acceptodds
Under review as a conference paper at ICLR 2027

Dissecting Long-range Dependency in Graph Implicit Models

Abstract

Capturing long-range dependencies in an effective manner has been one of the key bottlenecks for advances in graph neural networks. Recently, graph implicit models (GIMs) have gained popularity as they provide a means to capture long-range dependency by solving for embeddings that are fixed-point solutions to `infinitely deeply stacked' graph neural networks, giving an effectively unbounded receptive field. Prior work shows examples where standard graph neural networks demonstrably fail to capture long-range dependencies, while graph implicit models succeed. Yet, graph implicit models do not necessarily outperform standard graph neural networks in real-world long-range benchmarks. In this paper, we theoretically disentangle this conundrum. We first analyze why linear GIMs can solve the simple chain task through an aligned learning-dynamics surrogate. We then establish a quantitative vanishing-influence bound: under contraction and sensitivity bounds uniform in graph size, GIMs cannot fit exact scores or maintain a fixed positive classification margin on the generalized chain-parity task for sufficiently large . Motivated by this diagnosis, we propose , a GIM with per-node adaptive selection that retains intermediate representations while propagation continues. Benchmarking against implicit and standard graph models on synthetic chain-parity and four graph-property datasets shows that it achieves the best mean on three of the four graph datasets in our comparison. GLoRa and selection ablations provide further evidence for the proposed design.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Implicit Model, Parity Task, Adaptive Halting

Reviews

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

The paper revisits long-range dependency in graph implicit models, using an aligned learning-dynamics surrogate for the simple chain task and a vanishing-influence bound under contraction and sensitivity bounds uniform in graph size. AdaGIM adds per-node adaptive selection: a separate selected state retains intermediate representations while propagation continues, using a learned gate and residual factor.

Critical strengths
  1. The paper distinguishes vanishing influence from failure to classify. Its parity result concerns exact scores or a fixed positive margin, while correct classification with a vanishing margin remains possible. It also clearly limits the simple-chain result to an aligned unconstrained surrogate. These distinctions help interpret what the benchmarks actually demonstrate.
  2. The selection mechanism has a clear mathematical explanation. Under contractive propagation, the residual gate gives finite total write mass, allowing earlier states to retain weight. The paper also distinguishes the initialized trajectory limit from a unique algebraic fixed point and explains finite-unroll differentiation.
Critical weaknesses
  1. The main benchmark results need stronger validation. Table 1 uses a reduced training protocol, and the paper acknowledges unresolved seed counts, displayed spreads, and tuning allocations. This makes it difficult to judge whether the reported gains are reliable and the comparisons are fair.
  2. The benefit of last-write selection is not fully isolated. The JK-style baseline uses only trajectory averaging, leaving stronger learned readouts untested. In Table 3, the full gate beats MLP-only gating on only one of three datasets. Given the substantial memory overhead, comparisons with matched-capacity and matched-computation alternatives would help establish the mechanism’s practical value.
  3. The computation cost and memory cost of the proposed method are high.
Questions for authors
  1. With the same propagation, parameter budget and trajectory length, how does last-write selection compare with learned JK-style attention and node-adaptive stopping/readout? Does a node-shared gate work similarly?
1 comment

Thank you for the feedback.

Benchmark validation (W1)

We reran all ten models with the same input features and 20 tuning attempts per model and dataset. The validation-selected configurations are evaluated over five seeds, reporting mean ± sample standard deviation (Table R1 in our reply to Reviewer Zq9j). AdaGIM achieves the best mean on both Peptides tasks. We will release the configurations, per-seed results, and logs.

A separate comparison in Appendix G.1 uses approximately 500K parameters and 250 training epochs. AdaGIM achieves 0.6641 Peptides-func AP versus 0.6327 for GCN, supporting its effectiveness with longer training.

Learned readouts and shared gates (W2 and Q1)

We added learned JK-attention, JK-LSTM, and per-node ACT-style halting. These variants are trained separately end to end using the same backbone architecture, configuration, maximum propagation budget, and training schedule as AdaGIM. JK-attention uses the same scorer architecture and essentially the same parameter count. The long-solve baseline receives up to 100 iterations with tolerance 10−510^{-5}.

