{"id":"745aa9fc-2e39-4a63-b584-9f144368c5ca","arxiv_id":"2608.10668","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FITTER transfers temporal link prediction to knowledge graphs with unseen vocabularies by encoding relative snapshot order and relation interaction patterns instead of absolute names or dates.","lead":"FITTER is a graph neural network that learns temporal knowledge graph patterns on one graph and applies them to a different graph whose entities, relations, and timestamps were never seen during training. It combines established structural reasoning with a relative time encoding, and reports consistent gains over static baselines across 15 transfer scenarios.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-domain transfer results may violate the paper's own disjoint-vocabulary protocol for ICEWS-family settings; overlap between source and target vocabularies is not ruled out and must be verified.","rationale":"The paper's central claim has two load-bearing conditions: (1) the method is actually evaluated under disjoint vocabularies, and (2) the empirical comparison is fair. The reader's weakest assumption concerned ordinal temporal encoding and granularity transfer; that is a genuine conceptual limitation, but it is harder to settle decisively because FITTER's design deliberately treats relative ordering as the transferable signal, and the paper's time-shift theorem is correct for the encoding itself. The disjoint-vocabulary protocol issue is more directly checkable: the paper asserts 'no overlap in entities, relations, or timestamps' but does not describe any preprocessing to enforce disjointness, and the ICEWS-family datasets very likely share vocabularies and even timestamps under their standard splits. If the overlap check comes back empty, the concern does not land and the conditional verdict stands. If it comes back non-empty, the paper must be revised to report only truly disjoint settings or to show experimentally that overlap has no effect. I am not recommending rejection because the model is vocabulary-agnostic by construction, the paper includes honest appendices (GraphOracle leakage, limitations), and code/checkpoints are promised, so the flaw is fixable rather than fatal. The reader's rationale did mention the protocol concern, but did not make it the weakest assumption, hence 'partial' agreement.","tokens_in":857,"tokens_out":813,"duration_ms":121294,"concrete_test":"Download the released code and data, then compute for every one of the 15 source-target pairs the intersections V_train∩V_inf, R_train∩R_inf, and T_train∩T_inf using the exact splits used for Table 1, with special attention to ICEWS14↔ICEWS05-15. If any intersection is non-empty, re-run Table 1 after applying disjoint-vocabulary re-splitting (or remove the non-disjoint settings) and re-report MRR/H@1/H@10. If FITTER's advantage over ULTRA/TRIX disappears or shrinks materially, the fully-inductive cross-domain claim as stated is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 6 states that cross-domain evaluation ensures no overlap in entities, relations, or timestamps between Gtrain and Ginf. But for several of the 15 settings, e.g., train ICEWS14 → test ICEWS05-15 and train ICEWS05-15 → test ICEWS14, both graphs are ICEWS subsets. Under standard splits, ICEWS14's timestamps (2014) appear in ICEWS05-15's validation split (2014), and the entity/relation vocabularies overlap heavily. The paper reports no disjoint-vocabulary preprocessing, and Table 6 lists only standard dataset statistics. If overlap exists, the formal condition V_train∩V_inf = ∅, R_train∩R_inf = ∅, T_train∩T_inf = ∅ fails for these settings, so the claimed demonstration of fully-inductive inference on entirely unseen entities, relations, and timestamps is not established where it is needed most. This is the load-bearing premise for the 'first fully-inductive' claim. Target-validation tuning of α and k further allows the same-year validation to leak source temporal statistics into reported numbers, so the empirical support for 'consistently outperforms' is also at risk if overlaps are confirmed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FITTER, a vocabulary-agnostic structural model for temporal knowledge graph link prediction. It represents relations via ULTRA-style relation interaction graphs, encodes time via sinusoidal positional encodings over snapshot indices, and performs message passing over local and global entity graphs before scoring with an MLP. The authors claim FITTER is the first fully-inductive structural model for TKG link prediction, supporting cross-domain transfer to graphs with entirely unseen entities, relations, and timestamps. They prove a time-shift invariance property of the sinusoidal encoder and evaluate on 15 cross-dataset transfer settings across six benchmarks, reporting consistent improvements over INGRAM, ULTRA, and TRIX, plus