{"id":"171070fb-939e-4173-b6ab-f4d425c7bf34","arxiv_id":"2512.24787","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A hierarchical generative slate recommender that plans list-level preferences before decoding items achieves 5× speedup and gains in deployed A/B tests.","lead":"Tencent researchers built HiGR, a generative recommender that generates an entire video slate by first planning the list's theme and then choosing items, and they report it runs five times faster than prior generative recommenders while lifting watch time and video plays in live traffic. The significance is that generative recommendation, which usually suffers from slow token-by-token decoding, can now operate at industrial scale.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The >10% offline gain is not robustly established: it relies on a filtered proprietary test set, shrinks to 3–7% on public KuaiRec for key metrics, and the ORPO loss in Eq. (12) is written inconsistently, leaving the headline claim unverifiable without code/data and a corrected loss.","rationale":"The reader's verdict of CONDITIONAL is appropriate. My concern partly overlaps with the reader's rationale (proprietary filtered dataset, no code/error bars, non-standard ORPO) but differs from the stated weakest_assumption about CRQ-VAE prefix semantics. I think the semantic-consistency issue is a mechanism-level gap: even if prefixes are not perfectly semantic anchors, the model could still perform well by learning arbitrary preference embeddings. The evaluation and loss-formulation issues more directly jeopardize the central claim, because they determine whether the reported >10% offline gain and the ORPO contribution are credible and reproducible. I therefore partially agree with the reader's weakest_assumption but identify the evaluation/reproducibility concern as more load-bearing. The concern does not warrant rejection: it is addressable by releasing artifacts, unfiltered evaluation, and correcting Eq. (12). Thus the verdict remains CONDITIONAL (UNCHANGED).","tokens_in":18562,"tokens_out":10659,"duration_ms":101779,"concrete_test":"Release the code and, if possible, a random unfiltered evaluation sample; rerun the Table 1 comparison (HiGR-25M vs OneRec-25M) on KuaiRec and on the unfiltered Tencent logs with at least 5 seeds, reporting mean and 95% CI for NDCG@5/hit@5/recall@5. If the relative improvement over OneRec on the primary metric is below 10% on the unfiltered set, or the CI covers zero, the 'over 10%' claim should be revised. Separately, with released code, check that the implemented ORPO loss matches a corrected Eq. (12) using full-slate z_theta, not per-token z_theta.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the >10% offline improvement over SOTA. The evidence for it is weaker than the headline suggests. Appendix A states that all training data are filtered high-quality slate-level samples, where individual item durations and cumulative slate duration exceed thresholds; the offline test set similarly uses only slates with positive feedback. Such filtering can inflate both absolute and relative metrics if it correlates with model behavior. On the public KuaiRec benchmark (Table 1), HiGR-25M vs OneRec-25M improves NDCG@5 by only 5.2% (0.1574 vs 0.1496) and Impressions hit@5 by 2.7%; the >10% margin appears only on selected metrics of the proprietary dataset. No error bars, significance tests, or code are provided. A second, independent issue is Eq. (12): z_theta is defined as a sum over all tokens of a slate, but the equation writes z_theta(x, y+_t) and z_theta(x, y-_t), which is dimensionally inconsistent; the correct ORPO objective requires full-slate log-odds z_theta(x,y+) and z_theta(x,y-). As written, the post-training loss is not reproducible, and the reported ORPO gains cannot be traced to a well-defined objective. These are addressable reporting issues, not refutations, but they are exactly the load-bearing support for the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"HiGR proposes a two-stage generative slate recommendation framework for industrial deployment. It first tokenizes items into hierarchical semantic IDs with a Contrastive RQ-VAE (CRQ-VAE), then decouples generation into a coarse-grained slate planner and a shared fine-grained item generator (HSD), and finally applies an ORPO-style listwise preference alignment with three negative-slate objectives. The paper reports offline gains over state-of-the-art baselines of more than 10% on a Tencent industrial dataset, a 5× inference speedup over OneRec, and online A/B improvements of 1.22% in average watch time and 1.73% in average video plays. The