{"id":"8bd5fad9-57e3-4713-a54d-46d4218df10f","arxiv_id":"2501.15696","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"HyDRO distills graphs in hyperbolic space with random-walk spectral gap matching, improving link prediction and continual graph learning.","lead":"The paper introduces HyDRO, a graph distillation method that builds condensed graphs in hyperbolic space and uses random-walk spectral gap matching to preserve dynamic graph properties. It reports improved link prediction and continual learning performance over existing graph distillation methods on citation and social network benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The causal claim that spectral-gap alignment drives HyDRO's random-walk and CGL gains is unestablished: Lrw matches only λ2, §5.4 is qualitative heatmaps, and no ablation isolates Lrw.","rationale":"The paper's most distinctive contribution is the claim that spectral-gap alignment preserves random-walk properties and that this preservation drives the continual-graph-learning and task-generalization gains. The strongest claim in the abstract ties these gains to random-walk preservation. If Lrw is not actually responsible for the observed behavior, the central mechanism is unsupported even if the numbers in the tables happen to be reproducible. This is a load-bearing concern because it targets the causal link, not a table typo or a presentation issue. The reader identified the same weakest assumption: matching one scalar is assumed to be sufficient, with no ablation and only qualitative heatmap evidence. I agree with that assessment. The concrete ablation, combined with a quantitative commute-time error and a check of Asub's connectivity, would settle whether the spectral-gap term is doing the claimed work. Other issues, such as the abstract's overstatement of node-classification results, the narrative inconsistency in Table 2, and the absence of released code, are real but secondary; they affect clarity and reproducibility rather than the validity of the proposed mechanism. Therefore the current CONDITIONAL verdict remains appropriate: the method may work, but the central claim is not yet established by the presented evidence.","tokens_in":26073,"tokens_out":5108,"duration_ms":50703,"concrete_test":"On Cora and Citeseer, compare three configurations: (1) full HyDRO; (2) HyDRO with the Lrw coefficient set to 0 (same hyperbolic generator, same Lgm); (3) a Euclidean GCond-style distillation with the same Lrw term added. Report node-classification accuracy, link-prediction accuracy, CGL final accuracy, and a quantitative commute-time error between the condensed graph and a same-size subgraph of the original, e.g. ||L+_syn - L+_orig_sub||_F / ||L+_orig_sub||_F (with the 20,000 cap used in §5.4). Also report the number of connected components of Asub under the paper's sampling scheme. If configurations (2) and (3) match (1) on all metrics, the spectral-gap mechanism and hyperbolic space are not load-bearing; if (1) clearly beats (2) and (3), the stated mechanism is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 defines the random-walk loss as Lrw = |g_syn - g_sub|, where g = 1 - λ2 of a lazy walk matrix. This matches exactly one scalar, the second-largest eigenvalue, between a random sample of the original graph and the condensed graph. Section 5.4 validates random-walk preservation with commute-time heatmaps, but commute time is governed by the full pseudoinverse of the Laplacian, not by λ2 alone; two graphs with the same spectral gap can have arbitrarily different pairwise commute times. The paper offers no bound or theorem connecting |g_syn - g_sub| to any commute-time or diffusion error, and no ablation removes Lrw from the total loss in Eq. 6. Consequently, the reported continual-learning and link-prediction gains cannot be attributed to random-walk preservation; they may come from the hyperbolic parameterization or gradient matching. An additional unexamined detail is Asub: it is only described as sampled per epoch. If the sampled subgraph is disconnected, its lazy-walk spectral gap is identically zero, making the alignment target degenerate. These omissions leave the central mechanism asserted rather than demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HyDRO, a graph distillation method that generates condensed graphs by embedding node features in a Poincaré ball and aligning the spectral gap of the synthetic graph with a per-epoch sampled subgraph of the original graph, in addition to a gradient-matching loss. Experiments are reported for node classification, link prediction, neural architecture search, cross-architecture transferability, random-walk preservation via commute-time heatmaps, continual graph learning, membership-inference privacy, and denoising. The central claims are that HyDRO consistently outperforms state-of-the-art graph distillation methods on node classification and