acceptodds
Under review as a conference paper at ICLR 2027

What Governs Unified Comprehension and Generation on Graphs? A Controlled Study

Abstract

Vision and language have converged on unified models that both understand and generate, yet graph pretraining usually discards its generative decoder and deploys the encoder alone. We ask what governs a single graph model's ability to support both tasks. A controlled evaluation of a compact graph autoencoder across five domains and eight training conditions separates three effects. First, the generation degradation seen with frozen decoders comes from refitting the latent prior on the fine-tuned encoder. It shrinks as a weight anchor limits encoder drift, and the tested generator families shift the trade-off less than this prior deployment on three of five domains. Second, a null-calibrated learned metric exposes generation discrepancies that descriptor MMD hides and changes which conditions are not significantly dominated. Third, the comprehension gain of retaining the decoder, significant only on Synthetic and CiteSeer, is reproduced by a weight anchor with no generative loss on Synthetic but not on CiteSeer. Only an active generative loss reduces the mismatch between the current encoder and the deployed decoder relative to frozen decoders, on every domain under descriptor MMD and outside QM9 under C2ST, without pinning the encoder to its pretrained weights.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Learning, Comprehension and Generation, Control Evaluation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Ji4kOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The authors note that the typical approaches to graph learning essentially segment the tasks of graph generation and graph comprehension by discarding the decoding head after training, then training a new decoder for classification based on the latent space embeddings. In contrast, vision and language tasts have moved toward unified models for comprehension and generation. The focus of the authors work it to evaluate the whether the seperation of graph generation and comprehension is the ``correct'' choice. To evaluate this, the authors consider a variety of different approaches to the combined comprehension/generation tasks, mimicking those which appear in the literature, while using a common architecture for the auto-encoder training and comprehension heads. These networks are applied to 5 different sets of graphs of at most 24 vertices (synthetic stochastic block model, CiteSeer, Cora, PubMed, and QM9).

Critical strengths
  1. The question the authors attempt to address in interesting and important.

  2. The experimental design is thorough, well thought-out, and well documented.

Critical weaknesses
  1. In multiple places, the use of a non-specific referent (e.g., ``it'') compromises the clarity of the presentation to the point that several sections are incomprehensible.

  2. While there are numerous experiments presented in this work, representing the authors commitment to the scientific questions, due to the space limitations the analysis and interpretation of the often reduces to highlighting single values from large tables, rather than extracting broad themes and addressing how the results answer the stated research questions.

Questions for authors

None

2 comments

We thank the reviewer for finding the question interesting and important and the experimental design thorough, well thought-out, and well documented. We address your concerns one by one below.

W1 (Non-specific referents such as "it"). The revision replaces each non-specific referent in the main text with the noun it refers to. Three examples follow.

  • Sec. 4.3. "It matches Seq-reg on Synthetic and PubMed ..." now reads "The stronger anchor (Seq-anchor at λ_anc = 1) matches Seq-reg on Synthetic and PubMed ..."
  • Sec. 3. "It then fine-tunes for comprehension" now reads "The schedule then fine-tunes ...", "This tests whether Phase-1 pretraining matters" now reads "Joint-retained tests whether ...", and "It uses the same fixed Phase-1 decoder" now reads "Seq-anchor uses ...".
  • Abstract. "It shrinks as a weight anchor limits encoder drift" now reads "This degradation shrinks as a weight anchor limits encoder drift".

W2 (Single values from large tables rather than broad themes). The revision gives the verdict of every research question in one table and moves the cell-level values to the appendix.

  • One summary table. Table 8 gives the verdict of every research question on every domain, with a check mark where the paired CI is significant in the stated direction.
  • Cell-level values move out. Values that do not change a verdict are now in App. D.13.

The results answer the research questions through four themes, which Sec. 5 discusses.

