{"id":"edb82235-9f67-4a6b-8bbe-3a191e3e8a22","arxiv_id":"2607.03918","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Interleaving fixed log-scale gap tokens with semantic IDs, plus TA-FAMAE temporal regularization, consistently beats ReSID and other SID generative baselines on Amazon sequential recommendation.","lead":"ChronoSID adds discretized inter-interaction time-gap tokens and a time-gap prediction regularizer to semantic-ID generative recommenders. The change improves next-item prediction on Amazon domains, especially when users return after long delays.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Gains may partly reflect extra encoder capacity and token-type diversity rather than genuine temporal semantics of the fixed log-scale bins.","rationale":"The reader correctly flags the hand-chosen five-bin log-scale discretization as a weak assumption and rightly issues CONDITIONAL because of modest absolute gains, missing error bars, and unreproducible artifacts. That concern is real but secondary: Table 4 already shows moderate robustness inside the log-scale family, and the paper never claims the bins are uniquely optimal. The more load-bearing gap is causal attribution. The strongest claim is not merely “numbers go up,” but that the numbers go up because inter-interaction time is now represented. The experimental design never isolates the semantic content of Disc(Δt) from the incidental effects of longer encoder sequences and an extra token type. A constant- or shuffled-gap control is the minimal experiment that would settle whether the central interpretation holds. Until that (or an equivalent) control is reported, the verdict remains CONDITIONAL; the concern does not justify REJECT because the pipeline is clean, ablations are informative, and the long-gap diagnostic is directionally consistent with the motivation. I therefore leave the reader’s CONDITIONAL verdict unchanged while sharpening the weakest link from “lucky bins” to “uncontrolled capacity / token-type confound.”","tokens_in":21299,"tokens_out":689,"duration_ms":6504,"concrete_test":"On MI and VG, retrain the exact +Gap Tokens / ChronoSID configurations of Table 3 after replacing every real gap token with either (a) a single constant dummy token or (b) a randomly shuffled gap label drawn from the same 5-bin vocabulary (preserving length and type diversity). If R@10 / N@10 remain within ~1–2 % of the real-gap ChronoSID numbers, the temporal-semantics claim weakens; if they collapse toward ReSID, the claim is supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim attributes ChronoSID’s consistent gains over ReSID (same GAOQ, same T5 backbone, same beam) to temporal awareness via fixed log-scale gap tokens interleaved on the encoder (§3 Eq. 6, §4.3.2 Eq. 30). Ablations (Table 3) show gap tokens dominate TA-FAMAE, and Figure 3 shows larger relative gains on long target-side gaps. However, the design also lengthens every history from 3L to 4L tokens and injects a new discrete token type with its own embeddings. No control replaces real Disc(Δt) tokens by (i) a constant dummy gap, (ii) randomly permuted gap labels, or (iii) a non-temporal filler token that preserves length and vocabulary size. Without such a control, the observed improvements—and the long-gap pattern—could arise from extra encoder capacity or from learning a generic “separator / recency-bucket” signal rather than from the semantic content of the chosen {1h,1d,1w,1mo} thresholds. The sensitivity sweep over G (Table 4) only varies bin count under the same log-scale family; it does not break the temporal correspondence. Thus the attribution of gains to “temporal awareness” remains under-identified.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes ChronoSID, a lightweight temporal augmentation of the standard three-stage semantic-ID (SID) generative recommendation pipeline. It keeps Globally Aligned Orthogonal Quantization (GAOQ) identical to ReSID and injects time in two places: (i) TA-FAMAE, which adds a log-gap regression auxiliary loss to field-aware masked auto-encoding of item features (§4.1, Eqs. 12–15), and (ii) fixed log-scale gap-token interleaving on the T5 encoder input, expanding each history from 3L to 4L tokens while leaving the decoder target as the next item’s three-level SID (§3 Eqs. 3–7, §4.3.2 Eq. 30). On eight Amazon-2023 leave-one-out splits the method reports consistent Recall@K / NDCG@K gains over ReSID and other SID generative baselines (Table 2), with