{"id":"cb05a05a-5a1b-4208-903b-9bafb38c9dbd","arxiv_id":"2505.11803","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"VITA uses a (conjunction, start, end) time triplet and an encoder-decoder Transformer to predict missing entities, relations, times, and numeric literals in temporal hyper-relational knowledge graphs, outperforming baselines on four datasets.","lead":"This paper proposes VITA, a way to represent the validity period of facts in knowledge graphs as one of four time types (since, until, a period, or always valid) and a Transformer model that learns from this representation. On link prediction benchmarks it reports better accuracy than existing models, with the largest gains when predicting the time values themselves.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains rest on custom dataset construction that may favor VITA, and missing code/data plus omitted VITA hyperparameters make this untestable; a controlled re-run with time qualifiers stripped is needed.","rationale":"The reader's CONDITIONAL verdict is appropriate. The model itself is internally coherent: the encoder/decoder design, masked training, and ablations are plausible, and the versatile time triplet is a reasonable single-interval representation. The main weakness is evidential: the empirical comparison is run on datasets the authors constructed, and one step in §5.1.1 is under-specified (whether time-valued qualifiers are removed), so the flagship 75.3% could come from the benchmark transformation rather than from modeling temporal validity. I verified that 75.3% is the average of the four relative time-MSE improvements (69.6%, 65.5%, 67.2%, 98.9%), not a maximum as the abstract's 'up to' wording implies. Without code/data release and VITA hyperparameters, this cannot be resolved from the paper. Still, there is no clear logical flaw in the core representation or architecture, so the right outcome is to keep the paper conditional: require the controlled re-run and release before fully accepting the empirical claim. Hence the verdict is unchanged.","tokens_in":20153,"tokens_out":10878,"duration_ms":111604,"concrete_test":"Release the dataset construction scripts and final datasets, and run one controlled comparison: build Wiki/YAGO variants that keep original timestamps (no merging into Period), remove all time-related qualifiers (start time, end time) from the qualifier set, and give VITA and HyNT identical hyperparameter search budgets; then recompute time-prediction MSE. If VITA's advantage over HyNT does not persist, the gains are an artifact of the custom benchmark transformation rather than the versatile time representation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"VITA's central empirical claim is that modeling temporal validity explicitly is what produces large link-prediction gains (up to 75.3%). That claim is supported only by three custom datasets whose construction, described in §5.1.1, includes steps that can systematically favor VITA: (1) YAGO and Wiki facts are merged from timestamped/expanded forms into Period triplets, removing the redundancy that TKG baselines expect and making the 'compactness' comparison partly definitional; (2) qualifiers are crawled from Wikidata, and for YAGO time-containing qualifiers are used to update the time triplet, but the paper never states whether those time-valued qualifiers are removed from the qualifier set—if they remain, VITA receives the same time information through both the time triplet and the qualifier encoder, while the only time-prediction baseline (HyNT) receives it only as qualifiers; (3) baselines are handicapped asymmetrically (TKG baselines lose qualifiers, HKG baselines lose time). Since neither the datasets nor VITA's own hyperparameters are released, the 75.3% average time-prediction gain cannot be checked independently. This is not an internal inconsistency, but it is the load-bearing support for the superiority claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a versatile time representation for temporal hyper-relational knowledge graphs, encoding the temporal validity of a fact as a time triplet (c, t1, t2) with c in {Since, Until, Period, Invariant}, which covers the four common single-interval validity types including unbounded intervals. It then introduces VITA, an encoder-decoder model with separate encoders for the base triplet, the time triplet, and the qualifiers, a Time Value Encoder to handle both real-valued and infinite time values, and a TimeSpan Fuser that injects validity-span features into the base-triplet and qualifier decoders. The model is trained with a masked strategy and evaluated on entity, relation, time, and numeric-literal prediction over four datasets (Wiki, YAGO, wikipeople, ICEWS14), reporting consistent gains over a substantial set of TKG and HKG baselines, with the largest advantages in time prediction relative to HyNT. Ablation studies that remove the time triplet and the TimeSpan Fuser support the key design choices.","tokens_in":20414,"tokens_out":9194,"duration_ms":88650,"significance":"The central claim—that