  1. A generator is a prior plus a decoder, and under a frozen decoder the prior is what breaks (RQ2). Fine-tuning moves the encoder while the frozen decoder still expects the old latents, so re-estimating the prior on the moved encoder hurts generation. Reusing the prior saved before fine-tuning lowers C2ST significantly on all five domains (Table A.20). Any report of frozen-decoder generation should therefore state which prior was used.
  2. Comprehension gain and generation capability are separable (RQ3). Against discarding the decoder, keeping the generative loss improves accuracy significantly on Synthetic (+0.080) and CiteSeer (+0.046). An L2-SP anchor, which computes no generative loss, obtains the Synthetic gain without generative ability (Sec. 4.3).
  3. Only the active generative loss keeps encoder and decoder compatible without pinning the encoder (RQ4). Any encode-then-decode use needs the latents of the current encoder to decode well. Round-trip MMD is 0.003 to 0.273 for Seq-reg against 0.325 to 0.561 for Seq-freeze (Table 6), and a frozen decoder with the generative loss kept active reaches 0.008 to 0.279, so the compatibility comes from the generative gradient reaching the encoder (App. D.9).
  4. The generation metric changes the conclusions (RQ1). Descriptor MMD ranks a random Erdős-Rényi sampler second only to the generation-only model on QM9 (0.245 against 0.236 to 0.242). A classifier two-sample test calibrated against that random-graph baseline exposes the gaps MMD hides.

RQ5 then asks whether the generator family matters more than these interfaces. The tested prior families, a latent-diffusion prior and a VAE, shift C2ST less than prior deployment on three of five domains (Table A.22). The conclusion ends with the practical recommendations that follow from these themes.

  • Report the deployed prior with any frozen-decoder generation result.
  • Pair a decoder frozen after pretraining with the cached prior, and a decoder trained during fine-tuning with a refit prior.
  • Keep the generative loss active when an L2-SP anchor does not reproduce its classification gain, and whenever encoded inputs must be decoded.
  • Check compatibility by round-trip reconstruction, not by linear CKA.
  • Score generators with a learned discriminator against a random-graph baseline.
AuthorsOpenReview ↗

Kind Follow Up

Dear Reviewer Ji4k,

With the discussion deadline approaching (November 18 AoE), we wanted to briefly follow up on our rebuttal. We revised the wording for clarity and reorganized the results around the research questions. If you have a chance to take a look, we would appreciate hearing whether these changes address your concerns. Thank you again for your time and feedback.

Reviewer_tYQ8OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates the trade-off between graph comprehension and generation using a graph autoencoder across five domains and eight training conditions. It examines the effects of latent prior refitting, encoder drift, and decoder retention.

Critical strengths

The paper addresses a relevant problem in unified graph comprehension and generation.

Critical weaknesses
  1. Discards generative decoder and deploys encoder : No details are provided for this statement. The task is not clear for this statement.

  2. Abstract is not clear: Lacks motivation and explanation

  3. “Anchoring”, “active generative loss” is not explained in the introduction but mentioned in the contributions/findings. Meaning and its significance remains unclear to the reader.

  4. Citations and explanation missing for MMD [1] and C2ST [2] in Introduction.

    [1] Gretton, A., Borgwardt, K.M., Rasch, M.J., Schölkopf, B. and Smola, A., 2012. A kernel two-sample test. The journal of machine learning research, 13, pp.723-773.

    [2] David Lopez-Paz and Maxime Oquab. Revisiting Classifier two-Sample Tests. International Conference on Learning Represenatations, ICLR 2017

  5. Minor edit errors: Fig 1 overlaps with its caption.

  6. Paper not clearly written. Many terms are not explained or introduced before its mention.

  7. Line#138-139: Not clear : “Write Lcomp for the cross-entropy classification loss and Lgen for the graph reconstruction loss defined below.”

  8. What are the three generative mechanisms mentioned in line#172? SEQ-REG schedule is not defined in this section or prior to this section.

  9. Line#198: L2-SP anchor is not defined and not cited.

  10. Table 1 results are not clear.

  11. Datasets used for this study is on small graphs dataset.

Questions for authors

see above

4 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

We thank the reviewer for recognizing that the paper addresses a relevant problem in unified graph comprehension and generation, and for the detailed and specific points, which helped us improve the presentation. We address your concerns one by one below.

W1 (Discards generative decoder and deploys encoder). The statement refers to the standard pipeline of graph self-supervised learning, in which the decoder is used only during pretraining.

