acceptodds
Under review as a conference paper at ICLR 2027

GRAPHVQ: STRUCTURE-AWARE AUTOREGRESSIVE DECODING OVER CONTEXT-QUANTIZED GRAPH TOKENS

Abstract

Graph foundation models need a discrete token representation, but casting a graph as a generatable token sequence faces a structural obstacle: edges spanning beyond the serialization window cannot be emitted in one pass—so one-pass autoregressive generators systematically under-produce cycles—and a single global condition cannot tell candidate edges apart. GRAPHVQ removes both obstacles: node contexts—features plus a local edge mask under multi-order breadth-first serialization—are quantized into a shared codebook by a VQ-VAE with BCEcalibrated Bernoulli edge decoding, and a second-stage structure-aware decoder emits the global adjacency conditioned on token-derived pair features, whose necessity over any global-summary condition is formalized in a scoped impossibility result. The tokenizer reconstructs node features at 0.86–0.99 accuracy and decodes local edges at AUROC ≥ 0.89 (ECE ≤ 0.007). Under one same-split protocol on four datasets, pair conditioning improves orbit MMD 0.248→0.174 on PROTEINS and 3.4× on a ring stress test, and vanishes on a random-label control—the signature of attribute–topology coupling—so the gain is claimed exactly where attributes carry edge-relevant signal. GRAPHVQ ranks first among learned generators on PROTEINS, ties for first on SYN-COMM, and improves orbit MMD 2.7–17× over one-stage generation on three datasets, with seed-level bootstrap intervals confirming the rankings are not seed noise; on MUTAG the unweighted edge target under-generates and is reported as such. These results locate the structural control of autoregressive graph generation in the granularity of the condition: pair-level token context turns a quantized vocabulary into a usable capacity axis for distribution-faithful graph generation and future token-level pretraining.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph generation, vector quantization, autoregressive models, graph tokenization, graph foundation models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_q6UqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GraphVQ, a discrete token-based autoregressive framework for graph generation. The method first employs BFS serialization and VQ-VAE to encode node attributes and local structures into discrete tokens, then uses a GRU to model the token sequence. Finally, token-derived pairwise features are used to guide the generation of the full adjacency matrix, mitigating the limitations of local windows in capturing long-range edges. Experiments on four graph datasets demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The two-stage generation framework helps alleviate the limitations of local serialization windows in generating long-range edges, offering practical value.

  2. The experiments include multiple baselines and ablation studies, demonstrating competitive generation quality on several datasets.

Critical weaknesses
  1. Limited Novelty. The proposed method primarily combines existing techniques, including BFS serialization, VQ-VAE, GRU, and chunk-wise autoregressive generation. The core Pair-wise Conditioning mechanism also relies on conventional feature construction and mean pooling, lacking sufficient methodological novelty. Furthermore, pair-wise features within each chunk are averaged, meaning that the conditioning is not strictly performed at the pair-wise level. This further weakens the claimed novelty of the approach.

  2. Insufficient Validation of Core Contributions. The continuous baseline changes both the latent representation type and the prior family, making it difficult to isolate the contribution of VQ. Although the paper includes several conditioning ablations, these experiments primarily demonstrate the effectiveness of conditioning information, without sufficiently establishing the unique advantages of the proposed chunk-pooled decoder over alternative structure-aware edge generation mechanisms.

Questions for authors
  1. In Stage B, pair-wise features within the same chunk are mean-pooled, resulting in a shared conditioning representation. How does the model distinguish the edge probabilities of different node pairs under this setting? What specific role do the position-specific readout heads play in this process?

  2. The paper reports that unweighted BCE leads to significant under-generation of edges on MUTAG, while positive-class weighting substantially improves connectivity. Do the authors attribute this primarily to graph sparsity, class imbalance, or distribution mismatch between the token prior and the edge decoder?

