{"id":"903abd77-1f33-4460-a205-33680e6c3d87","arxiv_id":"2505.06282","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"IFL-GCL uses InfoNCE's similarity signal to resample semantically similar non-augmented pairs as positives and updates the contrastive loss, improving graph pre-training accuracy on several benchmarks.","lead":"This paper proposes treating graph contrastive learning as a positive-unlabeled learning problem and uses similarity scores learned by InfoNCE to find unlabeled positive node pairs, then retrains with a corrected loss. The method shows accuracy gains up to 9.05% on OOD benchmarks, but the theoretical justification relies on an assumption about label ordering that is likely violated in this setting.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 19 fails because IOD is violated on exactly the non-augmented pairs the method must rank: distinct-node semantic positives have p(o=+1|x)=0, so the InfoNCE density ratio cannot encode p(y=+1|x).","rationale":"The reader's weakest-assumption analysis targets IOD (Eq. 10), and my reading agrees: this is the load-bearing step. My stress test sharpens the objection by noting that in augmentation-based GCL, p(o=+1|x) is degenerate on D_U (zero for every distinct-node pair), so the IOD ordering condition cannot connect InfoNCE similarity to semantic positivity for any pair that the method must re-rank. The empirical evidence in Section 2.2, which the paper itself provides, is direct evidence of the violation. I considered whether the InfoNCE density-ratio step (Eqs. 15-18) is a separate failure, and it is also questionable once the corrected loss replaces the original InfoNCE objective, but the primary invalidation is IOD. The paper deserves credit for releasing code and for reporting consistent accuracy gains, and those empirical gains are not dismissed; however, the paper's title and central theoretical framing claim a 'free lunch' that the proof does not support. Since the reader already recommended REJECT and my analysis strengthens rather than changes that conclusion, the verdict should remain unchanged.","tokens_in":17569,"tokens_out":6054,"duration_ms":66377,"concrete_test":"Train the warmed-up GRACE/GCA encoder exactly as in Algorithm 1 before applying the corrected loss on Cora and GOODTwitch, then freeze it. For all non-augmented pairs in D_U, compute s_theta(n,n') and an oracle semantic score (label agreement or cosine similarity from the supervised encoder used in Section 2.2). Compute the Spearman rank correlation between s_theta and the oracle semantic score restricted to distinct-node pairs. Also, for the threshold t_s, compare the SameClassRatio of selected D_U^+ (Eq. 29) against the base rate of random non-augmented pairs. If the rank correlation is not significantly positive or the selected set has SameClassRatio at the base rate, Eq. 19 is falsified for the very pairs the method relies on, and the threshold-based resampling is not semantically guided.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Eq. 19: s_theta(n,n') preserves the order of p(y=+1|x). Its proof uses the IOD assumption (Eq. 10), which asserts that orderings of p(y=+1|x) match orderings of p(o=+1|x), and then uses InfoNCE to model the density ratio r(x)=p(x|y=+1,o=+1)/p(x) (Eq. 18). In the augmentation-based GCL setting, o=+1 is assigned only to same-node augmented pairs; for every distinct-node pair in D_aug^-, p(o=+1|x)=0. The set IFL-GCL must re-rank is exactly D_U = D_aug^-, so for any two distinct-node pairs x and xhat, p(o=+1|x)=p(o=+1|xhat)=0. The IOD equivalence in Eq. 10 then becomes vacuous: it cannot distinguish semantic positives from semantic negatives, and it is directly contradicted by the paper's own Section 2.2 evidence showing that some non-augmented pairs have supervised similarity comparable to or higher than augmented positive pairs. Therefore InfoNCE, trained on augmentation labels, models the augmentation-posterior density ratio, not the semantic positive posterior; Eq. 19 does not follow. This is not a small technical gap: it removes the entire bridge between learned similarity and semantic positivity. Additionally, even if Eq. 19 held for the original InfoNCE loss, the corrected loss in Eq. 27 changes the training objective, so subsequent resampling using s_theta is no longer covered by the proof. The empirical gains in Tables 1 and 2 may still be real, but they do not establish the claimed 'free lunch.