explicitly modeling the type and span of temporal validity, rather than treating time as an ordinary numeric qualifier, improves link prediction—is well-motivated and the proposed representation is simple and intuitively appealing. The paper's strengths include a broad comparison against state-of-the-art TKG and HKG baselines, consistent improvements across four datasets, and ablations that isolate the contributions of the time triplet and TimeSpan Fuser. If the results are reproducible, the work offers a useful building block for temporal hyper-relational KG reasoning. However, the empirical support depends on custom datasets whose construction may advantage VITA, the model's own hyperparameters are not reported, and code and data are not released, so the reported gains cannot currently be independently verified.","major_comments":[{"comment":"The dataset-construction procedure for YAGO and Wiki may give VITA an information advantage over HyNT. The text states that for YAGO \"a few collected qualifiers here also contain time information, we update the time information of the corresponding facts according to the (latest) qualifiers we collected from Wikidata,\" but it does not state whether these time-valued qualifiers are subsequently removed from the qualifier set. If they remain, VITA receives time information through both the time triplet and the qualifier encoder, while HyNT receives it only as qualifiers; the reported 75.3% average improvement in time prediction could then be partly due to redundant input rather than the representation. The authors should clarify the removal policy and report an ablation in which time-valued qualifiers are stripped from the qualifier set for both VITA and HyNT.","section":"§5.1.1"},{"comment":"VITA's own hyperparameters are never reported. The appendix provides hyperparameter settings for all baselines but not for VITA (embedding size, number of self-attention layers and heads, learning rate, batch size, dropout, training epochs, and the lambda value per dataset). Combined with the absence of code and dataset release, this makes the experimental results unreproducible. Please provide a complete hyperparameter table and release the datasets and code, or at minimum make the dataset-construction code and exact train/validation/test splits available.","section":"Appendix A / §5.1.2"},{"comment":"The time and numeric-literal prediction comparison appears to lack a common protocol. The paper does not describe how HyNT is configured for time prediction (for example, whether it predicts a softmax distribution over discretized numeric tokens or a real-valued regression) nor how the reported MSE is computed relative to the normalized time values used in TVE. A mismatch in the prediction target or evaluation scaling would make the large reported improvement (75.3% average relative MSE reduction) difficult to interpret. Please specify the exact prediction and evaluation procedure for each baseline on these tasks, including normalization and mapping back to the original time scale.","section":"Table 4 / §5.2.3"},{"comment":"The claim of \"up to 75.3%\" improvement is ambiguous and arguably inaccurate. From Table 4, the relative time-MSE improvements over HyNT are about 69.6% on Wiki, 65.5% on YAGO, 67.1% on wikipeople, and 98.9% on ICEWS14; thus 75.3% is the average improvement across the four datasets, not the maximum. Please report per-dataset improvements and adjust the wording to \"average improvement of 75.3%\" or state the actual maximum, so the abstract is not misleading.","section":"Abstract / §5.2.3"}],"minor_comments":[{"comment":"The \"Completeness\" property is overstated. A single time triplet (c, t1, t2) represents one contiguous validity interval, possibly unbounded; facts with multiple disjoint validity intervals (for example, two non-consecutive terms in office) cannot be represented. Please qualify the completeness claim by specifying that it holds for a single validity interval per fact.","section":"§3, Definition 3.1"},{"comment":"The paper states that infinity tokens are not predicted because the conjunction uniquely determines them, but Definition 3.3 includes t1 and t2 as possible missing elements. Please clarify whether test queries with missing infinity tokens are excluded from the time-prediction evaluation and how this affects the reported MSE.","section":"§4.3"},{"comment":"The time granularity for expansion-based baselines is set to 1 year, but ICEWS14 originally has daily timestamps. Please confirm that 1 year is the finest granularity used and discuss the sensitivity of the results to this choice, since coarser pooling could disadvantage baselines on short-duration events.","section":"§5.1.2"},{"comment":"On Wiki, GRAN achieves a higher MRR (0.9861) than VITA (0.9856), and on YAGO HyNT is higher (0.9483 vs 0.9462), yet the text describes VITA as among the best-performing. Please clarify the ranking or add a footnote stating that VITA is best on average but not on every dataset.","section":"Table 3"},{"comment":"There are several typographical and consistency issues: \"TKHG\" in Definition 3.3 should be \"THKG\"; \"TSE\" in Section 4.2 should likely be \"TSF\"; the spelling of \"TeRo/TeRO\" is inconsistent; and the reference numbering in the related-work text appears duplicated for [62]. A careful proofread would improve presentation.