4 comments
AuthorsOpenReview ↗

Response to Reviewer q6Uq (1/3)

We thank the reviewer for the two precise questions — they target exactly the two places where the implementation and the narrative needed to be pinned down, and the matched controls below answer them.

Q1: How does the model distinguish edge probabilities within a chunk?

Equation (9) computes a masked mean of the projected pair features rijr_{ij} over the valid positions of chunk qq; the GRU combines this chunk summary with the previous chunk's bits and its hidden state into a shared chunk state sqs_q. Equation (10) then applies a position-specific readout vector wbw_b to sqs_q: the BB candidate slots of the chunk receive logits wb⊤sqw_b^\top s_q and are sampled independently. The design is deliberately a chunk-level shared condition with position-wise Bernoulli outputs.

Two diagnostics confirm this reading:

  • Permuting the pair features within a chunk changes the logits by at most 4.8×10−74.8\times10^{-7} (Sec. 4.3) — the condition is chunk-granular by construction.
  • The position-wise readouts are not degenerate: over held-out MUTAG graphs, the within-chunk predicted-probability standard deviation is 0.0450.045 on average (0.0650.065 at the 90th percentile), and predicted probabilities correlate with the normalized node-pair distance (j−i)/n(j-i)/n at r=−0.294r=-0.294 (19 graphs, one checkpoint).

The revision uses the term chunk-pooled pair features throughout and adds an algorithm box for the Stage-B decoder.

AuthorsOpenReview ↗

Response to Reviewer q6Uq (2/3)

Weakness 2: What do VQ and pooling each contribute?

We agree the Continuous control in the submission changes the latent type and the prior family together (an autoregressive Gaussian prior instead of the categorical GRU), so it cannot isolate quantization. We therefore ran two matched suites under the revision's MUTAG configuration (density-targeted Stage B, w+=3w_+{=}3, temperature 1.0, same splits, 10k samples, seeds 0–4):

(a) Code-learning control. The VQ tokenizer is replaced by a same-dimension (32) continuous autoencoder whose latent is discretized post-hoc by k-means into the same K=32K{=}32 code ids, keeping the categorical GRU prior, the Stage-B decoder, and the training budget fixed (continuous_cat). This isolates the code-learning mechanism — joint VQ training vs. post-hoc clustering — while holding the discrete interface itself fixed.

MUTAG (density-targeted, 5 seeds)degreeclusteringorbitspectralcomp. sizeconn.
GraphVQ (VQ + pooled)0.135±0.0120.845±0.0260.359±0.0530.092±0.0090.085±0.0320.653
continuous_cat0.139±0.0150.844±0.0050.393±0.0260.086±0.0100.061±0.0240.703
continuous0.180±0.0400.898±0.0120.521±0.1170.104±0.0060.002±0.0020.821
Raw0.138±0.0120.851±0.0090.392±0.0500.101±0.0190.075±0.0200.704

On MUTAG, no significant difference was detected between joint VQ training and post-hoc clustering on any metric (orbit p=.20p=.20, degree p=.68p=.68; paired tt, 5 seeds); this does not establish equivalence. The three discrete-interface variants have similar observed degree MMDs (0.135–0.139); the continuous-prior control is nominally worse on orbit (0.521 vs. 0.359/0.393; paired p=.07p=.07). The one-stage, condition, and mask ablations separately support the contributions of token-derived conditioning and the Stage-B decoder. The matched code-learning comparison is limited to MUTAG and holds the discrete interface fixed; it does not isolate discretization vs. no discretization, since the continuous control also changes the prior family.