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IFL-GCL, a graph contrastive learning method that reframes GCL as a positive-unlabeled (PU) learning problem. The central idea is that InfoNCE's learned similarity s_theta(n,n') is a 'free lunch' estimate of the density ratio r(x)=p(x|y=+1,o=+1)/p(x), and that under an Invariance of Order (IOD) assumption this similarity preserves the ordering of the true positive-class posterior p(y=+1|x). The method uses this ordering to threshold s_theta and identify semantically similar non-augmented pairs D_U^+, then redefines the InfoNCE objective (Eq. 27) so that these pairs are treated as positive with confidence-weighted losses. Experiments on nine graph datasets, in both IID and OOD settings and with LLM-based features, report consistent improvements over GRACE and GCA, with gains up to 9.05%.","tokens_in":17883,"tokens_out":5867,"duration_ms":59730,"significance":"If the theoretical link were valid, the contribution would be significant: it would turn InfoNCE's density-ratio estimation into a principled semantic signal for debiasing GCL, and the proposed corrected loss would be a natural maximum-likelihood consequence. The paper has concrete strengths: the code is publicly available, the experimental protocol is standard, the improvements over directly related baselines are consistent across many datasets, and Section 4.4.1 provides an interesting analysis of the quality of the discovered D_U^+ pairs. However, the central theoretical claim is not sound. The IOD assumption is violated in exactly the GCL setting the paper studies, the density-ratio derivation conflates the data-generating process of augmentation-based contrastive pairs with independent node sampling, and the corrected loss changes the training objective so that the proof of Eq. (19) no longer covers the model actually used for resampling. The empirical results may still indicate a useful heuristic, but they do not establish the paper's advertised 'free lunch' claim.","major_comments":[{"comment":"The IOD assumption is violated by the GCL labeling process. In this paper, o=+1 is assigned deterministically to augmented same-node pairs, while every distinct-node non-augmented pair has p(o=+1|x)=0 (Eq. 1). For any two distinct-node pairs x and x' in D_U, the right-hand side of Eq. (10) holds in both directions because 0 <= 0, so Eq. (10) would force p(y=+1|x)=p(y=+1|x'). This is directly contradicted by the paper's own Section 2.2 and Fig. 2, which show that non-augmented pairs vary widely in semantic similarity, with some exceeding augmented pairs. Consequently, Eq. (11) and the central claim Eq. (19) do not follow, and the classifier in Eq. (20) has no theoretical basis for identifying D_U^+.","section":"Section 3.2.1, Eq. (10), (11), (19)"},{"comment":"The proof of Eq. (19) applies to a model trained with the original InfoNCE loss Eq. (3). After the warm-up stage, Algorithm 1 switches to the corrected loss Eq. (27), which is a different objective. The density-ratio interpretation of s_theta is not established for the model trained under this corrected loss, so using s_theta from subsequent iterations to select D_U^+ is not justified by the theorem. Moreover, the procedure is a self-training loop: samples selected by the model's own similarity scores are then upweighted to increase those scores. The paper does not discuss confirmation bias or provide any control for it, and the 'free lunch' claim depends precisely on the model being trained under the original InfoNCE objective.","section":"Section 3.2.2, Eq. (27) and Algorithm 1"},{"comment":"The identification p(x)=q(n)q(n') used in Eq. (17) does not reflect the actual sampling process in GCL. Contrastive samples are not obtained by independently sampling two nodes from their marginal distributions; they are constructed by generating two augmented views of the same graph and pairing nodes across views, so the marginal distribution of pairs is determined by the augmentation process and the graph structure. Therefore the density-ratio equality s_theta(n,n') ∝ q(n,n')/(q(n)q(n')) = r(x) in Eq. (18) is not established. The step from Eq. (16) to Eq. (18) requires a model of how x is sampled that the manuscript does not provide.","section":"Section 3.2.1, Eq. (17)-(18)"}],"minor_comments":[{"comment":"The phrase 'decently sorted' should be 'descendingly sorted' or 'sorted in descending order'.","section":"Section 2.2, caption of Fig. 2"},{"comment":"There are several typos: 'proof' should be 'prove' in 'We will now proof this', and 'Invariance of OrDer' should be 'Invariance of Order'.","section":"Section 3.2.1"},{"comment":"The negative set D_aug^- in Eq. (1) contains only cross-view pairs (u_i,v_j) and (v_i,u_j) for i≠j, but the denominator of Eq. (2) also includes intra-view pairs s_theta(u_i,u_j) for j≠i. Please clarify whether intra-view pairs