{"id":"bc0ddec3-c42c-48fc-9d79-b1f267eb68cc","arxiv_id":"2607.25169","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CondPSE, a learned polynomial-filtered structural encoder, beats GPSE on CSL and EXP expressivity benchmarks yet matches it on molecular transfer.","lead":"This paper introduces CondPSE, a graph structural encoder that applies polynomial graph filters and FiLM-style modulation to random node probes, then freezes it as input for downstream GNNs. It separates synthetic graphs that GPSE cannot, but on molecular property prediction it performs about the same as GPSE, showing that synthetic expressivity gains do not guarantee downstream gains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inherited GPSE baselines not rerun under CondPSE's exact seed/CV protocol; head-to-head 'comparable transfer' conclusion is vulnerable to small shifts.","rationale":"The reader's conditional verdict is appropriate. I examined other possible concerns—confounded comparison, fixed random probe, polynomial filter implementation—but none is as directly load-bearing as baseline comparability because the paper's stated conclusion is explicitly a head-to-head relative to GPSE. Even if all CondPSE numbers are correct, the conclusion 'no better than GPSE' requires the GPSE numbers to be trustworthy under the same protocol. The fact that the paper discloses the inheritance is a point in its favor and makes the fix straightforward. I therefore recommend no change to the reader's CONDITIONAL verdict.","tokens_in":9442,"tokens_out":5183,"duration_ms":51618,"concrete_test":"Rerun GPSE (and LapPE/RWSE/no-PE) under the exact CondPSE harness: same folds/seeds (ten-times stratified 5-fold for CSL/EXP; 10-seed GPS and 4-seed ZINC sweep), same adapter and hyperparameters from Table 5, same graph preprocessing, and identical GPU/software environment. Use paired seeds so each encoder sees the same seed for a given fold. If any GPSE number moves by more than one reported standard deviation, recompute all deltas; specifically check whether GPSE CSL remains below 60 and whether GPSE remains better than CondPSE on ZINC and PCQM4Mv2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Tables 1–3 inherit all non-CondPSE baseline numbers (GPSE, LapPE, RWSE, no-PE) from the GPSE paper or its public implementation, while CondPSE is evaluated under 'corresponding public GPSE evaluation settings' but with its own seed draws (Section A.3). The paper's central negative claim—that CondPSE's CSL/EXP gains over GPSE do not translate into downstream advantage—is thus a head-to-head comparison built on externally sourced reference points. The molecular differences are small relative to reported variance: ZINC MAE 0.0648±0.0030 vs 0.0649±0.0035, PCQM4Mv2 0.1196±0.0004 vs 0.1207±0.0006. A baseline shift of ~1% (well within seed sensitivity) could flip 'without surpassing' into 'surpassing' on one or more datasets, and the CSL gap (42.9 vs 97.3) could shrink if GPSE is rerun with more favorable seeds. The paper acknowledges the inherited baselines, but this does not reduce their load-bearing role: every headline comparison is relative to those numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CondPSE, a learned positional/structural encoding (PSE) encoder that applies a bank of K learnable polynomial graph filters to Gaussian node probes and refines the resulting branches with FiLM-style conditioning using cross-filter, local message-passing, and global graph-level signals. CondPSE is pretrained on OGBG-MolPCBA using the same node-level and graph-level pretraining targets as GPSE, then frozen and used as input encoding for downstream GNNs. On synthetic isomorphism benchmarks, CondPSE raises CSL accuracy from 42.9% (GPSE-reported) to 97.3% and EXP from 68.3% to 99.9%. On molecular benchmarks (ZINC, MolHIV, MolPCBA, PCQM4Mv2-subset) with a GraphGPS backbone, CondPSE performs comparably to GPSE without consistently surpassing it. A ZINC backbone sweep shows mixed relative performance. The authors conclude that stronger benchmark-level structural discrimination does not automatically translate into better downstream graph-level transfer.","tokens_in":9781,"tokens_out":6230,"duration_ms":58831,"significance":"If the central result holds, the paper makes a valuable diagnostic contribution: a frozen learned PSE encoder can separate structures that 1-WL-bounded message passing cannot (CSL/EXP), yet this expressivity does not by itself yield a downstream molecular advantage. This challenges the implicit assumption that stronger synthetic structural discrimination should transfer to real-world tasks. The paper is transparent about its limitations, provides full encoder and training configuration tables, and reports multi-seed results with