(b) Decoder comparison. Three alternative structure-aware decoders sharing the tokenizer, prior, candidate order, chunk width, and budget (density-targeted on MUTAG; unweighted tables for all four datasets are in the revision's Appendix C):

MUTAG (density-targeted, 5 seeds)degreeorbitconn.paramssample s/1k
Pooled (ours)0.1350.3590.65320.7K17.1
Per-pair0.1580.2910.45420.5K18.8
Within-chunk AR0.1680.3300.49326.9K42.7
GraphRNN-style0.1470.3290.73233.6K8.8

On MUTAG, no significant advantage over pooled was detected on orbit or degree (paired tt on orbit: p=.12/.55/.55p=.12/.55/.55 for per-pair/within-AR/GraphRNN-style). Pooled has the lowest observed degree MMD and sampling cost among the pair-conditioned heads; alternatives improve clustering/spectral, per-pair has lower observed orbit MMD (0.2910.291 vs. 0.3590.359, p=.12p=.12), and GraphRNN-style has higher connectivity (0.7320.732 vs. 0.6530.653). Pooled also has the lowest orbit MMD on the two largest benchmarks (PROTEINS 0.1750.175 vs. 0.2150.215–0.3370.337; SYN-COMM 0.0090.009 vs. 0.0100.010–0.0190.019). These quality–cost trade-offs motivate our choice; the full metric-level table is in Appendix C. (Training batch size 32 for all arms on MUTAG; the within-AR arm uses batch size 1 only for the largest PROTEINS graphs, which the revision states explicitly.)

AuthorsOpenReview ↗

Response to Reviewer q6Uq (3/3)

Response to Reviewer q6Uq (3/3)

Q2: Is MUTAG under-generation sparsity, class imbalance, or prior–decoder mismatch?

Three facts, without a claimed causal ordering among them:

  1. Class imbalance is structural. Positive bits are 12.1%12.1\% of all edge bits (384/3171); under true conditions the positive/negative conditional NLL is 1.831.83 vs. 0.1190.119 nats/bit — the positive class has an observed higher NLL. These values come from a single-checkpoint, unweighted diagnostic, not five-seed statistics or a calibration assessment.
  2. Density targeting is the principled response. With positive-class weight ww, the weighted-BCE optimum is q∗=wp/(wp+1−p)q^*=wp/(wp+1-p): w=3w{=}3 shifts the per-bit target for a p=0.1p{=}0.1 edge to 0.250.25, raising the sampled edge rate. The measured effect matches the intent: mean degree of generated MUTAG graphs rises to 2.922.92 vs. 3.193.19 in the training reference. The weight was selected in the diagnostic phase from w+∈{1,3,5}w_+\in\{1,3,5\} by training-reference structural metrics (Table C.3), not on test metrics. This is a design choice for imbalanced edge prediction: the revised main table uses it on MUTAG (connectivity 0.109→0.6530.109\to0.653, orbit MMD 0.598→0.3590.598\to0.359, mean rank 4.75→2.604.75\to2.60, five seeds), and the probability-oriented unweighted variant with validation temperature calibration is reported alongside in Appendix C.
  3. The decoder tracks its conditions; size is set by the prior. The oracle → reconstructed → generated NLL gap on MUTAG is 0.3713→0.3713→0.37670.3713\to0.3713\to0.3767 (unweighted diagnostic run, seed 0): the observed average gap is small. The condition-dependence diagnostic (Fig. B.1) confirms the condition is used (+13.6% positive-edge NLL under global shuffle). Separately, generated graphs average 13.113.1 nodes against a resampled target of ≈17.8\approx 17.8 because the token prior emits EOS earlier than the empirical length distribution; on MUTAG (max 28 nodes, below the generation cap of 64) the size gap therefore comes from the prior's EOS behavior, not from the edge decoder (Algorithm 1 now states that the sampled NN is an upper bound), and the revision reports this decomposition explicitly. (The NLL decomposition and the position-head diagnostic were computed under the unweighted diagnostic configuration; their findings are limited to that configuration.)

Thanks for the author's reply. The newly added control experiments provided valuable evidence for understanding the contribution of VQ. But I still have concerns about the novalty of the method. I maintain my score.