  • GraphMAE (Hou et al., 2022) pretrains a GNN encoder together with a decoder that reconstructs masked node features.
  • After pretraining, the decoder is deleted, and only the encoder is fine-tuned for downstream graph classification.
  • GPT-GNN (Hu et al., 2020) follows the same pattern with attribute and edge prediction.

The deployed model can therefore classify graphs but cannot generate them. The paper studies one model that keeps both abilities with one set of weights, classifying graphs and sampling new graphs that match a reference distribution. The revised abstract and Sec. 1 now state this pipeline and both tasks.

W2 (Abstract is not clear, lacks motivation and explanation). The abstract now opens with the motivation, that graph pretraining discards its generative decoder so the deployed model can no longer generate. It names the two tasks and what the eight conditions vary, and explains the prior, the learned metric, and the active generative loss where they appear. It reads as follows.

Vision and language have converged on unified models that both understand and generate, yet graph pretraining usually discards its generative decoder and deploys the encoder alone, so the deployed model can classify graphs but no longer generate them. We ask what governs a single graph model's ability to support both tasks, classifying graphs and sampling new graphs from one set of weights. A controlled evaluation of a compact graph autoencoder across five domains and eight training conditions, which differ only in how fine-tuning treats the decoder, and repeated with larger encoders, 128-node graphs, and a diffusion decoder, separates three effects. First, the generation degradation seen with frozen decoders comes from refitting the latent prior, the distribution from which latents are sampled, on the fine-tuned encoder. This degradation shrinks as a weight anchor limits encoder drift, and the latent-diffusion and VAE families shift the trade-off less than this prior deployment on three of five domains. Second, a null-calibrated learned metric, a classifier two-sample test scored against a random-graph baseline, exposes generation discrepancies that descriptor MMD hides and changes which conditions are not significantly dominated. Third, in the primary grid, the comprehension gain of retaining the decoder, significant only on Synthetic and CiteSeer, is reproduced by a weight anchor with no generative loss on Synthetic but not on CiteSeer. Only an active generative loss, which keeps the reconstruction loss during fine-tuning, reduces the mismatch between the current encoder and the deployed decoder relative to frozen decoders, on every domain under descriptor MMD and outside QM9 under C2ST, without pinning the encoder to its pretrained weights, and under descriptor MMD it does so in every added setting.

W3 ("Anchoring" and "active generative loss" not explained in the introduction). Both are now defined in a "Terms used throughout" paragraph in Sec. 1, before the findings list.

  • Anchoring (Seq-anchor) adds the L2-SP penalty λ_anc‖θ − θ⁽¹⁾‖² (Li et al., 2018) during fine-tuning, which pulls the encoder weights θ toward their pretrained values θ⁽¹⁾ without any reconstruction loss. Its significance is that it tests whether the benefit of keeping the decoder can be obtained by weight regularization alone.
  • An active generative loss (Seq-reg) keeps λ·L_gen in the fine-tuning objective, so the reconstruction loss keeps training the encoder and the decoder while the classifier is learned. Its significance is that, among the tested conditions, only an active generative loss keeps the fine-tuned encoder compatible with the decoder without pinning the encoder to its pretrained weights (Sec. 4.4).

W4 (Citations and explanation missing for MMD and C2ST). Sec. 1 now cites both at first mention, with one explanatory clause each.

  • Descriptor MMD is the kernel two-sample distance of Gretton et al. (2012) between degree, clustering, and spectral statistics of generated and reference graphs, as is standard for graph generation (You et al., 2018).
  • C2ST is the classifier two-sample test of Lopez-Paz and Oquab (2017), the held-out accuracy of a fresh graph network trained to tell generated from real graphs. A score of 0.5 means the two sets cannot be told apart.
AuthorsOpenReview ↗

Rebuttal (2/2)

W5 (Fig. 1 overlaps with its caption). Fig. 1 no longer overlaps its caption. The revision removes the negative vertical spacing around the figure.

W6 (Terms not explained or introduced before mention). Every technical term in Secs. 1 to 4 is now defined at first use or points to its definition, including prior, refit, cached, anchoring, active generative loss, C2ST, random-graph baseline, real-versus-real floor, update-matched, and "not significantly dominated". The new Table 1, at the start of Sec. 3, lists the eight conditions in one line each.