Table R4. Readout comparison. Seeds 0–2; means and reported variability. Parameter counts are for Peptides-func; bold denotes the best mean.

ReadoutParameters (Peptides-func)Peptides-func AP ↑Peptides-struct MAE ↓ZINC MAE ↓
AdaGIM (a=qσ~a=q\widetilde{\sigma})122,8940.5737 ± 0.00480.2883 ± 0.00470.4185 ± 0.0088
σ~\widetilde{\sigma} only122,8920.5649 ± 0.00570.2879 ± 0.00240.4076 ± 0.0032
qq only56,8440.5437 ± 0.00920.4000 ± 0.00250.4301 ± 0.0119
Last iterate HTH^T56,8420.5427 ± 0.00480.4504 ± 0.02650.4210 ± 0.0084
Long-solve endpoint56,8420.5707 ± 0.00930.4709 ± 0.01520.4391 ± 0.0116
Trajectory mean (JK-mean)56,8420.4798 ± 0.01890.4016 ± 0.00810.4166 ± 0.0183
JK-attention (gate scorer architecture)122,8920.5009 ± 0.00720.3631 ± 0.00230.4300 ± 0.0143
JK-LSTM255,7560.4922 ± 0.02540.3672 ± 0.00700.3796 ± 0.0173
Graph-shared gate122,8940.5715 ± 0.00520.2677 ± 0.00220.4438 ± 0.0076
Per-node ACT-style halting122,8920.4380 ± 0.01010.3761 ± 0.00420.4702 ± 0.0085
  • Learned readouts: AdaGIM improves over JK-attention by 0.0728 AP on Peptides-func and 0.0748 MAE on Peptides-struct. It also outperforms JK-LSTM, which has roughly twice the parameters, and ACT on both tasks (Welch p≤0.03p\le0.03).
  • Shared gates: Our gate MLP already shares parameters across nodes. Using one gate value for all nodes in a graph gives similar Peptides-func AP (0.5715 versus 0.5737) and improves Peptides-struct MAE to 0.2677. Nodewise gating performs better on ZINC.
  • Gate factors: Sigmoid-only gating also gives strong performance at the tested iteration budgets. Under contractive propagation, qq allows earlier selected states to retain weight after convergence (Appendix D; our reply to Reviewer fHKD, Q3).

Training and inference cost (W3)

There is a standard technique to reduce the memory usage. Gradient checkpointing reduces peak training memory on both Peptides tasks by approximately 76%, to about 1.5 GB. It saves memory by recomputing intermediate states during backpropagation.

Table R5. Cost on A40 GPUs. IGNN values are in parentheses; Tmax⁡T_{\max} is the maximum propagation budget per core. Inference time is averaged per graph within a batch.

Dataset (Tmax⁡T_{\max} × cores)Training seconds/epochTraining peak memory, MBWith checkpointing: seconds/epoch, MBBatched inference, ms/graph
Peptides-func (30 × 2)20.7 (4.0)6,223 (183)30.7, 1,5030.89 (0.29)
Peptides-struct (30 × 3)24.1 (2.2)6,565 (113)38.9, 1,5381.08 (0.17)
ZINC (20 × 1)4.3 (3.0)311 (38)8.4, 990.30 (0.26)
MolHIV (10 × 2)15.9 (10.8)374 (51)28.1, 1170.24 (0.20)

Checkpointing uses 1.48× and 1.61× the training time on Peptides-func and Peptides-struct, respectively. Batched inference averages 0.24–1.08 ms per graph in the tested configurations.

Our contributions are a theoretical analysis of long-range limitations under contractive propagation and the general AdaGIM framework, which separates propagation from adaptive selection. The experiments demonstrate the value of retaining intermediate representations for long-range prediction. Designing more efficient models within this framework, with lower training time and memory requirements, is a promising direction for future work.

Reviewer_fHKDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Implicit GNNs (IGNNs) are graph neural networks, where instead of having a model with given depth, the receptive filed is increased by iterating over the same, typically single layer model, possibly until convergence. The hope is that implicit models can address long-range dependent tasks, but this capability is not proved in the state of the art.

