{"id":"973e12e4-e32b-44cc-ba92-2ee8bd13a997","arxiv_id":"2608.06765","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LiFTER predicts future links in continuous-time dynamic graphs with a signed sum of grounded temporal rule executions, matching neural baselines on historical-negative forecasting while providing exact, replayable evidence for every score.","lead":"LiFTER is a neuro-symbolic predictor for time-varying networks that computes every link score as a signed sum of explainable rule executions over raw historical facts, so each prediction can be replayed and edited fact by fact. It matches or beats neural baselines on the harder historical-negative evaluation and achieves the best explanation accuracy and deletion fidelity among five systems, effectively turning link forecasting into a transparent microscope.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The validation-selected H=10 local fact window caps which historical evidence can ever ground a rule; longer-range queries are invisible to every term in Eq. (8).","rationale":"The paper's central claim has two parts: a grounded score/explanation identity and competitive forecasting. The identity part is well supported by construction, by Eq. (8), and by the verifier's corruption detection in Table 8; I do not see a soundness flaw there. The forecasting half is credible but rests on the finite H window chosen by validation. The reader identified this same dependency, and my read of Sections 4.1, 4.2, and 5.1 agrees: the locality premise is the least secure load-bearing condition. Because the paper already validates H on an interval disjoint from test and ships code, the correct disposition is a conditional acceptance with a direct held-out sensitivity analysis rather than rejection. No stronger objection surfaced after checking the negative-sampling description, the exact Shapley computation, and the explanation protocol.","tokens_in":17365,"tokens_out":13808,"duration_ms":128987,"concrete_test":"Re-run the full Table 2 protocol with H in {5, 10, 20, 40, 80} on the held-out test intervals only, reporting historical AUC/AP and dataset ranks for all four datasets. If the best test configuration differs from H=10 by more than seed-level noise, or if LiFTER's rank against the neural baselines changes, the H=10 conditional claim is brittle; if test metrics are flat across H, the locality objection does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 defines the executable fact database as the most recent H=10 facts adjacent to the query source and candidate endpoints (Eq. 1a), and Section 5.1 selects H=10 by chronological validation. All rule schemas in Section 4.2, including pair renewal and positioned recurrence, can ground only in this window. Therefore any decisive interaction older than the ten most recent local events is not representable by any term in Eq. (8), and the corresponding forecast and explanation are incomplete by construction. This is not an internal inconsistency, and the reported benchmark results may be correct, but the central claim of a grounded, editable CTDG predictor is only as general as this locality premise. The premise is established on four 32,768-event prefixes and is not demonstrated to transfer to streams with longer-range recurrence or higher source activity. The paper would be materially stronger with direct held-out evidence that the chosen H bounds test performance, not merely validation performance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"LiFTER is a neuro-symbolic predictor for continuous-time dynamic graphs. It represents each historical interaction as a grounded fact and defines a finite set of rule schemas (endpoint binding, pair renewal, positioned recurrence, ordered transitions). The candidate logit is exactly a signed sum of grounded rule executions plus a prior, and the paper reports that an independent verifier reconstructs all 19,664 test logits within tolerance 2e-5. LiFTER is evaluated on four CTDG benchmarks under a historical-negative protocol, where it ranks first on both AUC and AP for Reddit and LastFM, second on Wikipedia, and second in AUC and first in AP for MOOC. The same architecture is used as a microscope via exact Shapley decomposition over all 128 execution coalitions, query-regime analysis, and grounded-fact intervention. A synthetic experiment in Appendix A studies when a learned typed predicate vocabulary (K>1) is beneficial.","tokens_in":17651,"tokens_out":9930,"duration_ms":81145,"significance":"If the claims are confirmed, the paper's main contribution is a forecasting architecture whose output and explanation are the same object: a replayable, editable, grounded computation. The strongest