ConditionPhase 1Phase-2 decoderPhase-2 loss
C-onlynonenoneL_comp
G-onlynonetrainedL_gen (no classifier)
MAE-discardreconstructiondiscardedL_comp
Seq-freezereconstructionfrozenL_comp
Seq-anchorreconstructionfrozenL_comp + λ_anc·L_anc
Seq-regreconstructiontrainedL_comp + λ·L_gen
Naive-Decoupledreconstructiontrained, separate readoutsL_comp + λ·L_gen
Joint-retainednonetrained from scratchL_comp + λ·L_gen

W7 (Line 138 to 139 not clear). The sentence now reads "Let L_comp denote the cross-entropy loss of the classification head and L_gen the reconstruction loss of Eq. (1)."

W8 (Three generative mechanisms in line 172, Seq-reg schedule not defined). The three generative mechanisms are three ways of sampling the latent z before decoding.

  1. A Gaussian fitted to the encoded latents, which is the default.
  2. A latent diffusion model trained on the same latents and decoded by the same decoder, so it differs from (1) only in the prior.
  3. A variational bottleneck (VAE) trained with a KL term, which changes how the latent itself is parameterized.

The Seq-reg schedule is reconstruction pretraining followed by fine-tuning on L_comp + λ·L_gen. The paragraph now opens with "We compare three ways of sampling the latent z before decoding", states the schedule in place of its name, and follows the new Table 1, which defines every condition.

W9 (Line 198, L2-SP anchor not defined and not cited). L2-SP is the penalty λ_anc‖θ − θ⁽¹⁾‖² that pulls the fine-tuned encoder weights θ toward their pretrained values θ⁽¹⁾ (Li et al., 2018). The revision now defines and cites it at first use in Sec. 1 and again at line 198 in Sec. 3.2.

W10 (Table 1 results are not clear). Each cell of Table 1, which is Table 2 in the revision, is a C2ST score, where 0.5 means generated and real graphs cannot be told apart, 1.0 means they are trivially separable, and lower is better. The last column is a random-graph baseline, an Erdős-Rényi sampler with matched sizes and edge density. A generator that scores below this baseline captures structure beyond graph size and edge density. The revised table groups the columns by condition type (generation-only reference, active generative loss, frozen decoder, decoupled diagnostic, baseline), puts in bold every cell significantly below the baseline using the paired CIs of Table A.18, and states this reading rule in the caption.

  • On Synthetic, the conditions that keep the generative loss come close to real graphs (Seq-reg at 0.538) and beat the baseline (0.772). Frozen decoders with a re-estimated prior do not (0.802 to 0.837), unless a strong anchor keeps the encoder near its pretrained weights (0.569).
  • On the citation graphs, the same conditions beat the baseline and stay between 0.72 and 0.85.
  • On QM9, every condition scores 0.96 or higher, a gap that descriptor MMD does not show (Sec. 4.1).

W11 (Datasets used are small graphs). The paper's two interface findings, on prior deployment and on encoder-decoder compatibility, also hold on larger graphs. The 24-node budget serves the controlled grid of 8 conditions × 5 domains × 10 paired seeds, with every condition sharing one encoder. The paper already includes 64-node ogbn-arxiv ego-graphs (App. D.7), where Seq-reg reaches accuracy 0.310 against 0.275 for C-only and generation MMD 0.162 against 0.605 for Seq-freeze. The revision adds all eight conditions on the PROTEINS graph-classification benchmark with a 128-node budget, five times the main grid, with 10 paired seeds and the same tests (App. D.12, Tables A.31 and A.32).

  • Data. 1,113 graphs in 2 classes, with 39.1 nodes on average and at most 620, so that 3.9% of graphs exceed 128 nodes.
  • Refit penalty. Reusing the saved prior lowers the frozen decoder's C2ST by 0.178 [0.091, 0.283].
  • Compatibility. Round-trip MMD is 0.858 [0.794, 0.906] lower with the active generative loss than with the frozen decoder.

I read the author’s response and have revised my score.