This paper proves the contrary, by showing an example, where they prove that an implicit model can not reach 100% accuracy, and the accuracy decreases with the distance that should be covered. Then, a modification of implicit models is proposed, where representation update can be stopped with a learnable gate, while the propagation of the embeddings can continue in the implicit model itself. The results show that this learnable gate improves the performance of the ungated implicit model.

Critical strengths

The main strength of the paper is the first part, where it is proved for the (K,L) chain parity task that the accuracy of the well-established EIGNN model is upper bounded below one, and the gap depends on L, the length of the chain. I believe this result is interesting for the community working on implicit GNNs.

The proposed method with adaptive gating makes sense intuitively, and the paper proves that it keeps the good properties of EIGNN in terms of unique equilibrium and convergence. It also improves the accuracy of some long-range dependent tasks, including the chain parity task.

Critical weaknesses

The story of the paper is appealing. It shows first that the effects of distant features diminish in IGNNs, and then proposes a solution that can learn to stop the updates in the implicit model. However, I do not see it demonstrated that the gating directly addresses the diminishing effect of the distant inputs, and I do not see in the definition of the “Practical selection gate” that it would try to address distance dependence in any way. On the contrary, the authors seem to show in Appendix G4 that the probability of using the gating does not seem to relate to these distances.

It would also be very interesting to see a mathematical model of the gated EIGNN, similar to the results the authors have for the basic IGNN in Lemma 3.5 and Theorem 3.7.

Questions for authors

The notation G should be introduced on page 2.

What is the y-axis on Fig. 2?

The gating has two terms, q and sigma. I see that ablation study is performed on these, but I would appreciate an intuitive explanation why both (and specially q) is needed. It would be also interesting to see for what H^t, H^t+1 pairs does the sigma factor close the gate. Maybe it could be shown for the chain parity task. Would that demonstrate that the gates address the diminishing effect of the distant input?

Are we satisfied with the AdaGIM results in Table 2? Can it be explained why does the accuracy still fall for K > 1?

I find Appendix D4 with the results in Table 10 important, since this is the only part that tries to evaluate what AdaGIM actually does. However, the related text is very short and quite confusing. I would expect similar numerical results with better discussion already in the main paper.

2 comments

Thank you for recognizing the value of the chain-parity analysis and adaptive selection.

Gated EIGNN and the two gate factors (Q3)

AdaGIM's selection framework is independent of the propagation backbone. Eq. (5) already provides the general mathematical model:

Z^{t+1}=a^{t+1}\odot H^{t+1}+(1-a^{t+1})\odot Z^t, \quad a=q\odot\widetilde{\sigma}.$$ Here, $\bar f_\theta$ is the chosen propagation update. Choosing EIGNN as the backbone gives an instance of this same framework. $H$ propagates information, while $Z$ retains information for prediction. - **$q$ reduces overwriting as propagation stabilizes.** It becomes small when consecutive propagation states change little. - **$\widetilde{\sigma}$ learns how much of a new state to accept.** A small value keeps more of the previous selected state; a large value incorporates more of the new state. **Proposition 4.1 justifies this general framework:** for fixed initialization, if propagation converges and its update is continuous at the limit, the selected state also converges. Thus, adding selection preserves a well-defined limiting representation. Under contractive propagation, $q$ additionally ensures $\sum_t a_v^t<\infty$, allowing earlier accepted states to retain positive weight (Appendix D). With sigmoid-only gating, repeated updates instead give $Z^\infty=H^\star$. On the reported chains, the sigmoid is nearly constant, and $q$ accounts for the selection timing. We will show both factors separately. On real graphs, adding the learned factor to $q$ improves Peptides-func AP from **0.5437 to 0.5737** and Peptides-struct MAE from **0.4000 to 0.2883** (Table R4 in our reply to Reviewer wfbg). ## Multi-input parity (Q4) For $K>1$, the model must combine several bits to compute parity, making the task more demanding than transferring one bit. Our five-seed diagnostic extends the chain length from 20 to 30, with supervision on nodes 1–10. **Table R2. Chain-parity diagnostic.** Five seeds; accuracy on nodes 11–30 at the checkpoint with the highest overall test accuracy, using the same rule for every model. Bold denotes the best mean. | Model | $K=1$ | $K=2$ | $K=3$ | |---|---:|---:|---:| | GCNII | 0.670 ± 0.011 | 0.598 ± 0.014 | 0.534 ± 0.006 | | IGNN | **1.000 ± 0.000** | 0.505 ± 0.007 | 0.502 ± 0.006 | | EIGNN | **1.000 ± 0.000** | 0.502 ± 0.006 | 0.615 ± 0.014 | | MGNNI | **1.000 ± 0.000** | 0.773 ± 0.006 | 0.535 ± 0.052 | | GIND | 0.615 ± 0.058 | 0.573 ± 0.038 | 0.550 ± 0.010 | | Monotone-MIGNN | 0.575 ± 0.050 | 0.542 ± 0.062 | 0.500 ± 0.000 | | AdaGIM | **1.000 ± 0.000** | **0.855 ± 0.058** | **0.726 ± 0.116** | AdaGIM achieves the best mean for both multi-input settings, improving over the next-best models by **8.2 percentage points for $K=2$** and **11.1 for $K=3$**. These gains show improved performance on multi-input computation. Theorem 3.7 concerns exact parity scores and fixed positive prediction margins; this diagnostic measures classification accuracy at a fixed chain length. ## Clearer selection diagnostics (Q5) Table 10 reports each iteration's contribution to the selected representation, averaged over nodes. The chain analysis above shows how selection timing varies between nodes. We will move both analyses into the main text with clearer explanations. **Table R3. Selection weights over propagation iterations.** Weights are averaged over test nodes and normalized after excluding the initial-state weight. | Dataset | Iterations 1–5 | Iterations >5 | |---|---:|---:| | Peptides-func | 20% | 80% | | Peptides-struct | 77% | 23% | | ZINC | 44% | 56% | | MolHIV | 98% | 2% | Later iterations receive most of the weight on Peptides-func and ZINC; earlier iterations dominate on Peptides-struct and MolHIV. ## Notation and Figure 2 (Q1–Q2) We will introduce $G=(V,E)$ on page 2. Figure 2's y-axis shows the magnitude of the change in a target node's embedding after perturbing a source feature, on a logarithmic scale. We will specify the perturbation and averaging in the caption.

Thank you for the answers for my questions. I am happy with several of them, and I appreiate the results that show the effects of the different components. I can also increase my rating. I am still missing a deeper evaluation of the gating mechanisms (Q3 and Q5), so that it can be shown that it is not only the number of model parameters that increases the performance, but the mechanism itself.

Reviewer_Zq9jOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript investigates the effectiveness of graph implicit networks (GIMs) in capturing long-range dependencies in graphs. Specifically, the discrepancy between the theoretical unbounded receptive fields of GIMs and the seemingly meager empirical results wrt standard GNNs is subject of rigorous study. The work introduces a carefully constructed toy generalized (K,L)(K, L) chain-parity task to empirically study the modeling ability of GIMs as well as supporting theoretical discussion and analysis. They find that while GIMs have the capacity to propagate information from distant nodes, such an existence result does not lead to proper modeling in practice and indeed when distant signals must be more than just propagated but also combined or retained in non-trivial ways, such GIMs typically fail to model such behaviors.

To overcome these issues, they propose that propagating information and selecting representations for prediction should be treated as separate operations. This motivates a learned node-specific adaptive selection mechanism which enables prediction-relevant information in intermediate representations to be retrained before they become increasingly attenuated as GIMs reach a fixed point equilibrium and thus mitigate the aforementioned tension above. This is implemented as AdaGIM, which augments GIMs to two maintain for each node (1) a propagation state that continuously evolves during the optimization process, and (2) a selected state that retains information from earlier propagation iterations. Several other architectural and training features are also developed or adapted from other methods to enable stable performance and other desirable qualities. Theoretically, such a setup is shown to be desirable in a few respects, including convergence, and empirically the architecture generally results in better performance on real world and carefully constructed synthetic tasks.