evidence is the independent verifier reproducing all test logits, which is a machine-checkable claim, and the exact Shapley decomposition with residual 1.11e-16. The paper also ships public code, configuration scripts, and a data pipeline. The main limitations are the local H=10 fact window, which bounds what can be explained, and some underspecified training details; both are fixable and do not undermine the central identity in Eq. (8).","major_comments":[{"comment":"Eq. (1a) defines the fact database as all pre-query facts, but Section 4.1 states that the executable database contains only the most recent H facts adjacent to the query source and candidate endpoints. This discrepancy is load-bearing because the 'grounded' claim applies only to that bounded window; any decisive interaction older than the ten most recent local events is invisible to every term in Eq. (8). Please revise Eq. (1a) to describe the actual local database, and add held-out test evidence (e.g., test AUC vs H) that the chosen H is adequate, rather than relying only on the validation-based selection in Table 1.","section":"Section 3, Eq. (1a) and Section 4.1"},{"comment":"The setup paragraph states both 'history length 128' and later selects H=10 as the grounding capacity. Since Eq. (8) and the verifier claim depend on H, the exact experimental configuration must be unambiguous; please clarify what 'history length 128' refers to or correct it.","section":"Section 5.1"},{"comment":"The training objective uses sampled negative logits, but the paper never specifies how training negatives are sampled. This matters because the learned weights and the explanation traces depend on the training negative distribution, and only the evaluation protocol is described in Section 5.1. Please specify the training negative sampling procedure (e.g., uniform random, historical, or mixed).","section":"Section 4.4, Eq. (9)"},{"comment":"The entity vectors a_z and b_y are not defined. The paper should state explicitly whether they are static per-entity learned embeddings, time-dependent representations, or something else, and how they are trained; without this, the transition potentials and the reported parameter count cannot be reproduced.","section":"Section 4.2, Eqs. (6a)-(6b)"},{"comment":"The explanation comparison restricts every method to a shared evidence universe consisting of the most recent 10 events adjacent to the query source and candidate, which is exactly LiFTER's internal window. The post-hoc explainers were designed to explain models that see the full history; restricting them to LiFTER's window may disadvantage them and inflate LiFTER's relative explanation quality. Please justify this protocol or include a comparison with a larger evidence set.","section":"Section 5.6, Table 7"}],"minor_comments":[{"comment":"The row for H=10 lacks a space between the Reddit and MOOC values ('81.2585.09'), making the table hard to read.","section":"Table 1"},{"comment":"Several entries have missing spaces between numbers (e.g., MOOC row '−0.43%)85.07' and 'LastFMLiFTER'); please fix the formatting.","section":"Table 2"},{"comment":"The caption uses 'Mooc' but the paper elsewhere consistently uses 'MOOC'.","section":"Figure 4 caption"},{"comment":"The capped sum expression 'E_r = min(e4, sum_g e_r(g,q))' should be written as 'exp(4)' or 'e^4' for clarity.","section":"Section 4.3"},{"comment":"The term 'historical-candidate coverage' would be clearer as 'the percentage of queries for which a historical negative destination is available'.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The Section 5.1 'history length 128' appears to be an internal inconsistency; please verify against the released code before publication. The exact-verification claim is strong, but the local H=10 window should be tested for sensitivity on the held-out test set, not only on validation. The explanation comparison protocol also deserves scrutiny, as restricting all explainers to LiFTER's 10-event window may bias the comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read LiFTER. Short version: the core construction holds up. The rule language over single-relation CTDGs — endpoint bindings, pair renewal, position, ordered transitions — is genuinely new relative to TLogic/TILP/Neural-LP, and the logit identity of Eq. (8) is exact by construction. The independent verifier reproducing all 19,664 test logits with max error 1e-5 is real evidence, not a sketch. The paper also earns credit for the Shapley decomposition over all 2^7 coalitions, the fact-intervention analysis, and the honest negative result that K=1 suffices on these streams. The Appendix phase diagram is a nice controlled test for when typed predicates add value.