error bars. The ablation study cleanly isolates the contribution of the polynomial filter bank, and the authors explicitly disavow a definitive superiority claim on the backbone sweep. These strengths make the paper a useful reference for the PSE community.","major_comments":[{"comment":"All non-CondPSE baselines (No PE, GPSE, LapPE, RWSE) are inherited from the GPSE paper or its public implementation rather than rerun under the paper's exact seed/CV protocol. The headline comparisons — 'raises CSL accuracy from 42.9% to 97.3%' and 'performs comparably to GPSE without surpassing it' — both depend on these externally sourced numbers. Because CondPSE is evaluated with its own seed draws, a small shift in the GPSE baseline (e.g., CSL 42.9±7.9 under another seed protocol) could alter the claimed gains. The paper should rerun the baselines under identical folds, seeds, and downstream configurations, or at minimum provide cross-seed variance and paired comparisons where feasible.","section":"A.3, Tables 1–3"},{"comment":"CondPSE is pretrained on OGBG-MolPCBA alone, whereas GPSE is pretrained on a much larger and more diverse corpus (the GPSE paper uses a multi-dataset pretraining set). The paper does not control for this difference in pretraining data scale/diversity. The molecular transfer results could therefore reflect not only the encoder architecture but also the amount and composition of pretraining data. To support the claim that 'stronger CSL/EXP performance does not translate to downstream advantage,' CondPSE should be pretrained on the same pretraining corpus as GPSE, or the comparison should explicitly account for this confound.","section":"A.2, Tables 2–3"},{"comment":"The molecular differences between GPSE and CondPSE are not subjected to any significance test, and several are well within one standard deviation: ZINC MAE 0.0648±0.0030 vs 0.0649±0.0035; PCQM4Mv2-subset MAE 0.1196±0.0004 vs 0.1207±0.0006. The claim that CondPSE is 'without surpassing' GPSE is essentially an acceptance of the null hypothesis. With 10 seeds (and 4 seeds in the sweep), paired tests such as Wilcoxon signed-rank or a bootstrap over seeds are straightforward and would substantiate or qualify the conclusion. Without such tests, the central negative finding is not statistically grounded.","section":"Tables 2–3"}],"minor_comments":[{"comment":"The term 'NormalSE' is used without explicit definition at first mention. Clarify that it denotes standard i.i.d. Gaussian node probes, and please keep the terminology consistent ('NormalSE' vs. 'Gaussian probes').","section":"3.1"},{"comment":"The 'w/o filter bank' ablation is not described. Specify exactly what replaces the polynomial filter bank in this variant (e.g., a single linear transformation of A^0 Z, randomly initialized untrained filters, or a different propagation scheme) and report the parameter count if it differs.","section":"Table 1"},{"comment":"The statement 'several gaps are within reported variation' would benefit from concrete numbers. Indicate which differences are beyond one standard deviation and, ideally, add error bars or a significance test for the backbone sweep as well.","section":"Table 3"},{"comment":"The phrasing 'raises CSL accuracy from 42.9% to 97.3% relative to GPSE' could be read as a direct head-to-head experiment. Because the 42.9% value is inherited, consider rephrasing to 'compared to the GPSE-reported value' or 'exceeds the GPSE-reported accuracy.'","section":"Abstract"},{"comment":"The expressivity-generalization trade-off discussion is explicitly untested here, which is appropriate. The references [14,15] are cited as supporting theory; ensure they are publicly available and correctly described, especially [15], which is dated 2026.","section":"5"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-written and the core research question is interesting, but the quantitative claims rely on external baselines that are not rerun under the paper's protocol, and the pretraining data mismatch between CondPSE (MolPCBA-only) and GPSE (multi-dataset) is a serious confound. These issues are fixable but require additional experiments or a much more cautious framing. I would not publish the current version without either (a) rerunning GPSE and other baselines under identical conditions, or (b) substantially narrowing the claims to avoid head-to-head statements that the evidence does not support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe core of this paper is a clean, negative finding: a learned PSE encoder that crushes CSL/EXP (97.3% and 99.9%, vs GPSE's 42.9% and 68.3%) does not transfer any better than GPSE on molecular benchmarks. That is worth knowing. The architecture, CondPSE, is a sensible variant of GPSE — polynomial filter bank over Gaussian probes plus FiLM-style conditioning — and the ablations attribute most of the synthetic gain to the filter bank. Ten-run five-fold CV makes the CSL/EXP numbers credible, and the authors are explicit that the molecular results are mixed and that the ZINC sweep is a diagnostic, not a superiority claim.