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

The paper proposes GRAPHVQ, a three-stage framework that combines VQ-VAE based graph tokenization, a GRU autoregressive prior, and two stage generation with token derived pair features. The method aims to overcome serialization window limits and preserve global adjacency structure. Experiments on four datasets evaluate connectivity, structural MMD, and generation novelty.

Critical strengths

The paper addresses a relevant challenge in autoregressive graph generation by incorporating structural information through token-derived pairwise conditioning. The experiments include multiple baselines, and various metrics to assess generation quality.

Critical weaknesses
  1. Evaluation: The experiments cover only two real datasets, MUTAG and PROTEINS, with a maximum graph size of 64 nodes limiting its applicability to generalize across other datasets including generic domain graphs.
  2. The framework largely combines established techniques, including VQ-VAE based quantization, BFS serialization, GRU-based autoregressive modeling, and Bernoulli edge decoding. While their integration is reasonable, the introduction does not identify a sufficiently novel technical contribution beyond combining these existing components.
  3. Results: The Table1 do not demonstrate consistent superiority across datasets. GRAPHVQ achieves a relatively poor mean rank of 4.75 on MUTAG, compared to 2.95 for GraphRNN, and is outperformed by the simple Raw baseline on SYN-RING-ROLE (2.90 vs. 2.65), while tying it on SYN-COMM (2.40). Improvements are observed on PROTEINS, however, the overall results questions generalizability and advantage over the added architectural complexities.
  4. The paper does not report variance across the runs questioning the stability of the proposed method in comparison of the baselines.
  5. It would be beneficial to visualize the quality of the generated graphs in contrast to the ground truth/test dataset for the given method as well as compared to the baselines. Moreover, it is unclear from the visualization in the appendix as to why the two-stage contains too many disconnected components.
  6. Time and resource analysis are missing for the proposed approach.
Questions for authors

see above

3 comments
AuthorsOpenReview ↗

Response to Reviewer YEuL (1/2)

We thank the reviewer for the thorough assessment. We address each weakness with new evidence, and we hope the revised main-table protocol (MUTAG now uses the density-targeted objective, matching its 12.1% positive-edge rate) also clarifies the results picture.

W1: Two real datasets, ≤64 nodes

We added two real attributed-graph benchmarks — ENZYMES (600 graphs, up to 126 nodes) and NCI1 (4110 graphs, up to 111 nodes) — plus a 128-node arm covering both without truncation. The paper's central mechanism holds on both:

DatasetOne-stage orbit MMDGraphVQ orbit MMDPaired test (5 seeds)
ENZYMES0.469 ± 0.0470.210 ± 0.068p=.0012p=.0012
NCI10.749 ± 0.0140.475 ± 0.032p<.001p<.001

On ENZYMES, no significant difference in orbit MMD was detected between GraphVQ (0.2100.210) and selected strong controls (Raw 0.1910.191, p=.61p=.61; GraphARM 0.2020.202, p=.90p=.90); on NCI1, no significant difference was detected against Raw (0.4260.426, p=.063p=.063) or GraphARM (0.4170.417, p=.45p=.45), while GraphRNN is ahead on orbit there (0.3330.333, p=.0003p=.0003). Orbit MMD remains roughly unchanged in the 128-node arm (ENZYMES 0.210→0.2000.210\to0.200; NCI1 0.475→0.4760.475\to0.476). The tokenizer scales with the data: held-out edge AUROC 0.940.94 (ENZYMES) and 0.9980.998 (NCI1). We report the full 11-method tables and mean ranks (ENZYMES 3.85, NCI1 4.05) in the revision and describe the method as competitive within the TUDataset scope.

W2: Combination of established techniques

The components are indeed standard; the contribution is the mechanism connecting them, and each link is tested by a dedicated control: local windows miss long-range edges (span analysis, Fig. B.2); removing the tokenizer's local edge mask collapses generation (orbit 0.590.59–1.171.17, connectivity ≈0\approx 0); replacing pair features with a global mean worsens PROTEINS orbit MMD 0.174→0.2480.174\to0.248; and the random-label control is consistent with the attribute–topology mechanism, with the gain vanishing where labels carry no structural signal. The matched code-learning control on MUTAG detected no significant difference between joint VQ training and post-hoc clustering; it does not establish equivalence or robustness across datasets. The revised introduction is restructured around this causal chain.