\n\nNow the soft spots. The stress-test note lands: the grounding window H=10, chosen by chronological validation, caps what any rule can see. If a decisive interaction is older than the ten most recent local events, no term in Eq. (8) can represent it. That is a real limitation of the 'grounded microscope' claim, not a fatal inconsistency. The paper is open about H being a hyperparameter, but it never shows held-out evidence that H=10 bounds test performance. On high-activity nodes the window fills up in hours; longer-range recurrence is invisible. I'd want a sensitivity plot of test AUC/AP vs H, plus a statement of coverage (what fraction of test positives actually have their citation within H).\n\nSecond soft spot: the training-negative distribution is never specified. The evaluation uses historical negatives, meaning candidates are drawn from the source's pre-query destinations. But the loss in Eq. (9) samples negatives during training — random? historical? how many? That's a reproducibility gap. The paper says candidates are identical across models and seeds, which is good for comparison, but the training protocol matters for interpreting the 'competitive' claim.\n\nThird, smaller: the neural baselines are trained for 10 epochs, and the paper reports LiFTER's relative gains on LastFM as large. Given CRAFT and GraphMixer are known to need more epochs, I'd be cautious about claiming superiority over tuning. The paper is fair about it, but it undercuts the 'competitive' language.\n\nOverall: I don't see a load-bearing flaw. The forecasting claim is conditional on training protocol reproducibility; the explanation and verifier claims stand on their own. This paper deserves a serious referee. I'd bring it to reading group.\n\nRecommendation: engage.","headline":"A rare CTDG paper where prediction and explanation are the same object — the logit identity is real and verified, but the H=10 locality premise is the main thing to push on.","tokens_in":18101,"tokens_out":2002,"would_cite":true,"duration_ms":16861,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Future-link scores can be written as signed sums of grounded facts, making each forecast an auditable computation.","keywords":["continuous-time dynamic graphs","temporal link prediction","neuro-symbolic learning","grounded rule execution","rule induction","explainable artificial intelligence","verifiable computation","temporal graph forecasting"],"falsifier":"Take a held-out test query and locate a decisive fact older than the ten most recent facts around the source and candidate; if re-running LiFTER with that fact moved into the local window changes the score and trace while the ground-truth label is fixed, the locality premise is violated. Alternatively, hand-recompute any of the 19,664 test logits from raw history and frozen parameters: a discrepancy above $2\\times10^{-5}$ would contradict the verifiability claim.","tokens_in":17184,"feed_emoji":"🔬","tokens_out":12563,"duration_ms":103313,"temperature":0.7,"pith_summary":"LiFTER sets out to make continuous-time dynamic graph forecasting—predicting which entity will next interact with which other entity from a timestamped interaction stream—a verifiable grounded computation rather than a neural compression of history. The paper claims that every candidate score can be expressed as a prior plus a signed sum of rule executions, where each execution cites concrete historical facts, entity bindings, and temporal order, so the prediction and its explanation are the same forward object. If this is right, a forecast can be inspected, independently recomputed from raw history, and edited by removing or changing a cited fact, with no separate post-hoc explanation step. This matters because it dissolves the usual trade-off between accurate temporal link prediction and explainability: the explanation is not an approximation of the score, it is the score.","feed_headline":"Future-link scores are now signed sums of grounded facts","feed_subtitle":"LiFTER turns each prediction into an executable trace an independent checker can rebuild from raw history.","key_machinery":"The load-bearing mechanism is the grounded fact database $\\mathrm{Link}(u,v,t)$ plus a finite executor over it, with rule identity defined without semantic predicates: endpoint-bound unary rules test the six orientations in which a historical fact shares an argument with the query source or destination; pair-renewal rules require the same pair twice and compare the recurrence interval with the gap to the query; positioned-recurrence rules keep the source-local position $j$ of the repeated pair; and ordered one- and two-event transitions score a candidate