transductive comparisons and ablations. Code, checkpoints, and datasets are released, and the appendix documents and corrects two data-leakage bugs in the GraphOracle baseline.","tokens_in":22522,"tokens_out":6473,"duration_ms":67548,"significance":"If the central claims are correct, FITTER would be a meaningful contribution to cold-start TKG reasoning and Semantic Web interoperability: it has a dataset-size-independent parameter count, a clean local/global temporal decomposition, and an unusually transparent appendix that identifies and fixes baseline leakage. The release of code and checkpoints is also a strength. However, the central empirical claim currently rests on three unverified pillars: strict disjointness of training and inference vocabularies, the transferability of snapshot-index time encodings across heterogeneous granularities, and the statistical stability of the reported gains. The baseline set is also limited to static-KG fully-inductive models retrained on temporal data, so the 'first fully-inductive' claim is not benchmarked against any temporal structural fully-inductive method; this is a contextual caveat rather than an error. The paper has the potential to be a solid contribution, but the reported evidence does not yet establish the headline claim.","major_comments":[{"comment":"The fully-inductive condition defined in Section 2.1, V_train ∩ V_inf = ∅, R_train ∩ R_inf = ∅, T_train ∩ T_inf = ∅, is not demonstrated for several reported settings and is likely violated. For example, the row 'Trained on ICEWS14 → ICEWS05-15' and its reverse pair both involve ICEWS subsets: ICEWS14 is the 2014 slice of ICEWS05-15, so under standard splits the entity and relation vocabularies overlap heavily and the 2014 timestamps appear in ICEWS05-15's validation and test periods. Table 6 reports only standard dataset statistics and no overlap counts or a disjoint-vocabulary preprocessing step. Unless such preprocessing was applied, these settings are at least partially transductive, and the claim of fully-inductive inference on entirely unseen entities, relations, and timestamps is not established for the settings where the paper most needs it. The authors should report exact overlap statistics for all 15 settings or rerun the experiments with explicit disjointness filters.","section":"Section 6 / Table 6 / Table 1"},{"comment":"The experimental protocol states that Q_valid on the target TKG is used to tune inference-time hyperparameters, including the local window size k and fusion weight α. This contradicts the Table 1 caption's 'without fine-tuning' and weakens the transfer claim: the model is not applied to the target graph in a fully parameter-frozen manner. More seriously, in the overlapping ICEWS settings, the target validation set contains the same entities, relations, and timestamps as the source training graph, so tuning k and α on it can leak source temporal statistics into the reported target numbers. Even for genuinely disjoint settings, choosing hyperparameters on the target validation split is a form of target adaptation and should be disclosed as such, with an analysis of how sensitive the reported conclusions are to this choice.","section":"Section 6 / Table 1"},{"comment":"The transferable time representation is based solely on the snapshot index i, not on actual timestamps. Theorem 1 proves only that the Euclidean distance between sinusoidal embeddings of two indices depends on the index difference; it says nothing about whether equal index differences in different graphs correspond to equal real time intervals. Since GDELT uses 15-minute snapshots and YAGO uses yearly snapshots, 'one snapshot apart' has entirely different temporal semantics in the two graphs. The paper calls this 'granularity-agnostic', but the ordinal encoding cannot transfer duration-dependent temporal patterns such as 'within one week' or 'after four years'. The theoretical analysis therefore does not validate cross-domain transfer; it only verifies a trigonometric identity. A controlled experiment with rescaled or relabeled timestamps would be needed to support the granularity-agnostic claim.","section":"Section 5.1 / Appendix I, Theorem 1"},{"comment":"All results are reported as single runs without error bars, confidence intervals, or significance tests. Some of the reported advantages are small relative to the likely run-to-run variance: for example, trained on ICEWS05-15 and tested on GDELT, FITTER's MRR is 18.9 versus 17.9 for both ULTRA and TRIX, and on ICEWS18 FITTER's Hits@10 is 29.7 versus 32.1 for ULTRA and TRIX, i.e., FITTER is worse. The phrase 'consistently outperforms' is not supported without multi-seed results or at least a significance analysis. The authors should report mean and standard deviation over multiple seeds, or otherwise quantify the stability of the 15-setting comparison.","section":"Section 7 / Table 