Critical strengths
  1. The limitations of long-range information propagation in GIMs is characterized in a theoretically sound manner: (a) The aforementioned generalized chain-parity benchmark clearly distinguishes single-source propagation from multi-source parity computation. Moreover, Lemma 3.5 and Theorem 3.7 show a quantitative characterization of exponentially attenuating influence under appropriate contraction and Lipschitz assumptions. (b) Such a contribution is consequential -- it provides a way to reason about the discprepancy of theoretical unbounded receptive field size in GIMs and their seemingly contradictory inability to model complex long-range dependencies. (c) Theoretical analysis seems to be sound, properly qualified, etc. Moreover, certain analytical observations, such as the exponentially attenuating influence, serve as a motivation for architectural choices in the AdaGIM design.

  2. The inductive biases induced in the AdaGIM architecture are largely well-motivated and lead to interesting empirical results: (a) The explicit separation of signal representations due to ongoing implicit message passing and those retained for prediction is conceptually appealing and sound --- employing learned nodewise gates to select information from intermediate propagation states eschews the necessity of each node to maintain signal discrimation under the final equilibrium state restriction.

Critical weaknesses
  1. The theoretical diagnoses developed in Sections 2 and 3 do not seem to sufficiently establish that the primary mechanism of AdaGIM overcomes the identified limitation -- the vanishing influence analysis (ie Remark 3.6 and subsequent analysis) characterizes a limitation of uniformly contractive local equilibrium encoders. While AdaGIM does introduce a retained state with an input-dependent selection mechanism, the theoretical results dont seem to establish that such a construction construction would preserve a non-vanishing classification margin or, alternatively, fundamentally avoid the identified attenuation mechanism. Indeed, Appendix D seems to explicitly acknowledge that the convergence and uniqueness results do not provide such a guarantee.
Questions for authors
  1. Could the authors please clarify several experimental details such as the number of seeds that were used for experiments Table 1, how the reported means and error bars were computed, hyperparameter configuration, tuning budgets etc? There seems to be some contradictions or inconsistencies that may be also flagged in Appendix G. Quoting verbatim from the manuscript:

The experiment documentation gives conflicting seed counts for Table 1: five in its original caption, versus three for explicit baselines and a single seed (0) for validation-selected GIM configurations in the experimental setup. The setup also describes explicit baselines without Optuna tuning, whereas supplementary records describe six trials per model. Thus, the seed count, aggregation of the displayed spreads, and tuning allocation for this table remain unverified; complete selected configurations and run identifiers are unavailable

Am I reading this incorrectly or is there an explicit conflict in the experimental setup?

  1. Could the authors please provide a controlled comparison against stronger and/or modern intermediate-representation selection baselines, such as learned Jumping Knowledge aggregation or adaptive-depth graph networks?

  2. Could the authors please clarify what theoretical assumption allows AdaGIM's selected representation to behave differently from the encoders characterized by Theorem 3.7? Based on my read, theorem 3.7 establishes a vanishing-influence bound for uniformly contractive equilibrium encoders, whereas AdaGIM uses an initialization-dependent retained state and an input-dependent gate.

  3. Related to above, does the method improve primarily through the finite-length optimization and/or the representation retention, or is there evidence that it induces a change to the asymptotic distance-dependent sensitivity or classification-margin behavior?

[1] Xu, Keyulu, et al. "Representation learning on graphs with jumping knowledge networks." International conference on machine learning. pmlr, 2018. [2] Abbahaddou, Yassine, et al. "ADMP-GNN: adaptive depth message passing GNN." Proceedings of the 34th ACM International Conference on Information and Knowledge Management. 2025.

1 comment

Thank you for recognizing the theoretical contribution and the motivation for adaptive selection.

Q1: Experimental protocol

We reran all ten models with 20 tuning budgets (through Optuna) per dataset: 800 tuning attempts and 200 evaluation runs. All models use the same atom encoder and node features, without edge/bond features. Tuning uses seed 0 and validation selection. Selected configurations are evaluated on seeds 0–4, reporting the test score at the best validation epoch as mean ± sample standard deviation. We will release configurations, per-seed results, and logs.

Table R1. Recorded benchmark rerun. Five evaluation seeds; bold denotes the best mean.

