{"id":"f1019749-269c-44d5-b8a3-e27df36097c4","arxiv_id":"2607.14733","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GAttNHP pairs an attention-based event-chain encoder, a soft-group cross-chain excitation mask, and a non-crossing quantile head to predict entities and occurrence times in temporal knowledge graphs, with large gains on event-heavy datasets but not on all six.","lead":"GAttNHP forecasts future facts and their timing on temporal knowledge graphs by combining attention-based event histories, group-level excitation sharing, and ordered quantile prediction. It reports large gains on several benchmarks, but the abstract's six-dataset state-of-the-art claim is contradicted by the paper's own YAGO result.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The group-excitation mechanism is only defined within a 16-chain mini-batch (§3.4, Eqs. 15–21); no test-time batching protocol is specified, so the reported gains may depend on arbitrary batch composition rather than on learned cross-chain semantics.","rationale":"The reader's weakest_assumption locates the same vulnerability: cross-chain excitation is computed only among the B=16 chains in a mini-batch, so the learned Phi may reflect batch-sampling statistics rather than true mutual excitation. I sharpen this into an inference-time correctness issue: because the cross-attention memory pool contains only the other chains in the current batch, the model's output is not a well-defined function of a single query unless a deterministic test-time batching procedure is specified. No such procedure is given in §4 or §A.2.1. This is more load-bearing than the other valid concerns. The YAGO appendix contradiction affects the 'six-dataset SOTA' wording but not the mechanism; the median-vs-mean time comparison is a real evaluation bias but is fixable by reporting RMSE or median baselines; the NLL rectangle approximation is concerning but does not directly falsify the entity-ranking gains. The batch-dependence issue attacks the paper's named central contribution: the group attention module. The proposed test would settle it: if the model is stable across random batch compositions and B=1 matches B=16, then the concern does not land; if not, the reported gains cannot be attributed to semantic group-level Hawkes excitation. Since the issue is addressable by specifying and validating a test-time protocol, the conditional verdict is appropriate; no change to the reader's verdict is needed.","tokens_in":18931,"tokens_out":9444,"duration_ms":98295,"concrete_test":"Retrain GAttNHP on ICEWS14 with a fixed seed and evaluate the same test queries under 50 different random B=16 batch compositions, keeping all other components identical. Compute the standard deviation of MRR and the fraction of queries whose top-1/top-3 prediction changes across compositions; also evaluate the same queries with B=1. If MRR varies beyond the reported ±0.0064, if a non-negligible fraction of rankings change, or if B=16 does not beat B=1, the group-excitation result is batch-dependent and the central mechanism is not well-defined.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is the group-excitation module, and §A.4–A.5 attribute most entity-prediction gains to it. But that module is only computed inside a mini-batch of B=16 chains. Eq. (15) builds Phi_bar = W Phi W^T with W in R^{B×G}; Eq. (18) constructs the shared memory H_shared in R^{B×L×D}; Eq. (16) defines the Hawkes mask M_{q,k} using Phi_bar_{u,v}, which exists only for chain pairs in the same batch. During training, random batches mean Phi is estimated from which chains happen to co-occur, not from measured causal coupling; if mutually-exciting chains rarely share a batch, Phi_bar encodes sampling statistics. At inference, no batching protocol is given: random test batches give the same query different cross-chain contexts and different ranks, while B=1 would make the group term vanish. The ablation Base→Group thus conflates 'added access to other chains in the batch' with 'semantic group excitation.' The claimed mechanism—and the long-tail 'information borrowing' story in §A.5—is therefore not supported by the reported experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"GAttNHP models each (subject, relation) pair in a temporal knowledge graph as a marked event chain and treats TKG extrapolation as a marked temporal point process. The method has three components: (i) a self-attention history encoder (AttNHP) that computes a continuous-time chain embedding; (ii) a group-excitation module that softly assigns chains to G latent groups, learns a group-to-group excitation matrix Φ and decay vector γ, and uses these to form a Hawkes-derived mask in a cross-attention layer over a mini-batch of chains; and (iii) a Non-Crossing Quantile (NCQ) head, adapted from Wu et al. (2023), that predicts monotone quantiles of the inter-arrival distribution via a value/delta layer construction. The model is trained by joint NLL and pinball losses. Experiments report entity-prediction MRR/Hits@k