1"}],"minor_comments":[{"comment":"The text says the time span ranges 'from just 1 month (GDELT) to 189 years (YAGO)', but Table 6 lists GDELT with Times = 366, which is consistent with one year at daily granularity, not one month; Section 6 also describes GDELT as covering one year. This internal inconsistency should be corrected.","section":"Section 7.1 / Takeaway 1"},{"comment":"The proof of Theorem 3 claims that d_PE ≥ T guarantees solvability of M w = g for arbitrary g, using a Moore-Penrose right inverse. This is incorrect unless M has full row rank; d_PE ≥ T alone does not ensure that, and the Moore-Penrose inverse gives a least-squares solution, not an exact solution, when rows are dependent. The theorem needs a full-rank assumption on M or a construction of frequencies that guarantees it.","section":"Appendix I / Theorem 3"},{"comment":"The table reports ICL and GenTKG results as fixed regardless of training source, with only '–' for MRR; the absence of MRR and the fact that these models are evaluated only on extrapolation tasks should be stated more prominently in the main text, since it makes the comparison less direct than Table 2's layout suggests.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the disjoint-vocabulary violation for ICEWS-family transfer settings. If the authors cannot demonstrate strict disjointness or repair the protocol by filtering overlapping entities, relations, and timestamps, the paper's central claim would collapse for those rows. I believe major revision is appropriate because the issue is potentially fixable with re-experimentation and stricter reporting, but the revision must verify the formal condition from Section 2.1 for every setting and address the target-validation tuning and single-run issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper.\n\nFirst, it does something new: it defines and evaluates fully-inductive cross-domain temporal link prediction, where the target graph can have unseen entities, relations, and timestamps. That is a real gap, and the 15-setting evaluation is unusually broad. The paper is also honest in an unusual way: the appendix details two data-leakage bugs found in GraphOracle and reports corrected numbers. Credit where it is due. The ablations, the per-relation analysis, and the Strict Recurrency subset analysis are all solid, and the authors disclose inference cost and limitations.\n\nSecond, the central claim is not yet supported. The paper asserts that training and inference graphs have no overlap in entities, relations, or timestamps, but it never demonstrates this for the ICEWS-family rows. ICEWS14 and ICEWS05-15 are both ICEWS subsets; under standard splits, entities and relations overlap heavily, and 2014 appears in both. The stress-test note is correct: if no disjointness preprocessing was done, the rows where the paper most needs to prove fully-inductive behavior are themselves transductive in disguise. That matters more than the fact that the core ingredients (ULTRA's relation interaction graph plus sinusoidal encodings) are borrowed.\n\nA second soft spot: the protocol tunes k and alpha on target validation. That is hyperparameter adaptation, not pure zero-shot transfer. It is not fatal, but it should be described as a small amount of per-target tuning. Also, all results are single-run; no error bars. And the baselines are static KG models, so the temporal contribution is established mainly through ablation, which is decent but not a replacement for a real temporal inductive baseline.\n\nThe theory is true but weak. Theorem 1 is a trigonometric identity; Theorems 2 and 3 show the scorer can express periodic functions, not that the model transfers across granularities. The ordinal snapshot encoding is granularity-agnostic in the sense that it ignores duration, which is exactly why it may fail to transfer real temporal patterns. The paper should either validate this empirically or soften the claim.\n\nWho is this for? Researchers working on inductive reasoning for temporal KGs and cold-start semantic web applications. It deserves a serious referee, but the authors must verify the disjoint-vocabulary protocol and report seed-level variance. If the ICEWS-to-ICEWS rows are indeed overlapping, narrow the fully-inductive claim to the clearly disjoint rows.","headline":"A genuinely new setting and an honest leakage appendix, but the fully-inductive claim rests on a disjoint-vocabulary protocol that the ICEWS-to-ICEWS rows likely violate.","tokens_in":23069,"tokens_out":1741,"would_cite":false,"duration_ms":19948,"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":"This paper claims that a vocabulary-agnostic structural model, FITTER, can perform temporal knowledge graph link prediction on a target graph with entirely unseen entities, relations, and timestamps, and reports consistent gains over…","keywords":["temporal knowledge graph","inductive link prediction","cross-domain