W3: Table 1 does not show consistent superiority

With the revised density-targeted MUTAG configuration, the main table reads: first on MUTAG (2.602.60) and PROTEINS (2.852.85), tied-first on SYN-COMM (2.402.40), second on SYN-RING-ROLE (3.003.00 vs. Raw 2.552.55). The dataset-dependent pattern is itself informative and we now make it explicit: where labels directly encode structure (role/community ids in the synthetic suites), the Raw control is strong by construction; on the two real datasets, where attributes are noisy, the context-quantized pipeline leads. The MUTAG first place is robust to how the comparison is set up: re-ranking with the Raw and Continuous controls also trained under the density-targeted objective leaves GraphVQ first (2.652.65 vs. Raw 3.103.10). Against the Table 1 baselines, GraphVQ's degree MMD on MUTAG is significantly lower than every one (paired tt, p≤.02p\le.02), and no significant orbit MMD difference was detected against the strongest of them (GraphRNN, p=.33p=.33); the density-targeted Raw variant is close on degree (0.1380.138 vs. 0.1350.135) and nominally behind on orbit (0.3920.392 vs. 0.3590.359). The earlier fourth place was observed under the probability-oriented edge-target objective.

AuthorsOpenReview ↗

Response to Reviewer YEuL (2/2)

W4: Variance

The main comparisons and matched controls use five seeds. The ablations use three seeds, while the NLL and position-head diagnostics use a single checkpoint and the resource profiling uses seed 0; these are labeled separately. Table 1 retains the five-seed means. The revision adds mean±std Tables D.6–D.7 in Appendix D (p. 21) covering all ten compared methods on all four main datasets, plus the density-targeted Raw and Continuous controls on MUTAG; std is the sample standard deviation over seeds 0–4. For example, MUTAG orbit MMD is 0.359±0.0530.359\pm0.053 for GraphVQ and 0.401±0.0390.401\pm0.039 for GraphRNN. Appendix D also reports seed-level bootstrap 95% CIs (2000 resamples) and paired tt-tests (GraphVQ orbit CI: MUTAG [0.325,0.402][0.325,0.402]; PROTEINS [0.151,0.198][0.151,0.198]).

W5: Disconnected components

With the revised density-targeted MUTAG configuration, connectivity is 0.6530.653 (vs. 0.1090.109 unweighted; reference 1.001.00), and the residual gap is concentrated in isolated nodes rather than a broken main component. On PROTEINS, pair conditioning improves connectivity substantially over the unconditional control with identical size sequences: 1.541.54 vs. 3.003.00 components, 66.3%66.3\% vs. 54.2%54.2\% fully connected, largest-component fraction 0.9730.973 vs. 0.8150.815 (test reference: 1.071.07 / 94.6%94.6\% / 0.9710.971). A fixed-seed gallery with held-out test graphs, one-stage, and two-stage samples accompanies the table in the revision.

W6: Time and resources

End-to-end wall-clock (one training run + 2,000 generated graphs, seed 0, RTX 3090, GPU idle):

MethodMUTAG s/1kPROTEINS s/1kPeak memory (MUTAG / PROTEINS)Parameters
GraphVQ22.397.954 MB / 6.86 GB39.2K
GraphRNN273.21428.752 MB / 287 MB15.5K
GraphARM9.043.236 MB / 705 MB7.0K
GRAN (block size 4)7.115.320 MB / 61 MB3.5K
GraphVAE (one-shot)5.48.230 MB / 40 MB446K