ablations attributing most of the lift to gap tokens (Table 3), clearer relative gains on long target-side gaps (Figure 3), sensitivity on λ and G (Table 4), efficiency overhead (Table 5), popularity splits (Figure 4), and output-level SID accuracy with bootstrap CIs (Table 6).","tokens_in":21677,"tokens_out":1610,"duration_ms":17660,"significance":"Temporal intervals are a classical signal in sequential recommendation but remain under-used once both inputs and targets are discrete SID code sequences. ChronoSID’s controlled design—identical GAOQ, same T5 backbone and beam protocol as ReSID—makes the empirical comparison unusually clean for this literature. The experimental package is thorough (eight domains, component ablations, long-gap diagnostics, hyperparameter sweeps, popularity and output-level analyses, efficiency numbers). If the gains are genuinely driven by temporal semantics rather than encoder length or token-type diversity, the work supplies a simple, practical recipe that preserves the compact SID generation paradigm. That would be a useful, if incremental, contribution to generative recommendation systems.","major_comments":[{"comment":"Central attribution claim (§1, §3 Eq. 6, §4.3.2 Eq. 30, abstract): gains over ReSID are ascribed to “temporal awareness” via Disc(Δt) gap tokens. Ablations (Table 3) and long-gap groups (Figure 3) support usefulness of the injected tokens, but the design also lengthens every encoder sequence from 3L to 4L and introduces a new discrete token type with its own embeddings. No control replaces real gap labels by (i) a constant dummy gap, (ii) randomly permuted gap labels, or (iii) a non-temporal filler that preserves length and vocabulary size. Without such a control, improvements—and the long-gap pattern—could arise from extra encoder capacity or a generic separator/recency-bucket signal rather than the semantic content of the chosen {1h,1d,1w,1mo} thresholds. This is load-bearing for the paper’s main scientific claim and should be added (at least on MI/VG).","section":"§4.3.2, Table 3, Figure 3"},{"comment":"Discretization universality (§3, §4.3.2, Table 4): the method freezes G=5 log-scale bins with hand-chosen thresholds for all eight domains. Sensitivity only varies bin count G inside the same log-scale family; it does not test alternative partitions (linear, quantile, domain-adaptive, or learned). Given that absolute gains vary across domains (Table 2) and that the weakest modeling assumption is precisely this fixed binning, a short domain-adaptive or quantile-bin comparison (or an explicit limitation statement with a negative result) is needed before the gains can be confidently attributed to a general temporal mechanism rather than a binning that fits Amazon-2023 leave-one-out splits.","section":"§3, §4.3.2, Table 4"},{"comment":"Statistical reporting of main results (Table 2, §5.1): results are “averaged over five runs,” yet no standard deviations, paired significance tests, or confidence intervals appear in the main table (bootstrap CIs are given only for the MI output-level analysis in Table 6). For a multi-domain claim of “consistent” improvement over the direct baseline ReSID, at least pairwise significance (or mean±std) on the primary metrics should be reported so readers can judge whether small absolute lifts (e.g., BPC, TG) are reliable.","section":"Table 2, §5.1"}],"minor_comments":[{"comment":"Figure 1 caption and panel (b) use β / temporal decay coefficient without stating the exact functional form fitted (e.g., same-category rate ~ exp(−β·gap) or similar). A one-line equation would make the diagnostic reproducible.","section":"Figure 1"},{"comment":"Notation: Disc(·) is introduced in Eq. (4) and reused as Disc(t_l − t_{l−1}) in Eq. (28); the latter has an extra closing parenthesis. Also g^{(Disc(...))