are part of the negative set and how they relate to D_aug^-.","section":"Eq. (1) and Eq. (2)"},{"comment":"The text says 'Our methods occupy the best performances across all datasets apart from Computers' but then says 'our models rank second-best on all but PubMed and Photo'. These statements are inconsistent with the table: IFL-GC is best on PubMed, Photo, GOODTwitch, and GOODCora, while IFL-GR is best on Cora and GOODCBAS. Please rephrase to match the reported numbers.","section":"Section 4.2, discussion of Table 1"},{"comment":"The hyperparameter beta appears in Eq. (25) but is not listed in the inputs of Algorithm 1, and its role as a weighting exponent for all D_U^+ terms is not fully specified. Please define beta explicitly and include it in the algorithm's input list.","section":"Section 3.2.2, Eq. (25)-(27) and Algorithm 1"},{"comment":"The phrase 'show significantly improvements of IFL-GCL' should be 'show significant improvements of IFL-GCL' for grammatical correctness.","section":"Abstract"}],"recommendation":"reject","confidential_remarks":"The central theoretical claim of the paper, stated in the title and abstract as a 'free lunch', is unsound because the IOD assumption is violated in the augmentation-based GCL setup the paper itself formalizes. This is not a presentation issue but a load-bearing error: Eq. (19) is the bridge between InfoNCE similarity and semantic positivity, and without it the threshold-based resampling is an ungrounded heuristic. The empirical gains are consistent and the code is available, so a substantially revised manuscript might reposition the method as a purely empirical debiasing technique, but that would be a different contribution with a different evaluation and, likely, a different set of baselines. I therefore recommend rejection rather than major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The empirical package is real: IFL-GCL gives consistent gains over GRACE/GCA across nine datasets, including OOD and LLM-feature setups, and code is public. The theoretical claim that InfoNCE is a free lunch for semantic positivity does not hold as stated; the proof hinges on IOD, which is vacuous for exactly the pairs the method needs to rank.\n\nWhat is new: the PU framing is not novel—PU contrastive learning appears in refs [1,18,38]—but the dynamic resampling plus product-weighted corrected loss is a legitimate extension. Experiments are broad; IID gains are small but consistent, OOD gains on GOODTwitch/GOODCBAS are larger but noisy. The analysis of same-class ratio and supervised similarity of D_U^+ is a nice check.\n\nWhere it falls apart: Eq. 19 is load-bearing. IOD (Eq. 10) requires ordering of p(y=+1|x) to match ordering of p(o=+1|x). In augmentation-based GCL, p(o=+1|x)=0 for every distinct-node pair, so the assumption cannot rank semantic positives among D_U. The density ratio InfoNCE estimates is relative to augmentation-defined positives, not semantic positives. Once Eq. 27 replaces the InfoNCE objective, the earlier optimality no longer applies. So the free-lunch theorem is not established; the method is a self-training heuristic. That can work, and the evidence suggests it does, but the title overclaims.\n\nMinor: threshold t_s is dataset-dependent and selected per dataset; the paper does not say whether selection uses validation only. With a 1:1:8 split, leakage is a real worry. Many tunable knobs mean small IID gains could be tuning artifacts. Citations look fair.\n\nBottom line: I would send this to a serious referee rather than desk-reject. The empirical contribution and public code deserve scrutiny. But I would not accept the theory as is; either repair the bridge between InfoNCE similarity and semantic positivity or reframe the method as empirical debiasing. For GCL practitioners the recipe is easy to try; for theory readers it is a cautionary example.","headline":"A useful GCL debiasing recipe whose theoretical 'free lunch' bridge is broken; worth a serious referee, but not acceptance as-is.","tokens_in":18476,"tokens_out":7422,"would_cite":true,"duration_ms":75065,"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":"InfoNCE's learned similarity preserves the ordering of the positive-class posterior, so the contrastive score itself can act as a free semantic classifier inside graph contrastive learning.","keywords":["graph contrastive learning","InfoNCE","positive-unlabeled learning","semantic similarity","sampling bias","graph pre-training","out-of-distribution generalization","LLM-enhanced graphs"],"falsifier":"Train a GCL model on a graph with known class labels, then check whether the InfoNCE similarity $s_\\theta(n,n')$ consistently ranks same-class non-augmented pairs above different-class pairs after warm-up; if the ordering is violated, the free-lunch theorem fails. A more direct test: construct synthetic augmentations where only low-semantic pairs are labeled positive, and see whether thresholding $s_\\theta$ recovers the withheld high-semantic positive pairs.","tokens_in":17313,"feed_emoji":"📈","tokens_out":4661,"duration_ms":45929,"temperature":0.7,"pith_summary":"This paper claims that standard graph contrastive learning contains a hidden semantic signal: the InfoNCE similarity between two node representations preserves the ordering of the probability that the pair is semantically positive, even though training only labels augmented views as positive. On that basis, the paper argues GCL should be seen as a Positive-Unlabeled learning problem, where augmented pairs are labeled positives and non-augmented pairs are unlabeled rather than negatives. Using InfoNCE as a free density-ratio estimator, the method extracts semantically similar non-augmented pairs and redefines the maximum-likelihood objective into a corrected loss. If correct, this removes the need for extra supervision or a separate density-ratio estimator and directly mitigates the sampling bias that hurts GCL, with reported gains of up to 9.05% on one OOD benchmark and consistent improvements in LLM-enhanced graph settings.","feed_headline":"InfoNCE hides a semantic signal that fixes graph contrastive bias","feed_subtitle":"New loss treats similar non-augmented pairs as positives, boosting node classification by up to 9%.","key_machinery":"The load-bearing identity is the density-ratio interpretation of InfoNCE: the learned similarity $s_\\theta(n,n')$ is proportional to $r(\\mathbf{x}) = p(\\mathbf{x}|y=+1,o=+1)/p(\\mathbf{x})$, and by the Invariance of Order assumption this density ratio is ordered exactly like the true positive probability $p(y=+1|\\mathbf{x})$. This identity converts a contrastive score into a semantic classifier, allowing the method to mine $D_U^+$ by thresholding and then to reweight the InfoNCE objective so that discovered positives are promoted proportionally to their similarity.","core_discovery":"The central claim is the order-invariance identity of Eq. (19): for any contrastive samples $\\mathbf{x}=(n,n')$ and $\\hat{\\mathbf{x}}=(\\hat{n},\\hat{n}')$, $p(y=+1|\\mathbf{x}) \\le p(y=+1|\\hat{\\mathbf{x}})$ if and only if $s_\\theta(n,n') \\le s_\\theta(\\hat{n},\\hat{n}')$, where $s_\\theta$ is the exponential cosine similarity used in InfoNCE. The paper proves this by connecting InfoNCE's score to the density ratio $r(\\mathbf{x}) = p(\\mathbf{x}|y=+1,o=+1)/p(\\mathbf{x})$, which under the Invariance of Order assumption has the same ordering as the positive-class posterior. Therefore, thresholding $s_\\theta$ recovers unlabeled positive pairs $D_U^+$, which the paper then inserts into a corrected maximum-likelihood loss, Eq. (27), with confidence weights. The authors frame this as a free lunch: the contrastive score already carries semantic ordering, so no additional model or label is needed to correct sampling bias.","pith_inferences":["(Editorial inference) If the order-invariance theorem holds, any InfoNCE-trained representation already encodes an implicit semantic-pair classifier, which could be reused for tasks like unsupervised node retrieval, pairwise link scoring, or outlier detection without additional training; the paper does not explore these applications.","(Editorial inference) The free lunch is not fully free in practice because the method depends on a warm-up phase and a similarity threshold $t_s$; the reported hyperparameter sensitivity indicates that a too-low threshold can inject false positives and worsen bias, so the gains hinge on selecting these well.","(Editorial inference) The Invariance of Order assumption could be tested directly per dataset by comparing the ranking of augmentation-label probabilities against the ranking of true semantic positivity for held-out pairs; the paper's own OOD results suggest this ordering may be more reliable in some graphs than others, which would predict where the method helps most.","(Editorial inference) A natural extension, not studied in the paper, is to apply the same density-ratio interpretation to other self-supervised objectives that minimize alignment and uniformity; the proof specifically targets InfoNCE, so transferring it would require a new argument."],"forward_implications":["A GCL model can be warmed up with standard InfoNCE, then use its own similarity scores to resample semantically positive non-augmented pairs and train with the corrected