depends on the directed balances, which are not publicly observable. The concern is structural: MPFlow is trained on uniformly-sampled balances and evaluated on uniformly-sampled balances, creating a train-evaluation distribution match that inherently favors the learned policy over balance-agnostic heuristics like Betweenness. The paper itself cites Channel Balance Interpolation work (Vincent et al., 2024; Davis et al., 2025) that could provide more realistic balance priors, and even notes such priors 'can reduce variance in simulation and improve routing objectives,' yet does not use them or perform any sensitivity analysis. A sensitivity experiment with at least one non-uniform balance distribution—
  2. continued: e.g., a Beta-skewed distribution or balances informed by the cited CBI method—would substantially strengthen the central claim. Without it, the +8.59% uplift over Betweenness may be an artifact of the uniform sampling regime rather than a robust property of the learned policy. This is acknowledged in §7 but not tested.
  3. §6.4, Table 1: The comparison between MPFlow and the GCN baseline is confounded across at least four factors simultaneously: max vs. mean aggregation, edge-conditioned messages, access to edge features, and the hub-exclusion curriculum. The paper acknowledges this honestly (§6.4: 'so the mechanisms below are consistent with our results but not isolated by them'), but the ablation study (§6.2) only varies graph scale and hub removal—it does not isolate any of these factors. The curriculum is partially isolated in App. B (Table 9), which is valuable, but the contribution of max aggregation (the paper's central architectural claim) remains untested in isolation. A single-factor ablation replacing max with mean aggregation in the MPFlow architecture (holding edge features and curriculum fixed) would directly test the claim that max aggregation is 'matched to the min-cut structure of the.'
  4. continued: objective.' Without this, the paper's explanation for why MPFlow outperforms GCN remains speculative.
  5. §1, Abstract, §7: The production deployment (4640 channel-open decisions, 267.3 BTC) is cited prominently in the abstract and conclusion as evidence of practical relevance, but no performance metrics from production are reported. Without at least a retrospective comparison (e.g., observed routing capacity or relay volume of MPFlow-recommended channels vs. channels opened under the previous heuristic), the deployment serves as evidence of deployability but not of effectiveness. The authors should either report some production performance metrics or explicitly scope the deployment claim to 'deployed in production for peer recommendations' without implying validated real-world performance.
minor comments (7)
  1. Table 3 (§6.3): The cross-snapshot results report 'relative improvement over Random' rather than the paired uplift over Betweenness used in the main results (Table 1). This makes direct comparison between the two tables difficult. Consider reporting the paired uplift over Betweenness for the cross-snapshot study as well, or at least stating the Betweenness-relative gap explicitly in the text.
  2. App. F: The target set for max-flow computation comprises '50% of the nodes in the network, sampled randomly' per episode. This design choice is not discussed in the main text and could significantly affect the reward landscape. A brief justification in §4.1 or §5.3 would help.
  3. §4.1: The choice of k=5 is justified briefly ('In internal sweeps over K∈{1,3,5,10,20}, K=5 yielded the most reliable policy gradients'), but no data from these sweeps is shown. A table or figure summarizing this sweep would strengthen the justification.
  4. §5.2: GAT is reported only in the cross-snapshot study (§6.3) because it 'was added after the primary experiments were completed.' For consistency, consider noting this limitation more prominently or running GAT on the main 5k subgraph as well.
  5. Figure 1 caption: 'budget k=5×0.20 BTC' is slightly ambiguous; clarify whether this means 5 actions each of 0.20 BTC (total 1.0 BTC) or something else.
  6. Table 1: The 'd̄ vs Betw.' column is in units of 10^6 sat, while ΔF is in BTC. Consider using consistent units or making the conversion more prominent.
  7. §3 (Related Work): The paragraph on Pickhardt & Richter (2021) and related work is somewhat dense. A brief sentence clarifying how the max-flow objective relates to (or differs from) the min-cost flow formulation of Pickhardt & Richter would help readers from the LN routing literature.