\n\nThe soft spot is exactly what the stress-test note flags: every non-CondPSE baseline is inherited from the GPSE paper or its public implementation, not rerun under CondPSE's seed draws. The molecular differences are small enough that a one-percent baseline shift could flip 'without surpassing' into 'surpassing' on one or two datasets. That said, the qualitative message — huge synthetic gains, no clear downstream win — is likely robust, because even if GPSE's CSL were higher, the transfer gap to molecular tasks remains visible. But as written, the head-to-head numbers are not fully self-contained.\n\nThere are also smaller gaps: no significance tests on the molecular differences, no code/data release, and the 'target alignment' discussion is explicitly speculative. None of these are fatal, but combined with the inherited baselines, they keep the paper from being a fully closed empirical study. The most useful fix is to rerun GPSE under the same protocol or at least verify the inherited numbers reproduce; the second is to release the probe-seeded encodings.\n\nWho is this for? People working on learned positional/structural encodings who care about whether synthetic expressivity benchmarks predict transfer. It is a solid workshop-level contribution and a legitimate diagnostic. With the baselines verified and code released, it would be publishable at a main conference workshop or a short paper venue. I would send it to a serious referee, but I'd tell them the load is on the comparison. I would not cite it in the next twelve months until the baselines are confirmed.","headline":"A useful, honest negative result about learned PSE transfer, but inherited baselines soften the head-to-head conclusion.","tokens_in":10218,"tokens_out":2798,"would_cite":false,"duration_ms":25121,"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":"This paper shows that a learned structural encoder's strong discrimination of synthetic hard graphs does not automatically transfer to better downstream molecular prediction.","keywords":["Graph Neural Networks","Positional Encoding","Graph Expressivity","Transfer Learning","Polynomial Graph Filtering","Structural Encoding","Conditional Modulation","GPSE"],"falsifier":"Reproduce the full evaluation by running GPSE and CondPSE under identical random seeds, data splits, and CV protocol on CSL, EXP, ZINC, MolHIV, MolPCBA, and PCQM4Mv2-subset, and check whether CondPSE's synthetic gains persist and whether any molecular ordering changes; if CondPSE then clearly surpasses GPSE on a molecular task, the paper's central conclusion would need to be revised.","tokens_in":9325,"feed_emoji":"🧪","tokens_out":6295,"duration_ms":53137,"temperature":0.7,"pith_summary":"This paper introduces CondPSE, a learned positional/structural encoder that feeds random Gaussian node probes through a bank of learnable polynomial graph filters, creating multiple structural-response branches that are then refined by FiLM-style conditioning using cross-filter, local, and global signals. Pretrained on OGBG-MolPCBA to reconstruct structural and positional targets, CondPSE is frozen and supplied to downstream GNNs. On synthetic isomorphism benchmarks, it far outperforms the previous learned encoder GPSE (CSL accuracy 42.9→97.3%, EXP 68.3→99.9%), but on molecular property prediction it ends up comparable to GPSE, with a ZINC backbone sweep showing no consistent ordering. The paper's central claim is that stronger benchmark-level structural discrimination does not automatically translate into stronger downstream graph-level transfer, and it argues that downstream integration and pretraining-target alignment are likely responsible.","feed_headline":"Synthetic gains do not transfer for polynomial-filtered graph encoder","feed_subtitle":"CondPSE's sharp CSL/EXP gains over GPSE vanish on molecular benchmarks, with results depending on the downstream backbone.","key_machinery":"The load-bearing component is the polynomial graph filter bank: K learnable polynomial filters applied to i.i.d. Gaussian node probes, where each filter is a normalized mixture of adjacency-propagation powers A^i Z. This creates multiple structural-response branches, and each subsequent conditioning block applies FiLM modulation using cross-filter pairwise comparisons, a shared ResGatedGCN local pass, and graph-level statistics. The filter bank converts random probes into topology-sensitive encodings without any