from the destinations recently reached by the source. Each valid grounding receives temporal evidence $e_r(g,q)$ from a kernel that measures how closely its log-time gap matches the rule's learned preferred gap, and each rule aggregates these values into $E_r(q)$. Because every nonzero term in Eq. (8) names the facts and conditions that produced it, the complete forward pass is itself the execution trace, and the independent verifier re-runs exactly this grounding and scoring from raw history and frozen weights.","core_discovery":"The central claim is that a single-relation continuous-time dynamic graph, which has no semantic relation vocabulary, still supports a finite grounded rule language whose rule identity comes from query-relative entity bindings, pair renewal, source-local history position, and candidate-conditioned transitions. Trained weights make Eq. (8) exact: $s(X,Y,T_q)=b+\\sum_r w_r E_r+\\sum_{j,k} u_{j,k} R_{j,k}+\\sum_m c_{tr,m}$, with every non-prior term tied to specific facts, bindings, and temporal conditions. As evidence for the claim, the paper reports that an independent verifier recomputes all 19,664 test-prediction logits from raw history and frozen parameters with maximum error 0.0000131, detects deliberate corruption of cited facts, timestamps, entities, executions, or signed contributions 100% of the time, and that LiFTER ranks first on historical-negative AUC and AP for Reddit and LastFM while remaining within a few points of the strongest neural baselines on Wikipedia and MOOC. The paper also uses exact Shapley decomposition over all 128 execution coalitions to attribute dataset-level forecasting to component rule types: pair renewal on Wikipedia and Reddit, two-event transitions on MOOC, and one-event transitions on LastFM.","pith_inferences":["A consequence the paper leaves implicit is that the locality bound $H=10$ could become the main scaling constraint on longer-horizon streams; a hierarchical or global fact index that keeps older decisive facts inside the grounding set would extend the same signed, replayable ledger without changing the accounting.","The human-readable JSON certificate suggests a compliance use: deployed link-prediction systems could store per-decision fact ledgers and answer regulator or user questions such as 'why this pair, at this time' with a concrete replayable computation.","The reported competition between one- and two-event transitions on MOOC and LastFM raises a testable question about user behavior: whether coarse- and fine-grained transition evidence capture different intents, which could be checked with query-regime annotations beyond the median splits used in the paper.","The same grounded-rule construction could be carried over to relation-typed temporal knowledge graphs by letting the typed predicate vocabulary carry relation identity, potentially unifying the CTDG and TKG rule-learning settings."],"forward_implications":["Explanations no longer require a separate search: each prediction emits a trace of supporting and opposing facts with signed contributions, and the paper reports that three selected facts reproduce the original binary decision for 87.74% of queries.","Editing a cited fact re-executes the whole program, so counterfactuals and fact-level interventions are direct operations on the explanation rather than retraining or surrogate analysis.","The component decomposition gives dataset-level diagnoses: recurrence intervals dominate Wikipedia and Reddit, two-event transitions dominate MOOC, and one-event transitions dominate LastFM.","Forecasting quality remains competitive with neural baselines on historical negatives, so grounding does not appear to be bought at a large accuracy cost.","The optional typed predicate vocabulary should be used only when hidden event types are both identifiable from pre-event context and predictive of future outcomes; otherwise a single predicate is the more concise executable representation."],"supporting_citations":[{"why":"Supplies the memory-based TGN baseline that LiFTER must match on forecasting.","marker":"[1]"},{"why":"Supplies the GraphMixer baseline, the strongest neural result LiFTER trails narrowly on MOOC.","marker":"[3]"},{"why":"Supplies the candidate-conditioned CRAFT baseline and the strongest neural competitor on LastFM.","marker":"[15]"},{"why":"Supplies the repeated-edge memory baseline and the historical-negative evaluation protocol under which LiFTER ranks first.","marker":"[12]"},{"why":"Supplies the post-hoc T-GNNExplainer explanation system compared on ACC-AUC and deletion fidelity.","marker":"[5]"},{"why":"Supplies the TempME motif-explanation baseline for explanation-quality comparison.","marker":"[6]"},{"why":"Supplies