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The paper reports interesting results, but the empirical contribution rests on three custom datasets that are not released. The dataset-construction issue raised in major comment 1 is critical: if time-valued qualifiers are retained, the advantage over HyNT may be inflated. I would strongly encourage the editor to require the authors to release the datasets and code, or at minimum to perform the stripped-qualifier ablation and report VITA's hyperparameters, before the paper can be considered for acceptance. The missing hyperparameters and lack of code/data are also serious reproducibility concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"VITA is worth a serious look, not because the 75.3% number is meaningful, but because the four-type time triplet is a genuinely convenient way to talk about temporal validity in hyper-relational KGs. The paper makes a clear, well-structured case that time should be a first-class channel rather than another numeric qualifier. The design is straightforward: encode base triplet, time triplet, and qualifiers separately, let them interact through self-attention, and use a time-span fuser for entities and qualifiers. That's a sensible extension of HyNT-style transformers, and the infinity-token extension for Since/Until/Invariant is a small but useful trick. The ablations (w/o VTR and w/o TSF) support the design choices, and the case study is illustrative.\n\nThe soft spots are in the evidence. The evaluation runs on three custom datasets built by the authors, with a merging and qualifier-crawling pipeline that is not fully specified. The most concrete worry: when time-valued qualifiers found on Wikidata are used to update the time triplet, the paper doesn't say whether they are then removed from the qualifier set. If they remain, VITA sees the same time information through two channels, while HyNT sees it only as qualifiers. That asymmetry could inflate the reported time-prediction advantage. The merging of YAGO and Wiki from expanded timestamps into Period triplets also changes what the TKG baselines see, though the authors do average over expanded facts.\n\nReproducibility is the larger problem. No code, no data, and VITA's own hyperparameters are nowhere in the appendix. The baselines all get their hyperparameters listed; VITA doesn't. And there are no variance or significance numbers, which matters because most entity/relation gains are single-digit percentages. The abstract's \"up to 75.3%\" is also misleading: that is the average improvement in time prediction over a single baseline (HyNT), not the best case across tasks. On ICEWS14 the time MSE improvement is actually ~99%.\n\nNone of these are fatal to the central idea. The representation is clean and defensible, and the architecture is plausible. But as submitted, the empirical claim is untestable. A revision needs to release code and data, state the handling of time-valued qualifiers, report VITA's settings and seeds, and rephrase the 75.3% claim. I'd send it to peer review—the idea deserves scrutiny—but I wouldn't rely on the numbers until the pipeline is transparent. For my own work, I'd probably cite the representation but not the experimental results.","headline":"A clean time-triplet representation worth publishing, but the empirical advantage is untestable without code/data and the headline 75.3% is an average over a single baseline on custom data.","tokens_in":20945,"tokens_out":3576,"would_cite":true,"duration_ms":33302,"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":"A single time triplet $(c,t_1,t_2)$ represents since, until, period, and invariant fact validity, and VITA learns from time value plus timespan to outperform baselines by up to 75.3%.","keywords":["temporal knowledge graph","link prediction","hyper-relational knowledge graph","time representation","temporal validity","knowledge graph embedding","time value encoder","numeric literals"],"falsifier":"A reader could settle the claim by sampling facts directly from a public knowledge graph—open-ended start times, open-ended end times, and no time qualifiers—without the paper's merge-and-map pipeline, and comparing VITA's time-prediction error against the best qualifier-based baseline; if the reported 75.3% advantage shrinks to a few percent once the baseline is also given infinity handling, the versatile representation is not what carries the result.","tokens_in":19948,"feed_emoji":"🕒","tokens_out":11099,"duration_ms":99452,"temperature":0.7,"pith_summary":"The paper argues that existing ways of attaching time to knowledge-graph facts—temporal snapshots, granularity-based expansion, or treating timestamps as ordinary numeric qualifiers—either depend on ad-hoc interval choices or discard facts with infinite validity. It proposes