on six datasets and time-prediction MAE/quantile scores on three ICEWS datasets, with ablations and frequency-based analyses.","tokens_in":19136,"tokens_out":10144,"duration_ms":90118,"significance":"The paper is a serious attempt to bring continuous-time mutual-excitation modeling to TKG extrapolation, and it has real strengths: entity baselines are re-run with official code on the same chronological splits; the ablations separate the self-attention, group, and NCQ components; and the efficiency analysis (Table 5) shows that the group mechanism is cheap in parameters and memory. If the claims were fully supported, GAttNHP would be a useful addition to the TKG-forecasting literature and would provide a practical way to share statistical strength across event chains. However, the current manuscript overstates the empirical results, and the central group mechanism is underspecified at inference time; both issues must be resolved before the contribution can be assessed.","major_comments":[{"comment":"The abstract and Section 4.1's opening claim ('Across every metric and every dataset, GAttNHP substantially outperforms prior methods') are contradicted by Appendix Table 6. On YAGO, CENET achieves MRR 0.6060 and TLogic 0.5831, both above GAttNHP's 0.5130. Furthermore, time prediction is evaluated only on ICEWS14/18/05-15 (Tables 2-3), so the abstract's 'on six benchmark TKG datasets, GAttNHP improves ... on both entity prediction and time prediction' is not supported for time prediction on GDELT/WIKI/YAGO. Please restrict the claims to the settings actually evaluated and discuss the YAGO gap explicitly in the abstract or conclusion.","section":"Abstract/§4.1, Table 6"},{"comment":"Section 3.4 and Eqs. (15)-(18), (21) define the group-excitation term only over a mini-batch of B=16 chains: Eq. (15) forms Φ-bar = W Φ W^T with W in R^{B×G}, Eq. (18) constructs H_shared in R^{B×L×D}, and Eq. (21) applies the Hawkes mask inside a cross-attention softmax over that batch. The test-time batching protocol is not specified anywhere in Section 4 or Appendix A.2.1. Consequently, if test queries are batched arbitrarily, the same query gets different cross-chain contexts (and different ranks) across runs, and with B=1 the group term degenerates to within-chain attention. The ablation Base→Group (Figure 2, Section A.4.1) and the frequency analysis (Table 7, Section A.5) therefore do not isolate 'semantic group excitation' from the generic ability to attend to other chains present in the batch. Please specify the inference-time batching (e.g., fixed global context chains, one batc","section":"§3.4, Eqs. (15)-(18), (21); §A.2.1; §A.5"},{"comment":"Table 4 in Section A.1.1 reports ICEWS05-15 train/valid/test sizes of 38,692/46,092/46,275. This is inconsistent with the stated chronological 80/10/10 split (validation would be larger than training) and differs from the standard ICEWS05-15 split used in prior work. Please correct the dataset statistics and state whether all reported ICEWS05-15 results use the corrected split; as printed, the main-table ICEWS05-15 numbers cannot be reproduced.","section":"§A.1.1, Table 4"},{"comment":"Section 3.6 presents the Non-Crossing Quantile head as one of the paper's three contributions. The construction—value layer, delta layer, softplus increments, and centered cumulative sum to enforce monotonicity—appears to be the same NCQ module as DNet (Wu et al., 2023), which the text only credits with 'adapted from Wu et al. (2023).' Please state explicitly which elements are new to this paper (e.g., the application to TKG inter-arrival times, the joint loss, or an architectural change). If the head is identical, the contribution claim should be reframed as an adaptation/evaluation and DNet should be cited in the introduction and abstract where the contribution is announced.","section":"§3.6, Eqs. (24)-(26)"},{"comment":"Section A.4.1's comparison of NCQ vs. MSE in multi-task training reports a catastrophic entity-prediction collapse for the MSE variant (e.g., ICEWS14 MRR drops from 0.5118 to 0.1124), which is used to argue that NCQ is 'essential.' However, Section A.2.2 reports a hyperparameter grid for β but does not say that the MSE variant's β was re-tuned; a small β that is appropriate for the bounded pinball loss may be very different for the squared-error loss. Please tune β separately for the MSE variant, or report the β values used, so that the comparison is controlled.","section":"§A.4.1, §A.2.2"}],"minor_comments":[{"comment":"Section 3.7 says the final model is selected by minimum validation loss, while Section A.2.1 says 'We keep the checkpoint with the lowest training loss.' Please reconcile these statements.","section":"§3.7 vs §A.2.1"},{"comment":"The abstract's 'six benchmark TKG datasets ... on both entity prediction and time prediction' should be revised to reflect that entity prediction is evaluated on six datasets and time prediction on three.","section":"Abstract"},{"comment":"The line 'The appendix is preserved from the original manuscript:' at