eigendecomposition.","core_discovery":"The central discovery is an empirical separation between benchmark-level structural expressivity and downstream transfer for frozen learned PSE encoders: CondPSE, built from a polynomial filter bank and conditional modulation, can almost perfectly discriminate CSL and EXP graphs that 1-WL-bounded message passing cannot, yet those gains do not produce a downstream advantage over GPSE on molecular property prediction. The paper establishes this by freezing the encoder after pretraining and evaluating it under the same downstream protocols as GPSE, then dissecting the result with ablations and a backbone sweep.","pith_inferences":["The paper leaves unexplored whether fine-tuning (rather than freezing) CondPSE would let its strong structural signal translate into molecular gains; a natural extension is to run the downstream backbones with CondPSE weights trainable.","Because the downstream encodings are generated from a single fixed probe realization, the results could be probe-seed dependent; a testable extension is to average over multiple probe realizations when generating encodings and check variance.","The argument that expressivity can hurt generalization points to a measurable prediction: CondPSE encodings should produce larger intra-class variance than GPSE encodings on property-similar molecular graphs; a reader could compute this variance directly from the stored encodings.","The filter-bank design could be reused as a differentiable component inside a downstream GNN rather than as a frozen input encoder; the paper does not test that configuration."],"forward_implications":["Synthetic structural-discrimination benchmarks like CSL and EXP are not, on their own, reliable predictors of a frozen PSE encoder's downstream value; future encoder evaluations should report transfer results on multiple real-world tasks.","The same frozen encoding yields different relative performance across downstream backbones (e.g., CondPSE beats GPSE on GCN/GatedGCN/Transformer but loses on GIN/GINE on ZINC), so conclusions about encoder quality should be drawn from a backbone sweep, not a single architecture.","Because the polynomial filter bank accounts for most of the CSL/EXP gain, future PSE encoders may benefit from explicit multi-depth propagation responses rather than stacked message passing only.","Molecular transfer being close between CondPSE and GPSE suggests the bottleneck for frozen learned PSE encoders may lie in pretraining-target alignment and downstream integration rather than in structural expressivity.","The paper's results caution against equating higher expressivity with better generalization: overly fine structural distinctions could reduce the downstream model's ability to generalize within property-similar graph classes."],"fun_headline_variants":["Filter-bank encoder excels on synthetic graphs, fails to beat GPSE on molecules","Poly-filtered encoder: big synthetic wins, no real-world edge over GPSE","Structural expressivity doesn't guarantee molecular transfer for CondPSE","CondPSE's synthetic gains don't translate to molecular property prediction","Polynomial encoder beats 1-WL limits but not GPSE on real benchmarks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The headline comparisons rely on reusing GPSE, LapPE, and RWSE results from the GPSE paper or its public implementation under the assumption that they were generated under CondPSE's exact seed, split, and cross-validation protocol; if that assumption fails, the relative improvements and the 'no better than GPSE' conclusion could shift.","fun_headline_variants_meta":{"raw":{"variants":["Filter-bank encoder excels on synthetic graphs, fails to beat GPSE on molecules","Poly-filtered encoder: big synthetic wins, no real-world edge over GPSE","Structural expressivity doesn't guarantee molecular transfer for CondPSE","CondPSE's synthetic gains don't translate to molecular property prediction","Polynomial encoder beats 1-WL limits but not GPSE on real benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1206,"prompt_tokens":824,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":283}},"tokens_in":568,"tokens_out":382,"duration_ms":4674,"temperature":1.0,"reasoning_tokens":283,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T03:12:56.073811+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the full evaluation by running GPSE and CondPSE under identical random seeds, data splits, and CV protocol on CSL, EXP, ZINC, MolHIV, MolPCBA, and PCQM4Mv2-subset, and check whether CondPSE's synthetic gains persist and whether any molecular ordering changes; if CondPSE then clearly surpasses GPSE on a molecular task, the paper's central conclusion would need to be revised.","supporting_citations":[],"review_version":1}