loss, yielding accuracy gains up to 9.05% on GOODCBAS and consistent positive deltas over the GRACE and GCA baselines in both IID and OOD scenarios.","The same correction works when node features come from LLM encoders, and the reported improvement tends to grow with LLM scale, suggesting the method becomes more effective as semantic modeling improves.","The corrected objective imposes a stronger constraint on the discovered positives than a linear combination of likelihoods, because it multiplies the likelihoods with exponential confidence weights, which the paper argues leads to more thorough bias correction.","Because the method treats semantically similar cross-distribution pairs as positives, it reports larger improvements on OOD benchmarks than on IID ones, indicating a transferable-knowledge effect.","The theoretical result is not limited to a specific graph model: it applies to any InfoNCE-trained encoder with the same augmentation semantics, opening a generic upgrade path for existing contrastive pre-training pipelines."],"supporting_citations":[{"why":"Defines InfoNCE and establishes the density-ratio interpretation that the paper's free-lunch theorem builds on.","marker":"[25]"},{"why":"Supplies the Invariance of Order assumption and the corollary that density-ratio order equals positive-posterior order.","marker":"[17]"},{"why":"Surveys positive-unlabeled learning and the SCAR assumption that the paper contrasts with the more relaxed IOD used here.","marker":"[12]"},{"why":"Defines the GRACE baseline, one of the two InfoNCE-based models that IFL-GCL builds on and corrects.","marker":"[48]"},{"why":"Defines the GCA baseline, the other InfoNCE-based model that IFL-GCL augments with semantic guidance.","marker":"[49]"},{"why":"Introduces positive-unlabeled contrastive learning, providing the closest prior PU-contrastive framing that the paper extends.","marker":"[1]"},{"why":"Presents the heuristic negative-as-positive correction that the paper contrasts with its maximum-likelihood-derived corrected loss.","marker":"[42]"}],"fun_headline_variants":["InfoNCE's hidden semantics fix graph contrastive bias","Free lunch: InfoNCE corrects graph sampling bias","Graph contrastive learning gets a free semantic boost","InfoNCE unlocks semantic positives to beat graph bias"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the Invariance of Order assumption: the chance that a contrastive sample gets labeled positive by data augmentation is ordered exactly like its true semantic positivity, so a higher labeling probability always means a higher true positive probability.","fun_headline_variants_meta":{"raw":{"variants":["InfoNCE's hidden semantics fix graph contrastive bias","Free lunch: InfoNCE corrects graph sampling bias","Graph contrastive learning gets a free semantic boost","InfoNCE unlocks semantic positives to beat graph bias"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1335,"prompt_tokens":1080,"completion_tokens":255,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":696,"completion_tokens_details":{"reasoning_tokens":191}},"tokens_in":696,"tokens_out":255,"duration_ms":3070,"temperature":1.0,"reasoning_tokens":191,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:36:37.996212+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a GCL model on a graph with known class labels, then check whether the InfoNCE similarity $s_\\theta(n,n')$ consistently ranks same-class non-augmented pairs above different-class pairs after warm-up; if the ordering is violated, the free-lunch theorem fails. A more direct test: construct synthetic augmentations where only low-semantic pairs are labeled positive, and see whether thresholding $s_\\theta$ recovers the withheld high-semantic positive pairs.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Invariance of Order assumption and the corollary that density-ratio order equals positive-posterior order."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Surveys positive-unlabeled learning and the SCAR assumption that the paper contrasts with the more relaxed IOD used here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the GCA baseline, the other InfoNCE-based model that IFL-GCL augments with semantic guidance."},{"cited_title":"Positive Unlabeled Contrastive Learning","cited_arxiv_id":"2206.01206","evidence_quote":"Introduces positive-unlabeled contrastive learning, providing the closest prior PU-contrastive framing that the paper extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents the heuristic negative-as-positive correction that the paper contrasts with its maximum-likelihood-derived corrected loss."}],"review_version":1}