the start of Appendix A appears to be an editorial note rather than content and should be removed.","section":"Appendix preamble"},{"comment":"The NLL integral in Eq. (23) is approximated with a single-step rectangle rule; please provide the discretization used or a sensitivity check showing that a finer approximation does not change the results.","section":"Eq. (23)"},{"comment":"For the external time-prediction baselines GHT and GHNN, please state explicitly which point prediction (e.g., conditional mean, median, or sampled time) was used to compute MAE, since their native output distributions may differ from the median quantile used for NCQ.","section":"Tables 2-3"}],"recommendation":"major_revision","confidential_remarks":"The abstract's overstated SOTA claim and the inconsistent ICEWS05-15 statistics in Table 4 should be checked carefully by the editor. The NCQ overlap with DNet (Wu et al., 2023) also warrants attention: the construction appears to be reused under a new name, and the paper should clarify the incremental contribution. The central unresolved technical point is the inference-time batching for the group-excitation module; without a specified protocol, the main mechanism is not well defined at test time."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"GAttNHP is a clean combination of three existing building blocks—Yang et al.'s attention encoder, GNHP's group Hawkes structure, and Wu et al.'s non-crossing quantile head—and it delivers large MRR gains on the ICEWS datasets with baselines re-run on the same splits. That part is solid. But the abstract's claim of six-dataset SOTA is false: the appendix shows YAGO loses to CENET and TLogic, and the flagship group-excitation mechanism is only computed inside a mini-batch of 16 chains, with no test-time batching protocol. So the 'information borrowing' story is not yet supported by the experiments.\n\nCredit where it's due: the entity-prediction baselines are re-run on the authors' own chronological splits, which makes the ICEWS numbers credible. The ablation attributes most of the gain to the group module, and the frequency-aware analysis showing larger gains on tail chains is a nice diagnostic. The time-prediction head is compared fairly against RQS and RMTPP on quantile scores and coverage, not just MAE, which is more rigorous than most TKG time-prediction work. The O(G^2) group structure is a sensible way to avoid pairwise excitation, and the efficiency numbers are plausible.\n\nThe soft spots are real. First, the abstract overclaims; YAGO contradicts the SOTA statement. Second, the group module is confined to within a random batch of 16 chains. The learned Phi is estimated from whichever chains happen to co-occur, and at inference the same query gets different contexts depending on batch composition. The stress-test note is on target: the mechanism paper claims to address mutual excitation is not separable from arbitrary batch sampling. Third, the NLL integral is approximated by a single rectangle at the target time; that's crude and unvalidated, but probably not fatal. Fourth, there are small inconsistencies: ICEWS05-15's train count in Table 4 looks like a typo (38,692 vs the usual ~369k), and the ablation MRR in Figure 2 doesn't match Table 1 exactly. No code or seeds are released.\n\nThis paper deserves a serious referee—the empirical gains are large and the approach is worth engaging. The review should push for a defined test-time batching protocol, a corrected abstract, code/seeds, and a time-prediction comparison that doesn't stack the deck toward the median. Fix those, and this becomes a solid contribution in continuous-time TKG reasoning.","headline":"A useful assembly of known components with strong ICEWS gains, but the abstract overclaims on YAGO and the group mechanism is only defined inside a 16-chain mini-batch, so the core explanation is not yet supported.","tokens_in":19796,"tokens_out":3018,"would_cite":false,"duration_ms":27792,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G55","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Treating each subject–relation pair as an event chain, GAttNHP learns shared Hawkes excitation through latent groups and forecasts both the next object and its arrival time.","keywords":["temporal knowledge graph","extrapolation reasoning","Hawkes process","mutual excitation","self-attention","quantile regression","long-tail events","link prediction"],"falsifier":"Randomly permute chain identities within each mini-batch while preserving each chain's own event sequence, so group assignments become meaningless relative to the actual chains; if the group module's gains on tail chains persist unchanged, the learned excitation is not capturing semantically meaningful cross-chain structure.","tokens_in":18612,"feed_emoji":"🕒","tokens_out":2848,"duration_ms":27982,"temperature":0.7,"pith_summary":"The paper argues that a temporal knowledge graph is better modeled as a collection of mutually exciting event chains than as static snapshots, and