{"id":"50f8aac2-c98b-4da5-a268-13d3803a3499","arxiv_id":"2505.07171","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ReCDAP conditions a diffusion model on relation and positive/negative label information with attention pooling, reporting state-of-the-art few-shot knowledge graph completion scores on NELL and FB15K-237.","lead":"This paper introduces ReCDAP, a diffusion-based model for few-shot knowledge graph completion that explicitly models positive and negative triples and uses attention pooling to improve tail entity prediction. On NELL and FB15K-237 benchmarks, the authors report improved mean reciprocal rank and hits@k over prior few-shot knowledge graph completion methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on baseline numbers not re-run under a common protocol; direct comparison could change the ranking.","rationale":"The reader's weakest_assumption focuses on negative-tail replacement possibly creating false negatives. While technically valid, that concern has limited practical force on these datasets: with ~68k entities on NELL and ~14.5k on FB15K-237, a randomly replaced tail is a true negative with overwhelming probability, so contamination of the negative distribution is minuscule and unlikely to explain the reported gains. The more load-bearing issue is the comparative evaluation. The central claim is explicitly state-of-the-art performance, and that claim rests entirely on Table 2. The paper says baselines were not re-run, yet the table flags results as coming from official implementations; different original papers use different protocols for negative sampling, filtered metrics, and data splits. Without a controlled re-run, the reported MRR gaps could be due to protocol differences rather than method quality. The internal split inconsistency for FB15K-237 reinforces this concern. The proposed concrete test, rerunning all baselines under a single protocol, would directly settle whether the SOTA claim survives. Since the reader already assigned CONDITIONAL and flagged baseline concerns, my analysis does not change the verdict; it shifts the emphasis from negative sampling to evaluation comparability.","tokens_in":8999,"tokens_out":7016,"duration_ms":73058,"concrete_test":"Re-run GMatching, MetaR, FSRL, FAAN, GANA, CIAN, and NP-FKGC using their official implementations, the same pre-trained TransE/REFORM embeddings, the same filtered evaluation protocol, and the same train/valid/test relation splits as ReCDAP, over at least 5 seeds. Check whether ReCDAP's MRR remains above each baseline by more than the combined standard error; also verify the FB15K-237 test relation count (33 versus 13).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2 is the sole evidence for the headline claim that ReCDAP outperforms seven prior methods. The text in Sec. 4.2 states 'we used the results reported in those studies without re-running,' while the table note says * marks results from the authors' official implementation; these numbers were not produced under a single common evaluation protocol. Differences in negative-sample construction for ranking, filtered versus unfiltered metrics, validation relation splits, and pre-trained embeddings can alter MRR by several points, so the 0.045 MRR gap over NP-FKGC on NELL is not yet demonstrated. Section 4.1 also contains an internal inconsistency for FB15K-237: the text says split '75/11/13' but Table 1 lists 33 test relations. Without re-running all baselines under identical conditions, or at least reporting standard deviations and the evaluation code, the comparative SOTA claim is not falsifiable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ReCDAP, a few-shot knowledge graph completion model that combines a GNN-based global aggregator, a Bi-LSTM relation learner, a conditional diffusion module for modeling positive and negative support triple embeddings, a transformer-style attention pooler, and a TransE-style scoring function. The authors claim state-of-the-art results on NELL and FB15K-237 in the 5-shot setting, outperforming seven prior FKGC methods, with ablations isolating the contribution of the diffusion module and the attention pooler. The code is publicly released.","tokens_in":9204,"tokens_out":7466,"duration_ms":72161,"significance":"If the empirical results are confirmed, the paper's main contribution is the explicit modeling of positive and negative support distributions via a conditional diffusion process, a departure from prior work that uses negatives only as contrastive signals. The release of code, the ablation study, and the use of standard benchmarks are points in favor. However, the current evidence does not establish the headline SOTA claim: the diffusion model conditions on the clean input during the reverse process, undermining the generative interpretation, and the baseline comparison is not performed under a common evaluation protocol. The core idea is plausible and potentially useful, but the paper needs substantial revision to make the