a versatile time triplet $(c,t_1,t_2)$ with $c\\in\\{\\mathrm{Since},\\mathrm{Until},\\mathrm{Period},\\mathrm{Invariant}\\}$ and $t_1,t_2$ real numbers or $\\pm\\infty$ tokens, claimed to cover exactly the four kinds of temporal validity a fact can have. Building on this representation, VITA is an encoder–decoder model that learns time values and validity timespan separately and, under one masked training scheme, predicts missing entities, relations, time values, and numeric literals. The paper reports that VITA outperforms the best applicable baselines by 4.4–8.7% in entity prediction and by up to 75.3% in time prediction, with ablations showing the representation itself, not the architecture alone, drives much of the gain.","feed_headline":"One time triplet covers every fact validity type","feed_subtitle":"New model VITA turns open-ended validity into tokens and beats baselines by up to 75.3% on time prediction.","key_machinery":"The load-bearing object is the versatile time triplet $(c,t_1,t_2)$, whose conjunction $c$ selects one of four validity types and whose values are either real numbers or learned $\\pm\\infty$ tokens; completeness comes from allowing open-ended intervals instead of only closed periods. Two mechanisms carry the learning: the Time Value Encoder (TVE), which maps real-valued times through a min-max-normalized unit embedding and reserves dedicated embedding vectors for $\\pm\\infty$, and the TimeSpan Fuser (TSF), a translation-invariant learnable sinusoidal encoding, extended to infinity tokens, that injects the validity span into base-triplet and qualifier features. The design choice that prevents information leakage is that TSF features are fused into the entity and qualifier decoders but not into the time-triplet decoder, so predicting $t_1$ or $t_2$ must rely on context rather than on a transformed copy of the answer. These mechanisms together let a single masked training pass supervise all prediction heads from one factorized representation of the fact.","core_discovery":"The central discovery is representational: every temporal validity of a fact can be written exactly as a time triplet $(c,t_1,t_2)$, where the conjunction $c$ encodes the validity type—valid since $t_1$, valid until $t_2$, valid in the period $[t_1,t_2]$, or always valid—and the two time values may be real numbers or special $\\pm\\infty$ tokens. This turns a temporal hyper-relational fact into $\\{(s,r,o),(c,t_1,t_2),\\{(k_i,v_i)\\}\\}$ and removes the need to choose a snapshot interval or a timestamp granularity, because the fact is stored once at whatever precision the data has. The paper argues that the gains come from treating the time triplet as a first-class component: VITA separates the base triplet, time triplet, and qualifiers into three encoders, lets self-attention exchange information among them, and feeds the validity timespan into entity and qualifier prediction while deliberately withholding it from time prediction to avoid leaking the answer. The claimed result is that this design beats the best-performing baselines in all four link-prediction tasks, with the largest margin in predicting time values.","pith_inferences":["The paper leaves implicit that the time triplet is a representation layer, not a model; any hyper-relational encoder that adopts the triplet and infinity-token handling could plausibly inherit part of the reported advantage, independent of VITA's specific transformer blocks.","A testable extension: the benchmarks are dominated by Period facts (about 83–100%), with almost no Invariant cases on two of them, so a dataset where Since, Until, and Invariant are the majority would directly stress-test the completeness claim that the current experiments cannot fully confirm.","VITA is evaluated only in the interpolation setting (missing elements at historical timestamps); adapting the triplet representation to extrapolation, i.e., forecasting links after the training window, is a natural next step the paper does not pursue.","Treating open-ended validity as a token rather than a missing value suggests a model-agnostic normalization idea for temporal knowledge graphs: other numeric qualifiers that can be undefined or infinite might benefit from the same token-versus-value handling."],"forward_implications":["Temporal link prediction no longer needs a hand-chosen snapshot interval or granularity; one representation serves facts of any precision and any validity length.","Facts that are valid forever, or have only a start or only an end, become ordinary training examples instead of being dropped or artificially truncated, because $\\pm\\infty$ are explicit tokens.","A single masked model answers entity, relation, time, and numeric-literal queries from one training pass, so applications need one pipeline rather than separate specialists.","Storing each fact once instead of expanding it over its validity removes the oversampling bias and the out-of-memory failures that long-validity facts cause in expansion-style baselines.","The large reported margin in time prediction indicates that validity timespan