that cross-chain excitation can be captured cheaply by soft group assignment translated into an attention mask. If right, both what-happens-next and when-it-happens predictions improve substantially, with rare chains gaining the most because they borrow statistical strength from semantically similar chains. Evidence is reported on six benchmarks, with the largest gains on low-frequency chains.","feed_headline":"Grouped Hawkes attention beats six TKG benchmarks","feed_subtitle":"Casting each subject–relation pair as an event chain boosts link and time forecasts, with the biggest gains on rare chains.","key_machinery":"The central object is the Hawkes-derived analytical attention mask M_{q,k} = log(Φ̄_{u,v}) − γ̄_u Δt, which, when exponentiated inside softmax, becomes a group-Hawkes excitation–decay kernel Φ̄_{u,v} exp(−γ̄_u Δt) that modulates semantic similarity across chains. It is built from a soft group assignment W over G=4 latent groups and globally learned group-to-group excitation matrix Φ and decay vector γ, reducing quadratic cross-chain computation to O(G²).","core_discovery":"GAttNHP claims that a temporal knowledge graph can be recast as marked, mutually exciting event chains, one per subject–relation pair, and that a global group-level Hawkes prior can be converted into an exact attention mask that modulates semantic query–key similarity. Together with a non-crossing quantile head for inter-arrival times, this yields state-of-the-art entity prediction and calibrated time prediction on six TKG benchmarks, with ablations showing the group-excitation module is the primary driver and that its largest gains are on long-tail event chains.","pith_inferences":["The group-excitation gains on tail chains may be largely a semantic-smoothing or implicit data-augmentation effect; if so, a simpler group-mean intensity prior could reproduce much of the benefit, a hypothesis the paper does not test.","The NCQ head's advantage over MSE may stem as much from replacing quadratically exploding gradients with bounded pinball gradients as from genuinely better distributional calibration; a trimmed-MSE or Huber-loss control would separate these.","The framework suggests a general recipe for other sparse event-stream domains: treat each unit as a chain, let a small latent group structure mediate cross-unit excitation, and forecast with a monotone quantile head—potentially applicable to fraud detection, epidemic spread, or social media cascades."],"forward_implications":["Entity prediction on ICEWS14 improves to 0.5068 MRR, an absolute gain of 8.25 points over the prior best ECEformer, with Hits@1 nearly 10 points above TLogic.","Time prediction MAE is reduced by 47–70% compared with an internal MSE head on ICEWS datasets, while empirical 90% interval coverage stays close to nominal.","The group-interaction module's largest gains are on mid- and low-frequency chains: tail-chain MRR rises by 10.7 points over the self-attention-only variant.","The method is far lighter than strong baselines, using 2.8–8.5M parameters versus 88–107M and fitting on a single 16 GB GPU.","A mean-based time objective makes joint training unstable and collapses entity prediction, whereas the non-crossing quantile head stabilizes optimization."],"fun_headline_variants":["Group attention Hawkes process lifts TKG forecasts","Hawkes process with group attention wins on six TKG benchmarks","TKG reasoning gets a Hawkes boost: group attention excels","GAttNHP: group attention for better temporal graph forecasting","Long-tail events tamed: grouped Hawkes attention on TKGs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Cross-chain excitation is estimated from co-occurrences within a random mini-batch of 16 chains, so if truly coupled chains rarely appear together in the same batch, the learned group excitation matrix encodes batch-sampling statistics rather than genuine event dependencies.","fun_headline_variants_meta":{"raw":{"variants":["Group attention Hawkes process lifts TKG forecasts","Hawkes process with group attention wins on six TKG benchmarks","TKG reasoning gets a Hawkes boost: group attention excels","GAttNHP: group attention for better temporal graph forecasting","Long-tail events tamed: grouped Hawkes attention on TKGs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000603,"raw_usage":{"total_tokens":2656,"prompt_tokens":757,"completion_tokens":1899,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":1825}},"tokens_in":501,"tokens_out":1899,"duration_ms":12296,"temperature":1.0,"reasoning_tokens":1825,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T01:14:12.814809+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Randomly permute chain identities within each mini-batch while preserving each chain's own event sequence, so group assignments become meaningless relative to the actual chains; if the group module's gains on tail chains persist unchanged, the learned excitation is not capturing semantically meaningful cross-chain structure.","supporting_citations":[],"review_version":1}