Dear Reviewer,

Thank you for reading our response and updating your rating. Could you share which concerns remain after our rebuttal and revision, and which of them keep you from recommending acceptance? We would be happy to clarify any remaining points.

Reviewer_BT76OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies unified comprehension and generation on graphs. They study what parts matter for a graph model's capability to perform well on both generation and downstream classification. By conducting a controlled study of different setup version between encoders, decoders, and classifiers, they study how each connection between them affects the overall performance on generation and classification. They show that degradation from frozen decoders arises from latent prior mismatch. They also study the improvements in retaining the generator can improve performance in some datasets. They show their results on different graph datasets and study the different settings.

Critical strengths

Here are some strengths of the paper in my opinion:

  1. The paper studies an interesting idea of training a graph model that is good in both understanding and generating graphs?

  2. In their study, the analyze different research questions related to graph autoencoders, transfer learning, and finetuning by dividing the study into different phases where phase 1 focuses on generative pretraining and phase 2 focuses on comprehension fine-tuning

  3. They show that a single model can support both generation and encoding for classification. They also show that generative training can help improve classification performance in some cases and latent prior mismatch could be bad for generation.

Critical weaknesses

Here are some weaknesses of the paper in my opinion:

  1. The paper is not well written and makes reading the paper very difficult.

  2. The architectures present in the paper are relatively simple.

  3. The paper studies unified graph comprehension and generation but the discussion of graph autoencoders that use diffusion models as their decoders are not well studied [1]. Could the authors generalize this by adding works like [1]?

References

  1. Wesego, Daniel. "Graph Representation Learning with Diffusion Generative Models." arXiv preprint arXiv:2501.13133 (2025).
Questions for authors

Please look at the weaknesses section

4 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

We thank the reviewer for finding the idea of training one graph model that is good in both understanding and generating graphs interesting, and for recognizing that a single model can support both generation and classification and that latent prior mismatch can be bad for generation. We address your concerns one by one below.

W1 (The paper is not well written). The revision makes the paper easier to read in four places.

  • Abstract. The abstract now states the motivation and the two tasks, classifying graphs and sampling new graphs from the same weights, and explains the prior, the learned metric, and the active generative loss where they first appear.
  • Terms. A "Terms used throughout" paragraph in Sec. 1 defines the prior, the refit and cached priors, anchoring, the active generative loss, C2ST, and the random-graph baseline before the findings use them. The new Table 1 lists the eight training conditions in one line each.
  • Results. The new Table 8 gives every verdict per domain, and cell-by-cell readings of the tables moved to App. D.13.
  • Wording and layout. Non-specific "it" and "this" are replaced by the nouns they refer to, and Fig. 1 no longer overlaps its caption.

W2 (The architectures are relatively simple). The paper's findings on prior deployment and on encoder-decoder compatibility hold with larger and more expressive encoders. With a 19-times larger GIN both findings hold on all five domains, and with a GPS graph transformer on all three tested domains, the refit penalty on QM9 under generation MMD.

The original model is compact by design. It lets 8 conditions × 5 domains × 10 paired seeds share one encoder and one parameter count, which is what makes the paired, condition-by-condition comparisons possible. The two larger encoders keep the decoder, data, budgets, and tests unchanged.

  • Wide and deep GIN. Hidden width 256 instead of 64, 5 layers instead of 3, and latent size 32 instead of 8, for 1.05M parameters, 19 times the original. All five domains, 10 seeds.
  • GPS graph transformer (Rampášek et al., 2022). Each layer combines a GIN update with global self-attention over the nodes, for 0.76M parameters. Synthetic, CiteSeer, and QM9, 10 seeds.

The table counts the domains where each finding holds under paired-bootstrap 95% CIs.