claims falsifiable and technically consistent.","major_comments":[{"comment":"The reverse process is conditioned on c_global = Concat(r', z0, r_l), which includes the clean input z0 itself. In a standard DDPM the reverse transition p_theta(z_{t-1}|z_t) must not observe z0; here the model is trained to predict noise with access to the target, so Eq. (6)'s statement that 'new data is generated by sampling from the learned distribution p_theta(z0)' is not justified. At inference, if the condition contains the observed support embedding z0, the model refines that embedding rather than generating a new sample from the learned distribution; if z0 is not available, the conditioning is undefined. Please clarify the inference-time use of z0, or remove it from the condition and validate the generative claims; alternatively, reframe the module as a denoising representation refiner and avoid distributional claims.","section":"2.3.1 (Eq. 5)"},{"comment":"The comparative SOTA claim is not yet demonstrated. Section 4.2 states that baseline results were taken from prior studies without re-running, and only some entries (marked with *) come from official implementations. The baselines therefore differ in negative-sample construction, filtered/unfiltered ranking, relation splits, and embedding initialization, which can change MRR by several points. Moreover, no standard deviations or significance tests are reported, so the reported gap of 0.045 MRR over NP-FKGC on NELL may be within run-to-run variance. Please provide a unified reimplementation or at least report multi-seed results and significance tests.","section":"4.2, Table 2"},{"comment":"The negative support set is generated by randomly replacing the tail entity, and the paper does not filter replacements that are actually valid tails for the relation. Since the central contribution is the separation of positive and negative distributions, any false negatives contaminate the estimated negative distribution and can make the separation artifacts of the corruption process. Please quantify the fraction of corrupted triples, use a filtered negative-sampling strategy, or otherwise demonstrate that the results are robust to this contamination.","section":"2.3, negative triple construction"}],"minor_comments":[{"comment":"The text states FB15K-237 is split in the ratio 75/11/13, but Table 1 reports 33 test relations. Please correct the typo (75/11/33) and ensure the relation counts match the split.","section":"4.1 and Table 1"},{"comment":"The sentence 'The attention pooler was configured with a single head.' is duplicated; please remove one occurrence.","section":"4.2"},{"comment":"The notation for the L2 norm in the score definition is missing or malformed; please typeset Score = -||h' + r' - t'||_2 properly.","section":"Eq. (9)"},{"comment":"The last row 'w/o ReCD & AttnPool0.281' is missing a space between the variant name and the MRR value.","section":"Table 3"},{"comment":"The ablation study is reported only on NELL; please add FB15K-237 results or justify their omission.","section":"4.3"},{"comment":"The figure is very dense and the flow from z0 into the condition c_global is difficult to follow; a clearer diagram or pseudocode would help reproducibility.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a short SIGIR paper; the main risk is the mismatch between the generative framing and the actual use of z0 in the condition. I would ask the authors to clarify or reframe this before acceptance. The evaluation protocol also needs strengthening; without it, the SOTA claim is not falsifiable. I see no indication of misconduct, but the missing details matter."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a solid incremental paper for the few-shot KGC crowd. The specific combination — conditional diffusion with explicit positive/negative label conditioning plus attention pooling — is new as far as I know, and the ablations show each piece earns its keep. Code is public. That is real work and worth engaging with.\n\nThe soft spots are in two places. First, the diffusion reverse process conditions on z0 itself (c_global = Concat(r', z0, r_l), Eq. 5, Sec. 2.3.1). That means the model essentially autoencodes the input; it is not generating new triple embeddings in any meaningful sense. The 'distribution modeling' story is oversold, though the final scoring does go through a margin loss on held-out triples, so the method can still work as representation learning. The paper should acknowledge this directly.\n\nSecond, the SOTA claim. They took baseline numbers from prior papers or official implementations without re-running under a common protocol. Table 1 says 33 test relations for FB15K-237 while Section 4.1 says a 75/11/13 split, which would be 13 test relations if those are counts. That inconsistency alone needs fixing. The 0.045 MRR gap over NP-FKGC on NELL could easily shrink or vanish under a unified evaluation setup. I would not call the SOTA claim falsifiable until the authors release the exact evaluation code and a consistent split table.