same architecture is also evaluated on KuaiRec, where gains over OneRec-25M are smaller (roughly 2.7-6.5% on three of the five reported metrics).","tokens_in":1544,"tokens_out":2185,"duration_ms":56944,"significance":"The framework is a credible industrial contribution: the hierarchical decomposition is well motivated, the online A/B evidence is valuable, and the reported deployment at Tencent scale is a real strength. If the empirical claims withstand scrutiny, the paper demonstrates that coarse-to-fine generative slate planning can be both effective and efficient enough for production. However, the headline quantitative claim is not yet robustly established. The >10% offline margin appears mainly on the proprietary, filtered industrial test set, while the public KuaiRec results are substantially smaller and are reported without error bars or significance tests. In addition, the ORPO loss in Equation (12) is written inconsistently, so the reported post-training gains cannot currently be traced to a reproducible objective. These are addressable issues, but they are load-bearing for the paper's central claim.","major_comments":[{"comment":"The claim of over 10% offline improvement is not robustly supported. On KuaiRec, HiGR-25M improves over OneRec-25M by only about 2.7% on Impression hit@5, 3.6% on Effective View hit@5, and 5.2% on NDCG@5; the >10% margin appears on the proprietary industrial metrics. Appendix A also states that all training data consists of filtered high-quality slate-level samples with duration thresholds, and Section 4.1.2 describes a test set constructed from slates with positive feedback. Such filtering can inflate both absolute and relative metrics. The authors should state explicitly whether the same filtering is applied to the test set, report all metrics with standard deviations or confidence intervals, and perform significance tests on both datasets. Without this, the headline margin is not verifiable.","section":"4.2, Table 1, Appendix A"},{"comment":"The ORPO post-training objective is dimensionally inconsistent. Eq. (11) defines a token-level log-odds, and the text defines z as a sum over all tokens of a slate, i.e., a slate-level scalar. Eq. (12) then writes z_theta(x, y+_t) and z_theta(x, y-_t), which appear to be token-level quantities, and the first term uses y+_i<t without defining i. As written, the loss cannot be implemented or reproduced. The authors need to rewrite Eq. (12) with consistent notation, specifying whether the ORPO odds ratio is computed over the whole slate or over individual tokens, and how the supervised term is masked. Without a corrected objective, the ORPO gains in Table 4 are not attributable to a well-defined method.","section":"3.5.2, Eq. (12)"},{"comment":"The design assumes that the sum of SID embeddings is a reliable preference embedding for list-level planning. This assumption is central to HSD: the planner conditions on those sums, and errors in the planner propagate to the item generator. Table 2 reports internal consistency and concentration metrics for CRQ-VAE, but these are self-defined and are not validated against an external semantic benchmark such as item-item retrieval, taxonomy alignment, or human evaluation. The authors should provide a more direct test of the assumption, such as ablating CRQ-VAE against RQ-VAE within the full HiGR pipeline, or showing that the prefix space supports semantic interpolation or clustering that matches an external label set. Without such evidence, the hierarchical planning step is only weakly grounded.","section":"3.4.1, 4.3.1"}],"minor_comments":[{"comment":"No error bars, standard deviations, or significance tests are reported for any offline result, including the KuaiRec comparisons. This should be added at least for the main comparison table.","section":"4.1.2"},{"comment":"In Eq. (10), the generated SID is denoted s-hat^{d+1}_m while the previous text indexes the output SID sequence as s-hat^1_m, ..., s-hat^D_m. This off-by-one notation should be fixed.","section":"3.4"},{"comment":"The metrics Collision, Concentration, and Consistency are used to compare tokenization methods, but their formal definitions are not given in the main text or appendix. Please define them precisely.","section":"Table 2"},{"comment":"The manuscript text after Figure 4 contains a large run of garbled '/uni...' glyph sequences. This appears to be a rendering artifact and should be removed before submission.","section":"Appendix C / Figure 4"},{"comment":"The online A/B test compares HiGR against the incumbent multi-stage system, but the description does not specify the duration of the test, the number of