ModelPeptides-func AP ↑Peptides-struct MAE ↓ZINC MAE ↓MolHIV AUC ↑
GCN0.4837 ± 0.00450.3334 ± 0.00060.4979 ± 0.00470.7669 ± 0.0068
APPNP0.4022 ± 0.00270.3323 ± 0.00060.7016 ± 0.00190.7105 ± 0.0105
GCNII0.4192 ± 0.00570.3243 ± 0.00090.5254 ± 0.00510.7747 ± 0.0026
IGNN0.4141 ± 0.00600.3297 ± 0.00070.5255 ± 0.01750.7451 ± 0.0088
EIGNN0.4004 ± 0.00820.3289 ± 0.00050.6152 ± 0.00690.7387 ± 0.0058
MGNNI0.3678 ± 0.00820.3551 ± 0.00250.6027 ± 0.00280.7511 ± 0.0096
GIND0.3850 ± 0.00800.3338 ± 0.00340.5038 ± 0.00950.7073 ± 0.0184
Monotone-MIGNN0.4661 ± 0.00250.3361 ± 0.00120.4882 ± 0.00550.7802 ± 0.0094
IGNN-finite0.5190 ± 0.00660.3230 ± 0.00090.4962 ± 0.00740.7414 ± 0.0282
AdaGIM0.5243 ± 0.00780.3135 ± 0.00810.4782 ± 0.01220.7404 ± 0.0357

AdaGIM achieves the best mean on both Peptides long-range benchmarks and ZINC.

Q3: Trajectory retention

Theorem 3.7 analyzes predictions from the propagation equilibrium H⋆H^\star. AdaGIM predicts from a selected state ZZ that retains trajectory information. At H⋆H^\star, the retained-state update becomes the identity, so the augmented map does not satisfy the theorem's strict-contraction assumption. For fixed initialization and contractive propagation, Proposition 4.1 and Appendix D establish convergence and finite total write mass, allowing earlier states to retain nonzero weight. Our controlled readout comparisons and within-model perturbation measurements assess the benefit of this retention at the finite distances studied.

Q4: Retention and distant influence

Within trained Peptides-struct models, the selected state has fitted per-hop influence retention 0.620 versus 0.474 for the final propagation state. Its normalized response stays above 10−310^{-3} through 13 hops versus 9. Table R4 in our reply to Reviewer wfbg connects selection to prediction under a common backbone configuration. Appendix D establishes that retained trajectory information can persist after propagation converges.

Q2: Stronger selection baselines

Table R4 in our reply to Reviewer wfbg adds scorer-matched JK-attention, JK-LSTM, and per-node ACT-style halting, with each variant retrained under a common backbone configuration and training schedule. The results favor last-write selection on both Peptides tasks. We will also expand the discussion of ADMP-GNN.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of rebuttal

Dear Area Chair and Reviewers,

Thank you for your time, and thank you to the Area Chair for the discussion reminder. As the author discussion period closes, we would like to highlight the evidence added in our November 11 responses and ask whether any concerns remain.

Experimental reliability: Table R1 in our response to Zq9j reports standardized reruns of all ten models, with 20 tuning attempts per model and dataset and five evaluation seeds. The response specifies the input features, validation-based selection, and aggregation protocol.

Controlled selection comparisons: Table R4 in our response to wfbg adds learned JK-attention, JK-LSTM, and ACT-style halting under a common backbone configuration and training schedule. AdaGIM achieves better mean performance than these three alternatives on both Peptides tasks; JK-attention also uses the same scorer architecture and essentially the same parameter count.

Direct evidence about distant influence: Our response to Zq9j compares the selected and final propagation states within trained Peptides-struct models. The selected state has fitted per-hop influence retention of 0.620 versus 0.474, with normalized response remaining above (10^{-3}) through 13 hops versus 9. This directly examines the retention concern raised in the reviews.

Theoretical scope and computational cost: Our theoretical results support convergence and retention of trajectory information under the stated assumptions. Table R5 quantifies the computational overhead and the memory–time tradeoff from checkpointing.

We would appreciate knowing which concerns, if any, remain after considering this additional evidence, so that we can provide any needed clarification before discussion closes.

Thank you for your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.