FindingOriginal GIN (56K)Wide GIN (1.05M)GPS (0.76M)
Reusing the saved prior removes the refit penalty of the frozen decoder (RQ2)5/55/52/3 in C2ST, 3/3 in generation MMD
The active generative loss keeps encoder and decoder compatible (RQ4)5/55/53/3
Seq-reg's generator beats the random-graph baseline (RQ1)4/5 (all but QM9)4/5 (all but QM9)2/3 (all but QM9)
  • Refit penalty. With the wide GIN, the rise in C2ST when the prior is re-estimated instead of reused ranges from +0.375 [+0.312, +0.435] on Synthetic to +0.052 [+0.022, +0.098] on QM9. With GPS it is +0.345 on Synthetic and +0.146 on CiteSeer, and on QM9 generation MMD rises by +0.178 [+0.099, +0.264].
  • Compatibility. Round-trip MMD with the active generative loss minus the frozen decoder lies between −0.738 and −0.284 for the wide GIN and between −0.630 and −0.173 for GPS, all significant.

App. D.11 reports both encoders in full.

AuthorsOpenReview ↗

Rebuttal (2/2)

W3 (Graph autoencoders that use diffusion models as their decoders, works like [1]). The revision generalizes the study to such autoencoders, both in the discussion and in the experiments.

DDGAE [1] trains a graph autoencoder whose decoder is a discrete diffusion model over the adjacency, conditioned on the encoder output, and evaluates the learned embeddings with an SVM on PROTEINS and IMDB-B. Its decoder serves representation learning, without a prior over latents, sampling of new graphs, or a generation evaluation. Sec. 2 and App. A now discuss DDGAE together with the published work this line builds on, namely diffusion autoencoders (Preechakul et al., CVPR 2022), which pair an encoder with a conditional diffusion decoder, directional diffusion models for graph representation learning (Yang et al., NeurIPS 2023), and D3PM (Austin et al., NeurIPS 2021).

App. D.10 places a DDGAE-style decoder in our dual-use protocol, where the generator must be sampled from a prior and scored. The bilinear edge head is replaced by a discrete diffusion decoder over the adjacency (absorbing to the empty graph in T = 32 steps, a UNet denoiser conditioned on z and the step, 403,305 parameters), trained with the x0 cross-entropy term of the D3PM loss. Everything else is fixed, and generation-only, MAE-discard, Seq-freeze, Seq-anchor, and Seq-reg run on all five domains with 10 paired seeds (Tables A.27 and A.28).

  1. Encoder-decoder compatibility (RQ4) carries over. The active generative loss beats the frozen decoder in round-trip MMD on 5/5 domains and in round-trip C2ST on 4/5, the same pattern as with the bilinear decoder.
  2. The classification gain of Seq-reg over MAE-discard (RQ3) appears on 4/5 domains (Synthetic, CiteSeer, Cora, QM9), against 2/5 with the bilinear decoder.
  3. The dual-use model generates better. The diffusion decoder lowers Seq-reg's C2ST by 0.101 to 0.166 on CiteSeer, Cora, PubMed, and QM9 (Table A.28). On QM9 it beats the random-graph baseline, 0.797 against 0.975 (difference −0.178 [−0.245, −0.115]), which the bilinear Seq-reg does not.
  4. The refit penalty of the frozen decoder (RQ2) persists. Re-estimating the prior raises generation MMD significantly on 4/5 domains, so the frozen diffusion decoder still generates differently depending on which latents it receives. Under C2ST the penalty is significant on Synthetic (+0.073 [+0.012, +0.130]) and Cora (+0.059 [+0.027, +0.091]).
  5. Generator family (RQ5). Because this decoder generates better, swapping it in moves C2ST more than prior deployment does on Cora, PubMed, and QM9, by 0.063 [0.027, 0.102], 0.067 [0.031, 0.102], and 0.131 [0.058, 0.206] (Table A.28). App. D.10 reports this decoder-family comparison, next to the prior families of Sec. 4.5.

[1] D. Wesego. Graph Representation Learning with Diffusion Generative Models. arXiv preprint 2501.13133, 2025.

Reviewer_BT76OpenReview ↗

Reply to Authors

Dear Authors,

Thank you for your reply. Your reply has addressed some of my concerns. I hope you will still work on improving the readability of the paper if accepted for the camera-ready version. I have increased my score.

Dear Reviewer BT76,

We sincerely thank you for your constructive feedback and for increasing your score. We appreciate your recognition of our efforts to address the concerns. We will further improve the readability and overall presentation of the paper in the revision.

Sincerely,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.