\n\nMinor concerns: negative sampling by random tail replacement can create false negatives, and there are no error bars or significance tests. Neither is fatal, but both deserve mention in a revised version.\n\nWho is this for? People actively working on few-shot KGC, especially those using diffusion or negative sampling tricks. It deserves a serious referee — I would not desk-reject it — but it needs revision to address the conditioning issue and the evaluation protocol before the numbers can be trusted.\n\nRecommendation: engage with it as a borderline accept with revisions; do not take the SOTA table at face value yet.","headline":"Useful incremental FKGC paper; the method is plausible, but the SOTA claim and the 'generative' framing need scrutiny.","tokens_in":9682,"tokens_out":1985,"would_cite":false,"duration_ms":22111,"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":"ReCDAP treats negative triples as a real distribution, and by conditioning diffusion on relation and positive/negative labels, it achieves state-of-the-art few-shot knowledge graph completion on NELL and FB15K-237.","keywords":["few-shot learning","knowledge graph completion","conditional diffusion","negative triples","attention pooling","link prediction","relation-based conditioning"],"falsifier":"Measure the contamination rate of the random tail-replacement procedure on NELL and FB15K-237 (the fraction of constructed negatives that are actually true triples in the full KG) and check whether high-contamination relations show a performance drop; alternatively, replace the random negatives with type-constrained negatives that are guaranteed non-existent and see if the reported MRR gains change, which would indicate whether the claimed distributional separation depends on the cleanliness of the negative sample.","tokens_in":8812,"feed_emoji":"🔗","tokens_out":6490,"duration_ms":57001,"temperature":0.7,"pith_summary":"This paper claims that few-shot knowledge graph completion improves when the model treats non-existent triples as a second data distribution rather than as mere contrastive examples. It introduces ReCDAP, which feeds relation information, the triple's own embedding, and a positive/negative label into a diffusion model so that the denoiser learns separate latent distributions for true and false triples. On the 5-shot setting the method reports state-of-the-art MRR of 0.505 on NELL and 0.579 on FB15K-237, ahead of seven prior FKGC baselines. If correct, the result implies that explicit distributional modeling of negative evidence is a productive direction for data-scarce link prediction.","feed_headline":"Diffusion model over true and false triples sets KG record","feed_subtitle":"Separating false from true triples in a diffusion model lifts 5-shot tail prediction past prior methods.","key_machinery":"The central object is the relation-based conditional diffusion (ReCD) module: a DDPM whose reverse process is conditioned on a concatenated signal $c_{\\text{global}} = [r', z_0, r_l]$ made of the relation representation, the triple embedding, and a positive/negative label, with the noise-predicting network implemented as a U-Net using FiLM conditioning. The companion mechanism is an attention pooler, a transformer with a learnable query per group that pools denoised positive and negative embeddings separately into $z_{\\text{pos}}$ and $z_{\\text{neg}}$, which are concatenated to adjust the TransE scoring of candidate tails. Together they let the model approximate distinct distributions $q(z_0 \\mid \\text{positive})$ and $q(z_0 \\mid \\text{negative})$ instead of using negatives only as a ranking margin.","core_discovery":"ReCDAP's central claim is that a relation-conditioned denoising diffusion model, together with an attention pooler, can estimate and separate the latent distributions of positive and negative triples, and that feeding the pooled separation back into a TransE-style score improves tail prediction under 5-shot supervision. The model builds an extended support set by randomly replacing the tail of each support triple to create negatives, then conditions the reverse diffusion step on the relation representation, the initial triple embedding, and a positive/negative label. The attention pooler reads the denoised embeddings with separate queries for positives and negatives and concatenates the results to update the head and tail embeddings for scoring. The paper reports that this pipeline achieves MRR 0.505 and Hits@1 0.493 on NELL, and MRR 0.579 and Hits@1 0.491 on FB15K-237, the best numbers among the seven compared methods, and its ablations attribute the gain to both the diffusion separation and the attention pooling.","pith_inferences":["The random tail-replacement scheme for negatives is a weak negative sampler: on relations with many valid tails, some 'negative' triples will be true facts. A type-constrained or hard-negative variant might change the measured gains, and on high-degree