} vs g_l notation is slightly inconsistent between §3 and §4.3.2.","section":"§3 Eq. (4), §4.3.2 Eq. (28)"},{"comment":"Table 3 ChronoSID numbers on MI/VG are slightly higher than the corresponding ChronoSID entries in Table 2 (e.g., MI R@5 0.0419 vs 0.0417). Clarify whether these are different seeds, different checkpoints, or a reporting inconsistency.","section":"Table 2 vs Table 3"},{"comment":"Related work (§2.1) cites time-interval-aware self-attention (Li et al., 2020) and other temporal sequential models; a short explicit contrast—why gap tokens on SID sequences differ from continuous time embeddings in item-ID Transformers—would sharpen the positioning.","section":"§2.1–2.2"},{"comment":"Efficiency Table 5: training-time ratio on VG is 1.29× while MI is 1.07×; a brief note on sequence-length distribution or batching would help readers interpret the overhead.","section":"Table 5, §5.5"},{"comment":"Typos / polish: “Time-A ware” / “Field-A ware” spacing artifacts in the abstract and §4; “sequence-to sequence” missing hyphen in the abstract; “T emporal” in Figure 1(b) axis label.","section":"Abstract, Figure 1, §4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a clean, well-executed systems paper tightly built on ReSID (including shared co-author ecosystem and arXiv lineage). Novelty is real but incremental; the missing dummy/random-gap control is the main scientific gap. If the authors add that control and tighten statistical reporting, the paper is suitable for a solid IR/recsys venue. Scope fit is good for cs.IR; not a stretch for a top conference track either after revision."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know is that ChronoSID is a practical, low-overhead way to put inter-interaction time into the SID generative stack without breaking the compact three-code decoder. They keep ReSID’s GAOQ quantization frozen, add TA-FAMAE (masked feature recon plus log-gap regression on the target position), then interleave five fixed log-scale gap tokens (plus start) with the SID tuples on the T5 encoder only. Decoder target stays pure SID generation. That design choice is clean.\n\nWhat is actually new is the combination inside the generative SID pipeline, not time modeling itself (TiSASRec et al. already exist). The empirics are the strong part: eight Amazon-2023 domains, direct ReSID comparison, ablations that isolate gap tokens as the main driver, long-gap diagnostics that match the motivation, sensitivity on λ and G, popularity splits, and even output-level SID accuracy with bootstrap CIs. Efficiency numbers show only moderate encoder-length cost. Citation pattern is appropriate; they correctly position against TIGER/LETTER/ReSID and the older time-aware sequential work.\n\nSoft spots are real but proportionate. Absolute lifts are small (e.g., MI R@5 0.0388→0.0417). Main table lacks per-run error bars even though they average five seeds. No code or data release. The stress-test concern lands: lengthening every history 3L→4L and injecting a new token type could buy capacity or separator effects; they never replace real Disc(Δt) with constant, shuffled, or non-temporal fillers. Sensitivity only varies bin count inside the same log family. So the attribution to “temporal awareness” is plausible (especially the long-gap pattern) but not fully isolated. Fixed bins {1h,1d,1w,1mo} are a reasonable default, not a universal claim.\n\nThis is for people already building or evaluating SID generative recommenders. The central argument holds up under the evidence they present. It deserves a serious referee; I would not desk-reject. If you work in this subfield, read the ablations and Figure 3; otherwise it is skippable. I would cite it when discussing temporal inputs for generative SID models.","headline":"Clean, controlled extension of ReSID that adds gap tokens and a light temporal aux loss; modest consistent gains, especially on long gaps, but missing a dummy/random-gap control for capacity confounds.","tokens_in":22287,"tokens_out":567,"would_cite":true,"duration_ms":10683,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Inter-interaction time gaps, tokenized and interleaved with semantic item codes, fix temporal blindness in generative recommenders and lift next-item accuracy.","keywords":["Recommender Systems","Generative Recommendation","Semantic IDs","Temporal Modeling","Sequential Recommendation","Gap Tokens","Time-Aware Representation"],"falsifier":"Re-run the same leave-one-out Amazon splits with the identical ReSID quantization and T5 backbone, but replace the fixed log bins by random or equal-frequency bins of the same cardinality; if ChronoSID’s lift over ReSID disappears or reverses, the central claim that the temporal signal itself is responsible fails.","tokens_in":22189,"feed_emoji":"⏱️","tokens_out":860,"duration_ms":10503,"temperature":0.7,"pith_summary":"Semantic-ID generative recommenders turn next-item prediction into short code generation, but they usually feed the model only a static sequence of item codes and ignore how much time passed between clicks. The paper argues that those elapsed gaps carry real signal about whether a user is