Against the closest sequential baseline (GraphRNN), GraphVQ is 1212–15×15\times faster end-to-end at this budget, with better orbit MMD on PROTEINS (0.1750.175 vs. 0.3010.301). Block-autoregressive (GRAN) and one-shot (GraphVAE) models are faster, as expected. On MUTAG, training and sampling separate as: train 9.99.9 s, sample 17.117.1 s/1k (GraphVQ); train 18.618.6 s, sample 10.710.7 s/1k (one-stage); the pooled Stage-B decoder adds 20.720.7K parameters. The PROTEINS peak memory comes from materializing the full upper-triangle candidate tensor and is controlled by the chunk width BB; the revision reports batch sizes, NN, and BB for each run.

I read authors response and have revised my score.

Reviewer_fjbPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents GraphVQ. It’s a two-staged attributed-graph generator inspired by VQ-VAE. The method improves orbit MMD over one-stage generation on some datasets, but its advantage over simpler representations is less conclusive.

Critical strengths

The separation of local tokenization from full-adjacency generation is sensible and computationally compact. On PROTEINS, orbit MMD improves from 0.474 for One-stage to 0.175.

Critical weaknesses

a. Compare to VQ-VAE paper, the method GraphVQ seems no much additional novel ideas added.

b. It seems lack direct measurement metric to evaluate the quality of generated graph. For example, how to actually measure the quality of generated protein graph? The current measure seems far from actual application.

Questions for authors

a. The foundation-model motivation is reasonable, but the experiments establish only the utility of the tokenizer and decoder within small graph-generation benchmarks. How do you justify that it can be applied across different graphs, and finally foundation model?

1 comment
AuthorsOpenReview ↗

Response to Reviewer fjbP

We thank the reviewer for pressing on the two questions that decide the paper's value: what is new relative to VQ-VAE, and how the quality of a generated protein graph is measured directly.

W-a: What is new relative to VQ-VAE?

The method's difference from existing VQ-style graph pipelines is in how tokens are built and how they are decoded, and each link is verified by a control rather than asserted. Concretely, GQT tokenizes PPR-based node contexts with residual VQ and drives molecular generation with a GPT-style prior, and GFT builds a transferable tree vocabulary for pretraining and prompting; neither decodes a full adjacency from token-derived pair conditions:

  • Information source. Tokens encode local BFS context windows (node attributes plus a local edge mask), not whole-graph latents; removing the edge mask collapses generation (orbit 0.590.59–1.171.17, connectivity ≈0\approx 0).
  • Condition granularity. The decoder receives chunk-pooled pair features with position-wise readouts, not a single global code; replacing them with a global mean worsens PROTEINS orbit MMD 0.174→0.2480.174\to0.248.
  • Output range. Stage B enumerates the full upper triangle, so edges outside the serialization window can be emitted; this is what moves orbit MMD 0.474→0.1750.474\to0.175 over one-stage generation on PROTEINS (2.02.0–17×17\times across all four main datasets).

The random-label control is consistent with this account: the conditioning gain vanishes where labels carry no structural signal. The matched code-learning control on MUTAG (q6Uq thread) detected no significant difference between joint VQ training and post-hoc clustering; this finding is limited to MUTAG and does not establish equivalence. The novelty is this verified mechanism, not the components; the revision restructures the introduction around the chain and contrasts the design with these graph applications of VQ in related work.

W-b: How do you measure the quality of a generated protein graph?

In short: pair conditioning brings label-conditioned node degrees 3.8×3.8\times closer to the held-out reference than the unconditional control, and this holds under a size-range-restricted reference. We agree distribution-level MMDs are indirect, so we added a label-conditioned topology analysis: for each graph, the degree of each node type and the label–label edge mass, against the held-out test split (five seeds; pair and unconditional variants share identical sampled size sequences).