is a strong learnable signal, not a passive annotation."],"supporting_citations":[{"why":"The transformer-based hyper-relational baseline that represents time values as ordinary numeric qualifiers; it is the principal comparison target on all four prediction tasks and on the 75.3% time-prediction margin.","marker":"[6]"},{"why":"Supplies the translation-invariant sinusoidal time encoding that the TimeSpan Fuser extends to accommodate infinity tokens.","marker":"[57]"},{"why":"Establishes the hyper-relational representation of facts as base triplets plus qualifier pairs, which VITA adopts and augments with a dedicated time triplet.","marker":"[42]"},{"why":"Supplies the seed temporal-KG facts for the first benchmark, which the paper extends by crawling qualifiers and merging open-ended validity into Period triplets.","marker":"[24]"},{"why":"Supplies the yearly-expanded temporal-KG facts that are merged into Period triplets for the second benchmark.","marker":"[22]"},{"why":"Supplies the n-ary hyper-relational facts that are converted into base triplets with qualifiers and time information for the third benchmark.","marker":"[20]"},{"why":"Supplies the event triplets used as a qualifier-free benchmark, after merging facts to fit the versatile time representation.","marker":"[16]"},{"why":"An expansion-style temporal baseline whose out-of-memory behaviour on long-validity facts illustrates the computational cost the compact triplet avoids.","marker":"[38]"},{"why":"A temporal hyper-relational baseline incorporating time-invariant relational knowledge, against which VITA's handling of open-ended validity is compared.","marker":"[12]"}],"fun_headline_variants":["One time triplet fits every fact validity type","Time triplet replaces snapshots, beats baselines by 75.3%","Triplet time encoding handles since, until, period, always","No granularity choice: VITA encodes time as a triplet","Up to 75.3% better time prediction via triplet encoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, introduced with the dataset construction, is that the three custom benchmark datasets—built by crawling qualifiers and time information from public knowledge graphs and manually mapping relation identifiers, with about 0.2% of unmappable facts dropped—are fair and unbiased for comparing VITA against models designed for other time representations; if the construction favours VITA's triplet, the reported gains would not transfer.","fun_headline_variants_meta":{"raw":{"variants":["One time triplet fits every fact validity type","Time triplet replaces snapshots, beats baselines by 75.3%","Triplet time encoding handles since, until, period, always","No granularity choice: VITA encodes time as a triplet","Up to 75.3% better time prediction via triplet encoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000255,"raw_usage":{"total_tokens":1663,"prompt_tokens":1128,"completion_tokens":535,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":744,"completion_tokens_details":{"reasoning_tokens":449}},"tokens_in":744,"tokens_out":535,"duration_ms":5519,"temperature":1.0,"reasoning_tokens":449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:46:54.089396+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle the claim by sampling facts directly from a public knowledge graph—open-ended start times, open-ended end times, and no time qualifiers—without the paper's merge-and-map pipeline, and comparing VITA's time-prediction error against the best qualifier-based baseline; if the reported 75.3% advantage shrinks to a few percent once the baseline is also given infinity handling, the versatile representation is not what carries the result.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The transformer-based hyper-relational baseline that represents time values as ordinary numeric qualifiers; it is the principal comparison target on all four prediction tasks and on the 75.3% time-prediction margin."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the hyper-relational representation of facts as base triplets plus qualifier pairs, which VITA adopts and augments with a dedicated time triplet."},{"cited_title":"T-GAP: Learning to Walk across Time for Temporal Knowledge Graph Completion","cited_arxiv_id":"2012.10595","evidence_quote":"Supplies the seed temporal-KG facts for the first benchmark, which the paper extends by crawling qualifiers and merging open-ended validity into Period triplets."},{"cited_title":"InProceedings of the 58th annual meeting of the association for computational linguistics","cited_arxiv_id":null,"evidence_quote":"Supplies the yearly-expanded temporal-KG facts that are merged into Period triplets for the second benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the n-ary hyper-relational facts that are converted into base triplets with qualifiers and time information for the third benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"An expansion-style temporal baseline whose out-of-memory behaviour on long-validity facts illustrates the computational cost the compact triplet avoids."}],"review_version":1}