Circularity Check

0 steps flagged

No significant circularity: the max-flow reward is computed externally via push-relabel, baselines are standard heuristics, and the one self-citation (Davis et al. 2025 on balance interpolation) is acknowledged as future work rather than used as load-bearing support.

full rationale

The paper's central claim—that MPFlow outperforms Betweenness centrality on the max-flow objective—is evaluated against externally defined, standard metrics. The max-flow reward (rt = Ft - Ft-1) is computed via the push-relabel algorithm (Goldberg & Tarjan, 1988), an independent classical algorithm, not defined by the paper's own framework or agent. The baselines (Random, Degree, Betweenness) are standard graph heuristics from the literature (Freeman 1977; Newman 2010), not constructs of the authors. The hub-exclusion curriculum is a training strategy with no circular dependency on the evaluation metric. The one self-citation to Davis et al. (2025) / Vincent et al. (2024) on Channel Balance Interpolation appears in Related Work as a complementary approach the authors explicitly do not use, noting 'Our work differs by optimizing absolute max-flow uplift with a learned MPNN-PPO policy.' The paper acknowledges the uniform balance sampling as a limitation (§7) and cites CBI as a direction for reducing variance, but does not invoke it as a premise for any derivation. No 'prediction' or 'first-principles result' reduces to its inputs by construction. The derivation chain from problem formulation (MDP with max-flow reward) to evaluation (paired uplift vs. Betweenness) is self-contained against external benchmarks. The skeptic's concern about train-evaluation distribution match under uniform balance sampling is a correctness/external-validity risk, not a circularity issue—the reward is not defined in terms of the agent's output, and the evaluation metric is not constructed from the training procedure.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

No new entities are postulated. The paper uses existing graph-theoretic constructs (max-flow, min-cut) and existing ML architectures (MPNN, PPO).

free parameters (7)
  • k (budget) = 5
    Fixed at 5 based on internal sweeps over K∈{1,3,5,10,20}; chosen for 'most reliable policy gradients' (§4.1). This is a hyperparameter selected by performance, not derived.
  • per-action channel size = 0.20 BTC
    Fixed channel size per action, stated in Figure 1 and §5.3. Not justified or swept.
  • exclude_top_n_degree = 50
    Number of top hubs removed during training (§5.2). Chosen value; the ablation in Table 2 sweeps removal at evaluation time but not the training exclusion depth.
  • hidden dimension = 64
    MPNN hidden width (Table 11). Standard hyperparameter.
  • num message-passing layers = 2
    Depth of MPNN backbone (Table 11). Standard hyperparameter.
  • PPO learning rate = 0.002
    Adam learning rate (Table 12). Standard hyperparameter.
  • target set size = 50% of nodes
    The max-flow is computed from source to a random 50% of nodes (App. F). This fraction is not justified or swept.
axioms (4)
  • domain assumption Max-flow upper-bounds and serves as a valid proxy for deliverable payment volume on the Lightning Network
    Stated in §4.1: 'Max-flow therefore serves as a theory-grounded proxy for deliverable throughput.' This is the central modeling assumption. The paper acknowledges that validating transfer to realized payment success is future work (§7).
  • domain assumption Per-direction channel balances are uniformly distributed
    §5.1 and App. F: balances are sampled uniformly at random because they are not publicly observable. This assumption underlies every reward computation and evaluation.
  • standard math Push-relabel algorithm correctly computes max-flow
    Standard algorithm (Goldberg & Tarjan 1988); correctness is well-established.
  • domain assumption Lightning Network snapshots from a single production node are representative of the network
    §5.1: snapshots collected from a production node operated by Amboss Technologies. The paper does not validate that this node's view of the network is complete or representative.

pith-pipeline@v1.1.0-glm · 17458 in / 4303 out tokens · 200448 ms · 2026-07-10T02:37:50.662883+00:00 · methodology