transfer","vocabulary-agnostic","graph neural network","sinusoidal positional encoding","relation interaction graph","knowledge graph embedding"],"falsifier":"Construct two synthetic temporal knowledge graphs with identical snapshot counts, relation interactions, and recurrence statistics but different true time scales (for example, hourly versus yearly snapshots), train FITTER on one, and evaluate on the other; if transfer MRR stays flat the ordinal assumption survives, and if it drops sharply the relative-snapshot representation is not sufficient for cross-domain time.","tokens_in":22084,"feed_emoji":"🕸️","tokens_out":5789,"duration_ms":57165,"temperature":0.7,"pith_summary":"FITTER addresses a gap in temporal knowledge graph reasoning: existing link prediction models assume the entities, relations, and timestamps seen at inference time are already known at training time, so each model is tied to one graph and one vocabulary. The paper tries to establish that a purely structure-driven model, with no dataset-specific entity, relation, or timestamp embeddings, can be trained on one temporal knowledge graph and then applied to a completely different graph without retraining. It represents time by relative snapshot order using sinusoidal positional encodings, and relations by their interaction patterns with other relations, so nothing in the representation depends on the names of entities, predicates, or dates. Across 15 cross-dataset transfer scenarios spanning six benchmarks that differ in domain, granularity, and time span, FITTER is reported to consistently outperform the inductive baselines. If this is right, a single trained model could be dropped onto a newly constructed temporal knowledge graph and rank plausible facts there, which would make cold-start and large-scale temporal reasoning substantially more practical.","feed_headline":"One model transfers temporal link prediction across unseen graphs","feed_subtitle":"Trained on one event graph, it tops inductive baselines on five other graphs without retraining.","key_machinery":"The machinery has three coordinated pieces. First, a vocabulary-agnostic relation interaction graph treats relations as nodes and connects them by head-to-head, head-to-tail, tail-to-head, and tail-to-tail interaction types; message passing on this graph produces query-conditioned relation representations that work for relation names never seen in training. Second, a sinusoidal temporal encoding $\\mathrm{TE}(i)$ over snapshot index $i$ is time-shift invariant, meaning $\\|\\mathrm{TE}(\\tau_2)-\\mathrm{TE}(\\tau_1)\\|$ depends only on $\\tau_2-\\tau_1$, which is what lets a one-step temporal relation transfer across graphs with different timestamp vocabularies. Third, temporal-aware message passing runs on both a global entity graph and a local snapshot-window graph, fusing the two with a learned weight $\\alpha$, and a multilayer perceptron scores the fused representation together with the temporal encoding. Together these components eliminate the need for any dataset-specific entity, relation, or timestamp embedding.","core_discovery":"FITTER's central claim is that temporal link prediction can be made fully inductive by replacing absolute vocabularies with relative structure. Relation names are encoded through a relation interaction graph whose edges are structural interaction types, so a predicate is described by how it connects to other predicates rather than by a learned name embedding. Timestamps are encoded by sinusoidal positional encodings of their snapshot index, and the paper proves a time-shift invariance property: the Euclidean distance between two such encodings depends only on the difference in snapshot indices, not on the absolute values. Message passing over a global entity graph and a local windowed entity graph injects these temporal encodings into every propagation step, and the two views are fused before an MLP scores candidate quadruples. On the paper's evaluation, FITTER outperforms the compared inductive baselines on all 15 cross-domain transfer settings while recovering roughly 71-96 percent of the MRR of fully trained transductive models on in-domain benchmarks, using a model whose parameter count does not grow with the vocabulary.","pith_inferences":[],"forward_implications":["A model trained on one temporal knowledge graph can be applied directly to a newly built target graph with no shared entities, relations, or timestamps, enabling cold-start link prediction.","Training can be done on small, inexpensive graphs and transferred to large dense ones; the paper reports a model trained on the small ICEWS14 graph reaching 26.1 MRR on GDELT without retraining.","Because the approach is purely structural, cross-domain temporal transfer is available even when the target graph has no textual descriptions or language-model assistance.","The same fixed-size model handles both