link prediction, and that spectral-gap alignment preserves random-walk dynamics, which in turn improves continual graph learning.","tokens_in":26442,"tokens_out":4051,"duration_ms":34002,"significance":"If the causal mechanism were established, using a spectral-gap alignment term in hyperbolic graph distillation would be a worthwhile contribution to graph condensation, since dynamic/random-walk properties are rarely targeted in this area. The breadth of evaluation is a strength: results are averaged over ten runs with standard deviations, and the benchmark suite spans many tasks. However, the node-classification claim overreaches relative to Table 1, and the paper provides no ablation or theory isolating the Lrw term, so the central mechanism remains assumed rather than demonstrated. The contribution is therefore conditional on a revision that recalibrates the claims and either proves or convincingly ablates the spectral-gap mechanism.","major_comments":[{"comment":"The abstract claims HyDRO \"consistently outperforming state-of-the-art methods in both node classification and link prediction tasks,\" but Table 1 shows HyDRO ranks second or third on Cora, Citeseer, Ogbn-arxiv, Flickr, and Reddit across most reduction rates (e.g., Cora 1.30%: GEOM 82.5±0.4 vs HyDRO 81.5±0.4; Reddit 0.05%: GDEM 93.1±0.2 vs HyDRO 91.8±1.4; Flickr 0.10%: GDEM 49.9±0.8 vs HyDRO 47.3±0.5). The conclusion repeats this overstatement (\"overall best performance on node classification tasks\"). The claims should be corrected to state that HyDRO is competitive in node classification and leads in link prediction, or the abstract should be qualified accordingly.","section":"Abstract and §5.1 (Table 1)"},{"comment":"The random-walk loss Lrw = |gsyn − gsub| matches only one scalar, the spectral gap (second eigenvalue) of a lazy walk matrix, between the synthetic graph and a per-epoch subgraph sample. Section 5.4 validates random-walk preservation with qualitative commute-time heatmaps, but commute time is governed by the full pseudoinverse of the Laplacian, not by λ2 alone; two graphs with identical spectral gap can have arbitrarily different pairwise commute times. No bound or theorem in the paper connects |gsyn − gsub| to any commute-time or diffusion error, and Lrw appears in Eq. (6) without an ablation that removes it. Consequently, the reported gains in continual graph learning and link prediction cannot be attributed to random-walk preservation; they could arise from the hyperbolic parameterization or the gradient-matching loss. Please add an ablation study with Lrw removed and include a quantitative commute-time error metric (e.g., mean or median relative error across node pairs) in addition to heatmaps.","section":"§3.2 and Eq. (6)"},{"comment":"The paper only states that Asub is \"sampled from the real graph A at each epoch\" without specifying the sampling procedure or any connectivity guarantee. If the sampled subgraph is disconnected, its lazy random walk spectral gap is exactly zero (since λ2 = 1), making the alignment target degenerate and potentially driving Lrw to push the synthetic graph toward a disconnected or trivial structure. Please specify how Asub is sampled (e.g., uniform node sampling with induced edges, random edge sampling, or a GraphSAINT-style sampler) and report whether the sampled subgraphs are connected in practice, or otherwise justify the choice.","section":"§3.2, Asub sampling"},{"comment":"The continual graph learning results are presented only as line plots without error bars or a table of means and standard deviations, making it impossible to judge whether HyDRO's apparent advantage over DM and GCDMX on Citeseer and Arxiv is statistically meaningful. Please provide numerical results with standard deviations for the CGL experiments, consistent with the reporting format used elsewhere in the paper.","section":"§5.5 and Figure 2"}],"minor_comments":[{"comment":"DBLP is listed twice in the dataset enumeration (\"DBLP ... and DBLP\"); one occurrence appears to be an error, since Wiki-CS and Coauthor-Physics are introduced separately later in the same paragraph.","section":"§4, Datasets"},{"comment":"The list of structure-based methods contains a duplicated method name \"GCDM\" in sequence; the intended second method is unclear and should be corrected.","section":"§5.5"},{"comment":"The text refers to \"HyRo\" where \"HyDRO\" is clearly intended; please fix this typo.","section":"Appendix G"},{"comment":"Several cells in Table 2 use inconsistent decimal formatting (e.g., \"64.89±1.26\" and \"64.61±4.2\"), and the dense layout makes mapping columns to methods difficult; please reformat the table for clarity.","section":"Table 2"},{"comment":"There are minor language issues: \"denosing\" should be \"denoising,\" and \"seminar papers\" should be \"seminal papers\" in the experimental-settings description.","section":"§5.1 and §5.7"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid experimental breadth and the link prediction results are genuinely promising. The main obstacles are the overclaim in the abstract regarding node classification and the lack of any ablation or theoretical support for the spectral-gap mechanism, which is the paper's distinctive contribution. I would not reject on the basis of disagreement with current consensus; the claims and mechanism can be fixed within the scope of a revision. I also suggest asking the authors to reconcile their GDEM reproduction comments with the GC-Bench and GCondenser benchmark reports, since they rely on third-party observations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: HyDRO is a legitimate new combination—hyperbolic embeddings plus gradient matching plus a spectral-gap alignment term—and it performs well on link prediction and several secondary tasks. The paper is worth engaging, but the current draft overstates node classification performance, contains a false narrative claim about Table 2, and never isolates the Lrw term that is supposed to be the main novelty. The causal story is not established.\n\nWhat is actually new: as far as the citations show, this is the first graph distillation method to work in hyperbolic space, and the first to try to preserve random-walk behavior via spectral gap alignment. That is a real idea. The link prediction results in Table 2 are mostly strong, and the continual learning results (Figure 2) suggest the method does something useful even if the mechanism is unclear. The MIA and denoising evaluations are a plus; most distillation papers ignore those dimensions.\n\nSoft spots, in proportion:\n1. Abstract says 'consistently outperforming state-of-the-art methods in both node classification and link prediction.' Table 1 shows HyDRO is second behind GEOM on Cora, Citeseer, and Ogbn-arxiv, and second behind GDEM on Flickr and Reddit. It wins Pubmed, but 'consistently outperforming' is not what the table says. Fix the abstract.\n2. The text around Table 2 claims that except for HyDRO, all methods degrade from NC to LP. On Cora at 1.3%, HyDRO drops from 80.3 NC to 74.5 LP—a downgrade. The claim is simply false.\n3. The central mechanism: Lrw = |g_syn - g_sub| matches one scalar, the second eigenvalue of the lazy walk matrix. The validation uses commute-time heatmaps, but commute time depends on the full pseudoinverse of the Laplacian, so matching λ2 gives no guarantee about commute times. No ablation removes Lrw from Eq. 6, so the reported gains cannot be attributed to random-walk preservation; they may come from the hyperbolic parameterization or gradient matching. Also, Asub is sampled per epoch without checking connectivity; a disconnected sampled subgraph has spectral gap zero, making the target degenerate. This needs at least a theoretical bound or an ablation, ideally both.\n4. No code or data released, which slows verification.\n\nThe paper is not circular and the external comparisons are standard. The core method is plausible. But the claims need to be aligned with the evidence before this is publishable.\n\nWho this is for: graph distillation and continual graph learning researchers. A serious referee would find enough substance here to work with. My recommendation: send it to peer review, but make clear the authors need to correct the overclaims, add an ablation of Lrw, and release code. If the ablation shows Lrw is unnecessary, the paper still contributes a hyperbolic distillation baseline, but the main selling point changes.","headline":"HyDRO is a plausible new hyperbolic graph distillation method whose headline claims run ahead of its own tables and whose random-walk mechanism is asserted rather than shown.","tokens_in":26829,"tokens_out":3429,"would_cite":false,"duration_ms":28731,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"HyDRO claims that distilling graphs in hyperbolic space while aligning the spectral gap of random walks yields condensed graphs that beat Euclidean baselines and retain dynamic behavior for continual learning.","keywords":["graph distillation","graph condensation","hyperbolic embeddings","Poincaré ball","spectral gap","random walk","continual graph learning","link prediction"],"falsifier":"Run HyDRO on Cora, Citeseer, PubMed, and Flickr with the random-walk loss term removed, keeping the same hyperbolic generator, gradient matching, and hyperparameters. If node-classification, link-prediction, and continual-learning results stay within one standard deviation of the full model, the spectral-gap alignment is not responsible for the reported advantages.","tokens_in":25822,"feed_emoji":"🕸️","tokens_out":9326,"duration_ms":77013,"temperature":0.7,"pith_summary":"Graph distillation shrinks a large network into a small synthetic graph that can train graph neural networks cheaply. This paper argues that prior distillation methods lose information because they