the TGIB information-bottleneck explanation baseline for comparison on shared evidence.","marker":"[7]"},{"why":"Supplies the SIG self-interpretable subgraph baseline and the deletion-fidelity evaluation convention.","marker":"[8]"},{"why":"Supplies TLogic as the temporal-rule-learning contrast that uses relation-typed TKG walks.","marker":"[10]"},{"why":"Supplies TILP as the temporal inductive-logic baseline whose language-bias principle LiFTER adapts to relation-free CTDGs.","marker":"[11]"}],"fun_headline_variants":["Each future link score becomes a verifiable sum of grounded facts","Neuro-symbolic LiFTER lets every prediction be rebuilt from raw history","Forecasts you can trust: LiFTER's rules are fully inspectable","Link prediction becomes a glass box: LiFTER exposes the facts","Predictions you can audit: LiFTER's signed fact sums"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that all evidence needed for a future link sits in the most recent $H=10$ facts near the query source and candidate endpoints—a capacity chosen by validation on four datasets, not demonstrated to generalize.","fun_headline_variants_meta":{"raw":{"variants":["Each future link score becomes a verifiable sum of grounded facts","Neuro-symbolic LiFTER lets every prediction be rebuilt from raw history","Forecasts you can trust: LiFTER's rules are fully inspectable","Link prediction becomes a glass box: LiFTER exposes the facts","Predictions you can audit: LiFTER's signed fact sums"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000939,"raw_usage":{"total_tokens":4060,"prompt_tokens":1036,"completion_tokens":3024,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":2930}},"tokens_in":652,"tokens_out":3024,"duration_ms":20293,"temperature":1.0,"reasoning_tokens":2930,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:29:58.881956+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out test query and locate a decisive fact older than the ten most recent facts around the source and candidate; if re-running LiFTER with that fact moved into the local window changes the score and trace while the ground-truth label is fixed, the locality premise is violated. Alternatively, hand-recompute any of the 19,664 test logits from raw history and frozen parameters: a discrepancy above $2\\times10^{-5}$ would contradict the verifiability claim.","supporting_citations":[{"cited_title":"Temporal graph networks for deep learning on dynamic graphs,","cited_arxiv_id":null,"evidence_quote":"Supplies the memory-based TGN baseline that LiFTER must match on forecasting."},{"cited_title":"Do we really need complicated model architectures for temporal networks?","cited_arxiv_id":null,"evidence_quote":"Supplies the GraphMixer baseline, the strongest neural result LiFTER trails narrowly on MOOC."},{"cited_title":"Future Link Prediction Without Memory or Aggregation","cited_arxiv_id":"2505.19408","evidence_quote":"Supplies the candidate-conditioned CRAFT baseline and the strongest neural competitor on LastFM."},{"cited_title":"Towards better evaluation for dynamic link prediction,","cited_arxiv_id":null,"evidence_quote":"Supplies the repeated-edge memory baseline and the historical-negative evaluation protocol under which LiFTER ranks first."},{"cited_title":"Ex- plaining temporal graph models through an explorer-navigator framework,","cited_arxiv_id":null,"evidence_quote":"Supplies the post-hoc T-GNNExplainer explanation system compared on ACC-AUC and deletion fidelity."},{"cited_title":"TempME: Towards the explainability of temporal graph neural networks via motif discovery,","cited_arxiv_id":null,"evidence_quote":"Supplies the TempME motif-explanation baseline for explanation-quality comparison."},{"cited_title":"Self-explainable temporal graph networks based on graph information bottleneck,","cited_arxiv_id":null,"evidence_quote":"Supplies the TGIB information-bottleneck explanation baseline for comparison on shared evidence."},{"cited_title":"Invariant Graph Representations for Continuous-Time Dynamic Graphs Under Distribution Shifts","cited_arxiv_id":"2405.19062","evidence_quote":"Supplies the SIG self-interpretable subgraph baseline and the deletion-fidelity evaluation convention."},{"cited_title":"TLogic: Temporal logical rules for explainable link forecasting on tempo- ral knowledge graphs,","cited_arxiv_id":null,"evidence_quote":"Supplies TLogic as the temporal-rule-learning contrast that uses relation-typed TKG walks."},{"cited_title":"Temporal induc- tive logic reasoning,","cited_arxiv_id":null,"evidence_quote":"Supplies TILP as the temporal inductive-logic baseline whose language-bias principle LiFTER adapts to relation-free CTDGs."}],"review_version":2}