PROTEINSType-0 degreeType-1 degreeType-2 degreeSame-label edge frac.
Held-out test reference3.723.773.950.665
GraphVQ, pair condition3.394.114.400.754
GraphVQ, unconditional4.504.926.330.748

Pair conditioning reduces the mean absolute error of label-conditioned degrees from 1.441.44 to 0.370.37 edges (macro-averaged over the three node types, computed on the pooled per-label means), a 3.8×3.8\times reduction toward the reference. (The same-label edge fraction is essentially unchanged by conditioning, 0.7540.754 vs. 0.7480.748: the gain is in per-type degree, not in edge-mass composition.) Because generated graphs average 22.422.4 nodes vs. 39.539.5 in the reference, we additionally ran a size-range-restricted reference sensitivity check: the 88 test graphs falling inside the generated samples' 5–95% size range (mean 25.125.1 nodes). Under that restricted reference the error is 0.4000.400 (pair) vs. 1.491.49 (unconditional), versus 0.3740.374 vs. 1.441.44 under the full reference — the pair-over-unconditional advantage (3.73.7–3.8×3.8\times) holds under both. The size deficit itself decomposes into the generation-time node cap of 64 (target 39.5→31.239.5\to31.2) and the token prior's length behavior (31.2→22.431.2\to22.4); both are now described in the paper, and size control is listed as a limitation with a concrete fix direction. These are benchmark structural metrics on TUDataset labeled graphs; biological validity (folding, function) is out of scope and not claimed.

Q-a: Connection to a graph foundation model

The current evidence establishes a tokenizer/decoder interface that operates unchanged across six graph distributions (four main + ENZYMES/NCI1), up to 126 nodes, with roughly unchanged orbit MMD in the 128-node arm — including held-out edge AUROC ≥0.94\ge 0.94 on the new datasets. Cross-domain codebook sharing, frozen-tokenizer transfer, and pretraining at scale are the next steps, and the revision states the concrete conditions under which they would be tested (frozen tokenizer, cross-graph code usage, reconstruction and generation metrics) rather than claiming them now.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common response

We thank all three reviewers for the detailed and constructive feedback. During the discussion period we ran a series of controlled experiments targeting the entanglements the reviews identified, with the same splits, metrics, and evaluation code as the main table. The main comparisons and matched controls use five seeds (0–4); three-seed ablations and single-checkpoint diagnostics are identified separately. The new evidence covers: (i) a matched code-learning control (joint VQ training vs. post-hoc k-means discretization, discrete interface held fixed); (ii) a four-decoder comparison with matched loss, candidate order, chunk width, and budget; (iii) two additional real datasets (ENZYMES, NCI1, up to 126 nodes) and a 128-node arm; (iv) seed-level mean±std and bootstrap CIs for the full grid; (v) connectivity/component statistics and galleries; (vi) profiling with training and sampling separated; (vii) a chunk-level pair-probability diagnostic; (viii) an NLL decomposition of the MUTAG under-generation; and (ix) a label-conditioned topology analysis of generated PROTEINS graphs, including a size-range-restricted reference sensitivity check.

Two notes on the MUTAG protocol. First, the §4.2 sentence and the §4.3 comparison “0.36 vs. 0.61” in the submission mixed two protocols (density-targeted vs. unweighted); both are now reported in separate, labeled rows. Second, the revision adopts a revised density-targeted configuration (w+=3w_+{=}3) as the main MUTAG configuration — an additional choice supported by the matched controls below, with the original unweighted variant retained in Appendix C. Under the matched protocol, GraphVQ vs. Raw on MUTAG is 0.3590.359 vs. 0.3920.392 orbit MMD.

A summary of the revised headline result: with the density-targeted Stage-B objective on MUTAG — a principled response to its 12.1% positive-edge rate, derived below — GraphVQ ranks first on MUTAG (mean rank 2.60, recomputed over the full participant set under the pre-registered rule), first on PROTEINS (2.85), and tied-first on SYN-COMM (2.40), while remaining second on SYN-RING-ROLE where labels directly encode structure.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.