0 comments
read the original abstract

We address liquidity placement in the Bitcoin Lightning Network (LN): given a fixed budget, which channels should a node open to maximize its routing capacity? We cast this as a budget-constrained combinatorial optimization problem on graphs, selecting $k$ edge additions that maximize $s$--$t$ max-flow, a theory-grounded measure of routing capacity, and solve it with graph reinforcement learning. Our lightweight agent combines a message-passing policy network with proximal policy optimization (PPO) and action masking, and is trained under a hub-exclusion curriculum: the network's top hubs are removed from training subgraphs, forcing the policy to learn capacity-aware placement rather than hub attachment. In extensive experiments on real Lightning Network snapshots, our method consistently outperforms strong heuristic baselines on the max-flow objective across multiple seeds and unseen graphs. The agent has been deployed in production for peer recommendations, executing 4640 channel-open decisions that cumulatively allocate 267.3 BTC over $16 million across 30 managed nodes.

Figures

Figures reproduced from arXiv: 2607.08703 by Emanuele Rossi, Harrison Rush, Jesse Shrader, Simone Antonelli, Vikash Singh, Vincent Davis.

Figure 1
Figure 1. Figure 1: Liquidity placement as learned sequential graph construction. Left: MPFlow maps each state st to a peer at (budget k=5 × 0.20 BTC); reward: marginal max-flow rt = Ft − Ft−1. Right: two max-aggregation MPNN layers, a masked-softmax actor, and a max-pooled critic, trained with PPO. 1 arXiv:2607.08703v1 [cs.LG] 9 Jul 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: MPNN–Max actor–critic used in MPFlow. Inputs are node features X ∈ R N×F , edge features E ∈ RM×Fe (log(1+x) then z-score), and the directed edge list (self-loops added). The encoder has two edge-conditioned convolutions (NNConv) whose weights W(eij ) are produced by small edge MLPs; messages are aggregated by element-wise max, each followed by ReLU; a final LayerNorm is applied to the node embeddings. The… view at source ↗
Figure 3
Figure 3. Figure 3: PPO training flow. We collect K on-policy trajectories by sampling actions from a masked policy head and append transitions (st, at, log πold,t, rt, maskt, donet, st+1) to the rollout buffer. After collection, the value head (no grad) provides Vt, Vt+1 to compute TD residuals δt = rt + γ(1 − donet)Vt+1 − Vt, advantages At = GAE(δt; γ, λ), and targets Rt = At + Vt; At is standardized. PPO then runs for mult… view at source ↗
Figure 4
Figure 4. Figure 4: Paired uplift relative to Betweenness on the 5k subgraph (n=1000 paired episodes; bars are mean paired % uplift, whiskers are 95% CIs). MPFlow is the only policy whose CI excludes zero on the positive side; Degree, GCN, and Random all underperform Betweenness. Values match [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Robustness under graph size. Mean increase in max-flow (sat) after allocations vs. number of nodes included (ranked by degree) [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Robustness under targeted hub removal (5k subgraph). Mean increase in max-flow (sat) after allocations vs. number of highest-degree hubs removed [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: and [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages · 7 internal anchors

  1. [1]

    Learning Combinatorial Optimization Algorithms over Graphs

    URLhttp://arxiv.org/abs/1704.01665. arXiv:1704.01665. Vincent Davis, Vikash Singh, and Emanuele Rossi. Channel balance interpolation in the lightning network via machine learning. IEEE ICBC 2025 Workshop: NextGenDLT (Program Listing),

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    doi: 10.1145/48014.61051

    ISSN 0004-5411. doi: 10.1145/48014.61051. URLhttps://dl.acm.org/doi/ 10.1145/48014.61051. Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907,

    work page doi:10.1145/48014.61051

  3. [3]