operate in Euclidean space, which poorly represents the tree-like, heavy-tailed geometry of real networks, and because they preserve only static spectral structure. HyDRO instead generates the condensed graph inside the Poincaré ball, a model of hyperbolic geometry, and adds a loss term that matches the spectral gap of the lazy random walk between the synthetic graph and a sampled real subgraph. The authors report that the resulting condensed graphs achieve the highest link-prediction accuracy on almost all test sets, remain competitive on node classification and architecture search, and preserve random-walk properties well enough to improve continual graph learning. If these results hold, graph condensation can be made to carry dynamic diffusion information, not just static labels and structure.","feed_headline":"Spectral-gap matching makes distilled graphs preserve random walks","feed_subtitle":"Graphs condensed in hyperbolic space outdo Euclidean baselines and keep diffusion dynamics for continual learning.","key_machinery":"The load-bearing object is the spectral-gap alignment loss $L_{rw} = |g_{syn} - g_{sub}|$, where $g_{syn} = 1 - \\lambda_{2,syn}$ and $g_{sub} = 1 - \\lambda_{2,sub}$ are derived from the second-largest eigenvalues of the lazy random walk matrices of the condensed graph and a normalized, per-epoch sampled real subgraph ($\\epsilon = 1/2$). The spectral gap controls mixing time ($t_{mix} \\sim 1/\\Delta$) and diffusion speed, so the paper uses this single scalar as the carrier of random-walk information. The second component is the Poincaré ball generator: hyperbolic neural networks with Möbius linear transformations and Riemannian gradient updates produce the synthetic adjacency matrix, motivated by the claim that hyperbolic space fits power-law, tree-like network geometry without the distortion Euclidean embeddings incur.","core_discovery":"HyDRO's central claim is that a condensed graph should reproduce not only the original graph's task-relevant features but also its random-walk behavior, and that this can be done by generating the graph in hyperbolic space and aligning one spectral quantity. The method samples nodes, embeds their features in the Poincaré ball, builds edge embeddings by concatenation, maps them through Möbius linear layers, and reshapes the output into a symmetric adjacency matrix. Training minimizes $L_{total} = L_{gm} + L_{rw,norm} + \\beta L_{reg}$, with $L_{rw} = |g_{syn} - g_{sub}|$ comparing the spectral gaps of the synthetic and per-epoch sampled real graphs. The paper reports that this procedure outperforms all benchmark methods on link prediction, ranks at or near the top on node classification and neural architecture search, and yields commute-time heatmaps closer to the original graph than those of spectral baselines such as SGDD and GDEM. It further reports stronger continual graph learning under the Condense-and-Train framework and a favorable privacy-utility balance under membership inference attacks.","pith_inferences":["A natural next test is ablating $L_{rw}$ while keeping hyperbolic generation and gradient matching; if performance is unchanged, the reported gains come from the geometry rather than the random-walk alignment.","Matching the full eigenvalue spectrum or the commute-time matrix directly might preserve diffusion behavior more completely than a single spectral-gap scalar, but the paper does not explore this.","If spectral-gap alignment is indeed the causal mechanism, then Euclidean distillation methods could adopt the same loss and potentially narrow the gap with HyDRO, separating geometry effects from dynamics effects.","The privacy and robustness benefits are plausibly tied to hyperbolic decision surfaces, but the results shown do not isolate this cause; a targeted study varying curvature while fixing the distillation setup would test it."],"forward_implications":["Condensed graphs produced this way can serve as multi-task substitutes, transferring from node classification to link prediction without the accuracy drop seen in other methods.","Preserving the spectral gap preserves diffusion efficiency, which means distilled graphs can simulate information spread and support continual graph learning on evolving networks.","Hyperbolic generation changes the privacy and robustness profile of distillation, lowering membership-inference leakage relative to utility and degrading less under feature, structural, and adversarial structural noise.","The approach remains competitive on standard node-classification benchmarks, so the dynamic-preservation gains do not come at the cost of core task performance.","Because the spectral-gap term is dataset-agnostic, the pipeline extends to large inductive graphs such as Flickr and Reddit without changing the distillation objective."],"supporting_citations":[{"why":"Supplies the gradient-matching objective and the standard graph-condensation