relations the separation could degrade.","The same conditional-diffusion-plus-pooling recipe could transfer to other few-shot structured prediction problems—such as sparse recommendation or drug interaction prediction—where negative examples are abundant and cheap to fabricate.","The authors note memory inefficiency on high-degree datasets like Wikidata; an inductive or low-resource test would clarify whether the distributional separation generalizes beyond the two observed benchmarks.","One testable extension: use the learned positive/negative distributions to generate candidate tails directly (sampling from $p_\\theta(z_0)$) rather than only re-scoring existing entities, which could turn the model into a generative completion engine."],"forward_implications":["The reported SOTA numbers imply that distributional modeling of negatives outperforms contrastive-only negative use for FKGC, suggesting future work should invest in richer negative modeling rather than stronger margin losses alone.","Because the relation learner and attention pooler are relation-agnostic modules, the same conditioning design can be applied to other relation-aware prediction heads (e.g., rotational or bilinear scoring) without changing the diffusion core.","The ablations show that removing the positive/negative label hurts more than removing the support negative set entirely, so the label-injected separation is a key driver of the gain; direct corollary: injecting coarse group identity into any generative encoder may help few-shot prediction.","The method's reliance on an iterative T=100 diffusion sampler makes inference slower than one-pass models; the paper's own conclusion notes this, so any deployment would need to trade speed for the accuracy gain."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model that the conditional diffusion module is built on.","marker":"[5]"},{"why":"Shows a diffusion model can learn the distribution of knowledge graph fact embeddings, the direct precedent ReCDAP extends.","marker":"[9]"},{"why":"Provides the global aggregator and Bi-LSTM relation learner, plus the dataset splits and the strongest baseline the method must beat.","marker":"[11]"},{"why":"Introduced the support negative set and serves as a baseline that uses negatives only as a contrastive signal.","marker":"[12]"},{"why":"Defines the TransE translation scoring used by the score calculator.","marker":"[1]"},{"why":"Provides FiLM conditioning, the mechanism that injects the relation condition into the U-Net noise predictor.","marker":"[14]"},{"why":"Basis of the attention pooler, which uses transformer-style attention with learnable queries.","marker":"[18]"},{"why":"Supplies pre-trained TransE embeddings and the one-shot/few-shot evaluation protocol and datasets.","marker":"[22]"},{"why":"Supplies error-aware pre-trained embeddings used for the FB15K-237 experiments.","marker":"[20]"}],"fun_headline_variants":["Diffusion splits true/false triples for few-shot KG","ReCDAP: Attention pooling + diffusion for KG completion","Conditional diffusion separates positive from negative triples","ReCDAP: Diffusion model disentangles triple kinds in KG","Attention pooling over diffusion improves few-shot KG completion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that randomly replacing a support triple's tail entity produces a genuinely non-existent triple, so the set of 'negative' triples truly represents a separate distribution from positive facts; if some replacements are actually valid tails, the two distributions blur and the separation mechanism becomes unreliable.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion splits true/false triples for few-shot KG","ReCDAP: Attention pooling + diffusion for KG completion","Conditional diffusion separates positive from negative triples","ReCDAP: Diffusion model disentangles triple kinds in KG","Attention pooling over diffusion improves few-shot KG completion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001162,"raw_usage":{"total_tokens":4823,"prompt_tokens":971,"completion_tokens":3852,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":3773}},"tokens_in":587,"tokens_out":3852,"duration_ms":28432,"temperature":1.0,"reasoning_tokens":3773,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:22:34.638009+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the contamination rate of the random tail-replacement procedure on NELL and FB15K-237 (the fraction of constructed negatives that are actually true triples in the full KG) and check whether high-contamination relations show a performance drop; alternatively, replace the random negatives with type-constrained negatives that are guaranteed non-existent and see if the reported MRR gains change, which would indicate whether the claimed distributional separation depends on the cleanliness of the negative sample.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies error-aware pre-trained embeddings used for the FB15K-237 experiments."}],"review_version":1}