    Autoopen.https://docs.lightning.engineering/lightning-network-tools/ lightning-terminal/autoopen, 2024a

    Lightning Labs. Autoopen.https://docs.lightning.engineering/lightning-network-tools/ lightning-terminal/autoopen, 2024a. Accessed 2025-08-28. Lightning Labs. Getnodemetrics (betweenness centrality) — lnd api.https://lightning.engineering/ api-docs/api/lnd/lightning/get-node-metrics/index.html, 2024b. Accessed 2025-08-28. Lightning Labs. lnd/autopilot/agen...

    work page 2025

  4. [4]

    Continuous control with deep reinforcement learning

    URLhttp: //arxiv.org/abs/1509.02971. arXiv:1509.02971. Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous Methods for Deep Reinforcement Learning, June

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    Asynchronous Methods for Deep Reinforcement Learning

    URLhttp://arxiv.org/abs/1602.01783. arXiv:1602.01783. Satoshi Nakamoto. Bitcoin: A Peer-to-Peer Electronic Cash System

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [7]

    Optimally Reliable & Cheap Payment Flows on the Lightning Network

    URLhttps://arxiv.org/abs/2107.05322. Rene Pickhardt, Sergei Tikhomirov, Alex Biryukov, and Mariusz Nowostawski. Security and privacy of lightning network payments with uncertain channel balances.arXiv preprint arXiv:2103.08576,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [8]

    Security and Privacy of Lightning Network Payments with Uncertain Channel Balances

    URLhttps://arxiv.org/abs/2103.08576. Joseph Poon and Thaddeus Dryja. The Bitcoin Lightning Network: Scalable off-chain instant payments

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [9]

    Proximal Policy Optimization Algorithms

    URLhttp://arxiv.org/abs/1707.06347. arXiv:1707.06347. István Seres, András Gulyás, József Stéger, and András Benczúr. Topological analysis of the lightning network. InFinancial Cryptography and Data Security Workshops,

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [10]

    doi: 10.48550/arXiv.1809. 05088. Petar Veličković and Charles Blundell. Neural algorithmic reasoning.Patterns, 2(7):100273,

    work page doi:10.48550/arxiv.1809

  10. [12]

    Peng Wang, Hong Xu, Xin Jin, and Tao Wang

    URLhttps://arxiv.org/abs/2405.12087. Peng Wang, Hong Xu, Xin Jin, and Tao Wang. Flash: Efficient Dynamic Routing for Offchain Networks, June

    work page arXiv

  11. [13]

    Flash: Efficient Dynamic Routing for Offchain Networks

    URLhttp://arxiv.org/abs/1902.05260. arXiv:1902.05260. Yifei Yang, Dongmian Zou, and Xiaofan He. Graph Neural Network-Based Node Deployment for Throughput Enhancement.IEEE Transactions on Neural Networks and Learning Systems, 35(10):14810–14824, Octo- ber

    work page internal anchor Pith review Pith/arXiv arXiv 1902

  12. [14]

    doi: 10.1109/TNNLS.2023.3281643

    ISSN 2162-2388. doi: 10.1109/TNNLS.2023.3281643. URLhttps://ieeexplore.ieee.org/ document/10154469. Conference Name: IEEE Transactions on Neural Networks and Learning Systems. 11 A Additional Figures 12 Figure 2:MPNN–Max actor–critic used in MPFlow.Inputs are node featuresX∈RN×F, edge featuresE∈RM×Fe (log(1+x)then z-score), and the directed edge list (sel...

    work page doi:10.1109/tnnls.2023.3281643 2023

  13. [15]

    Their seminal work laid the foundation for analyzing flow networks by formulating the Max-Flow Min-Cut theorem and proposing the Ford-Fulkerson algorithm for computing the maximum flow in a network (Ford & Fulkerson, 1956). It involves determining the maximum amount of flow that can be sent from a source nodesto a sink nodetin a flow network, represented ...

    work page 1956