still in the same intent or has drifted. ChronoSID keeps the existing three-stage pipeline (representation, quantization, generation) and adds two light temporal pieces: an auxiliary time-gap prediction loss while learning item embeddings, and fixed log-scale gap tokens interleaved with the semantic IDs that the generator sees. On eight Amazon review domains this consistently beats the strongest same-pipeline baseline and other generative methods, with the largest lifts when the wait until the next interaction is long.","feed_headline":"Time-gap tokens lift generative recommenders on Amazon data","feed_subtitle":"Interleaving log-scale wait tokens with semantic item codes beats static SID baselines, especially after long absences","key_machinery":"ChronoSID’s gap-token interleaving: each historical interaction is preceded by a discrete log-scale gap token (<1 h, 1 h–1 d, 1 d–1 w, 1 w–1 mo, ≥1 mo, plus a start token) so the T5 encoder input becomes gap-SID-SID-SID repeated, giving the generator explicit temporal context without changing the decoder target or the quantization stage.","core_discovery":"Static semantic-ID sequences are temporally incomplete; injecting discretized historical inter-interaction gaps as encoder tokens, plus a light time-gap auxiliary loss on item representations, improves generative next-item recommendation while preserving compact SID generation, and the gains are clearest under long-gap conditions where preference drift is more likely.","pith_inferences":["If fixed bins already help, learned or adaptive temporal vocabularies could capture domain-specific rhythms (seasonality, shopping cycles) without enlarging the decoder.","The same gap tokens could condition not only next-item codes but also next-time prediction, turning the generator into a joint item-and-timing model.","Long-gap robustness may matter most in sparse catalogs or cold-start return visits, suggesting a natural stress test beyond Amazon leave-one-out."],"forward_implications":["SID-based generative recommenders that ignore elapsed time systematically under-perform when users return after long absences.","Gap-token injection alone accounts for most of the lift; the representation-level time-gap loss is a smaller complementary regularizer.","Encoder length grows only from 3L to 4L tokens while the decoder target stays a three-code SID, so the accuracy gain comes at moderate extra cost.","The same interleaving pattern can be dropped onto other SID pipelines that keep frozen item codes and a sequence-to-sequence generator."],"fun_headline_variants":["ChronoSID injects log-scale gap tokens into SID generative recs","Time-gap tokens lift generative recommenders especially on long absences","Discretized inter-interaction gaps improve semantic-ID next-item generation","TA-FAMAE plus gap tokens cure temporal blindness in SID recommenders","Interleaving wait tokens with SIDs sharpens generative recs under drift"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"A single hand-chosen set of five fixed log-scale time bins works well enough across product domains that the measured gains can be credited to temporal awareness rather than to a lucky binning.","fun_headline_variants_meta":{"raw":{"variants":["ChronoSID injects log-scale gap tokens into SID generative recs","Time-gap tokens lift generative recommenders especially on long absences","Discretized inter-interaction gaps improve semantic-ID next-item generation","TA-FAMAE plus gap tokens cure temporal blindness in SID recommenders","Interleaving wait tokens with SIDs sharpens generative recs under drift"]},"model":"grok-4.5","effort":"low","cost_usd":0.003432,"raw_usage":{"total_tokens":1166,"prompt_tokens":795,"num_sources_used":0,"completion_tokens":100,"cost_in_usd_ticks":34320000,"prompt_tokens_details":{"text_tokens":795,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":271,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":795,"tokens_out":100,"duration_ms":3025,"temperature":1.0,"reasoning_tokens":271,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T23:01:20.275173+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the same leave-one-out Amazon splits with the identical ReSID quantization and T5 backbone, but replace the fixed log bins by random or equal-frequency bins of the same cardinality; if ChronoSID’s lift over ReSID disappears or reverses, the central claim that the temporal signal itself is responsible fails.","supporting_citations":[],"review_version":1}