setup that HyDRO extends.","marker":"[Jin et al., 2021]"},{"why":"Provides the hyperbolic neural-network layers, including Möbius transformations and logarithmic/exponential maps, used to generate the condensed graph.","marker":"[Ganea et al., 2018]"},{"why":"Defines the spectral gap and its relation to random-walk convergence, the quantity HyDRO aligns.","marker":"[Hoffman et al., 2021]"},{"why":"Supports the claim that larger spectral gaps imply more efficient random-walk exploration.","marker":"[Coja-Oghlan and Lanka, 2009]"},{"why":"Supplies the Condense-and-Train continual graph learning framework on which HyDRO reports state-of-the-art results.","marker":"[Liu et al., 2023b]"},{"why":"SGDD is the Laplacian-energy baseline whose static spectral preservation HyDRO compares against for dynamic properties.","marker":"[Yang et al., 2024]"},{"why":"GDEM is the eigenbasis-matching baseline that HyDRO outperforms on most node-classification and link-prediction tables.","marker":"[Liu et al., 2023a]"},{"why":"Defines the commute-time computation used to evaluate random-walk preservation.","marker":"[Qiu and Hancock, 2007]"},{"why":"Provides the commute-time theory in large graphs that motivates using commute time as a size-insensitive random-walk metric.","marker":"[V on Luxburg et al., 2014]"},{"why":"Supplies the distortion argument for why Euclidean embeddings of tree-like graphs are inaccurate, motivating hyperbolic space.","marker":"[Linial et al., 1995]"}],"fun_headline_variants":["Hyperbolic distillation preserves random-walk dynamics","Spectral-gap matching in hyperbolic space for graph distillation","Graph condensation guided by random walks in hyperbolic space","HyDRO: hyperbolic embeddings for walk-preserving graph distillation","Distilled graphs keep walk features with hyperbolic spectral gap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that matching one number—the spectral gap between the condensed graph and a randomly sampled piece of the real graph—is enough to preserve random-walk and commute-time behavior, and that this preservation is what drives the continual-learning gains.","fun_headline_variants_meta":{"raw":{"variants":["Hyperbolic distillation preserves random-walk dynamics","Spectral-gap matching in hyperbolic space for graph distillation","Graph condensation guided by random walks in hyperbolic space","HyDRO: hyperbolic embeddings for walk-preserving graph distillation","Distilled graphs keep walk features with hyperbolic spectral gap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1596,"prompt_tokens":946,"completion_tokens":650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":574}},"tokens_in":562,"tokens_out":650,"duration_ms":6413,"temperature":1.0,"reasoning_tokens":574,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:02:03.033512+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HyDRO on Cora, Citeseer, PubMed, and Flickr with the random-walk loss term removed, keeping the same hyperbolic generator, gradient matching, and hyperparameters. If node-classification, link-prediction, and continual-learning results stay within one standard deviation of the full model, the spectral-gap alignment is not responsible for the reported advantages.","supporting_citations":[{"cited_title":"Hyperbolic neural networks","cited_arxiv_id":null,"evidence_quote":"Provides the hyperbolic neural-network layers, including Möbius transformations and logarithmic/exponential maps, used to generate the condensed graph."},{"cited_title":"Spectral gaps of random graphs and applications","cited_arxiv_id":null,"evidence_quote":"Defines the spectral gap and its relation to random-walk convergence, the quantity HyDRO aligns."},{"cited_title":"The spectral gap of random graphs with given expected degrees","cited_arxiv_id":null,"evidence_quote":"Supports the claim that larger spectral gaps imply more efficient random-walk exploration."},{"cited_title":"Does graph distillation see like vision dataset counter- part? Advances in Neural Information Processing Sys- tems, 36,","cited_arxiv_id":null,"evidence_quote":"SGDD is the Laplacian-energy baseline whose static spectral preservation HyDRO compares against for dynamic properties."},{"cited_title":"Graph simplification and matching using commute times","cited_arxiv_id":null,"evidence_quote":"Defines the commute-time computation used to evaluate random-walk preservation."},{"cited_title":"Hitting and commute times in large random neighborhood graphs","cited_arxiv_id":null,"evidence_quote":"Provides the commute-time theory in large graphs that motivates using commute time as a size-insensitive random-walk metric."},{"cited_title":"The geometry of graphs and some of its algo- rithmic applications","cited_arxiv_id":null,"evidence_quote":"Supplies the distortion argument for why Euclidean embeddings of tree-like graphs are inaccurate, motivating hyperbolic space."}],"review_version":1}