interpolation targets and extrapolation targets, so a single vocabulary-agnostic encoder covers settings that previously required separate transductive architectures.","Parameter count is independent of dataset size, so adding more temporal graphs does not require additional per-dataset embedding tables.","editorial extension: The ordinal snapshot encoding treats 'one snapshot apart' as the same temporal relationship whether the snapshots are 15 minutes or one year apart; a natural boundary test is to build synthetic graphs with identical snapshot counts and structure but different real durations and see whether transfer survives.","editorial extension: The paper's per-relation analysis suggests the temporal encoding matters most for relations whose events cluster in short windows, implying that for temporally uniform relations, structural transfer alone may be doing most of the work.","editorial extension: The time-shift invariance theorem implies the model can represent periodic patterns whose periods align with snapshot counts, but it cannot represent absolute elapsed time; systems that need real-duration reasoning would need an auxiliary duration channel."],"supporting_citations":[{"why":"Supplies the vocabulary-agnostic relation interaction graph that FITTER extends with temporal modules.","marker":"[5]"},{"why":"Supplies the query-conditioned message-passing engine used for entity propagation.","marker":"[36]"},{"why":"Supplies the temporal tensor scoring used inside FITTER's temporal message function.","marker":"[13]"},{"why":"Serves as a fully-inductive baseline in the cross-domain comparison.","marker":"[17]"},{"why":"Serves as a fully-inductive zero-shot transfer baseline that FITTER is evaluated against.","marker":"[34]"},{"why":"Defines the Strict Recurrency pattern used to attribute FITTER's gains to the temporal encoding.","marker":"[7]"},{"why":"Supplies the multi-frequency sinusoidal encoding analysis that motivates the temporal representation.","marker":"[27]"}],"fun_headline_variants":["Fully inductive temporal link prediction on unseen graphs","No retraining: one model predicts links on unseen temporal graphs","Vocabulary-agnostic temporal reasoning transfers across domains","Time-shift invariant model predicts on unseen graphs without retraining","Cross-domain temporal link prediction with zero retraining"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one snapshot of separation carries the same temporal meaning in every graph, so relative snapshot position can stand in for absolute time; if real duration or granularity affects the patterns being predicted, this encoding cannot transfer them.","fun_headline_variants_meta":{"raw":{"variants":["Fully inductive temporal link prediction on unseen graphs","No retraining: one model predicts links on unseen temporal graphs","Vocabulary-agnostic temporal reasoning transfers across domains","Time-shift invariant model predicts on unseen graphs without retraining","Cross-domain temporal link prediction with zero retraining"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000801,"raw_usage":{"total_tokens":3515,"prompt_tokens":933,"completion_tokens":2582,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":2505}},"tokens_in":549,"tokens_out":2582,"duration_ms":16109,"temperature":1.0,"reasoning_tokens":2505,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:44:52.869939+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct two synthetic temporal knowledge graphs with identical snapshot counts, relation interactions, and recurrence statistics but different true time scales (for example, hourly versus yearly snapshots), train FITTER on one, and evaluate on the other; if transfer MRR stays flat the ordinal assumption survives, and if it drops sharply the relative-snapshot representation is not sufficient for cross-domain time.","supporting_citations":[{"cited_title":"Advances in neural information processing systems34, 29476–29490 (2021) 20 J","cited_arxiv_id":null,"evidence_quote":"Supplies the query-conditioned message-passing engine used for entity propagation."},{"cited_title":"In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=rke2P1BFwS","cited_arxiv_id":null,"evidence_quote":"Supplies the temporal tensor scoring used inside FITTER's temporal message function."},{"cited_title":"In: International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Serves as a fully-inductive baseline in the cross-domain comparison."},{"cited_title":"In: The Third Learning on Graphs Conference (2025) FITTER: Vocabulary-Agnostic Cross-Domain Inference on TKGs 19","cited_arxiv_id":null,"evidence_quote":"Serves as a fully-inductive zero-shot transfer baseline that FITTER is evaluated against."},{"cited_title":"In: IJCAI (2024)","cited_arxiv_id":null,"evidence_quote":"Defines the Strict Recurrency pattern used to attribute FITTER's gains to the temporal encoding."}],"review_version":1}