users in each arm, or whether the differences are statistically significant. Adding these details would strengthen the online claim.","section":"4.6"}],"recommendation":"major_revision","confidential_remarks":"The paper has real industrial value and the online deployment is a strong point. My main concern is that the central offline claim is supported only by a filtered proprietary benchmark and an inconsistently written loss. I would be willing to accept after the authors fix Eq. (12), report error bars and significance on KuaiRec, clarify the test-set filtering, and add a direct validation of the preference-embedding assumption. The garbled text in Appendix C should also be cleaned, though that is purely editorial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi — quick take on the HiGR paper. It's a legitimate attempt at an industrial-scale generative slate recommender, and the core architecture is genuinely novel: a residual-quantized VAE with prefix-level contrastive constraints, a two-stage coarse-to-fine decoder, and ORPO-based listwise alignment. The efficiency story holds together — decoding M items in two stages with a shared item generator cuts the sequential steps, and the reported 5× speedup is consistent with the complexity analysis. The online A/B gains (1.22% watch time, 1.73% video plays) are plausible for a deployed system of this sort.\n\nThe soft spots are real, though. The headline 'over 10% offline improvement' comes only from the proprietary Tencent dataset, which is filtered to high-quality slates with positive feedback. On the public KuaiRec benchmark, the gains over OneRec are more like 3–7% depending on the metric — still positive, but a different story. No error bars, no significance tests, no code or data ('pending internal approval'), so an independent group can't verify the central claim.\n\nEq. (12) is also miswritten: z_θ is defined as a sum over all tokens in the slate, but the ORPO loss uses z_θ(x,y+_t) and z_θ(x,y-_t), mixing the per-token and full-sequence levels. It should be z_θ(x,y+) and z_θ(x,y−). That's fixable, but as it stands the post-training objective isn't reproducible from the text. The consistency metric in Table 2 is never defined either, which weakens the argument that the prefixes are semantically anchored.\n\nNone of this is fatal. The architecture is coherent, the ablations make sense (context embedding and ORPO both help), and the scaling-law figure is a nice touch. The paper would benefit from a corrected Eq. (12), a clear definition of Consistency, and at least one run with error bars on KuaiRec. If the authors can provide those, the core claims are likely to survive.\n\nI'd send this to peer review — it's the kind of industrial deployment paper the community needs to see — but I'd ask for the missing artifacts and a more careful write-up. The reader's CONDITIONAL verdict is about right.","headline":"A novel and coherent industrial generative slate recommender with a real speedup, but the headline offline margin rests on a filtered proprietary test set and Eq. (12) needs a fix.","tokens_in":19515,"tokens_out":4890,"would_cite":true,"duration_ms":46912,"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":"HiGR shows that slate recommendation can be treated as a coarse-to-fine generation problem: first plan the list's preferences, then decode individual items.","keywords":["Slate recommendation","Generative recommendation","Semantic IDs","Residual quantized VAE","Hierarchical decoding","Preference alignment","ORPO","Industrial recommender systems"],"falsifier":"A direct test of prefix semantic consistency: take a held-out set of items, group them by their first-level code, and ask human raters or an external semantic-similarity benchmark whether items sharing a prefix are more similar than items with different prefixes. A null result would mean the planner's preference embeddings are not semantically grounded. A second test would be to ablate the contrastive loss and show that planner quality and prefix consistency do not degrade, which would contradict the paper's mechanism.","tokens_in":18446,"feed_emoji":"⚡","tokens_out":4057,"duration_ms":44014,"temperature":0.7,"pith_summary":"HiGR argues that generative slate recommendation can be made fast and accurate enough for industrial use by structuring the item token space and splitting generation into two stages. Its central claim is that a contrastively trained residual quantizer produces semantic IDs whose prefixes are reliable meaning anchors, and that a two-stage decoder can then plan the whole slate at the level of preference embeddings before decoding each item. If this holds, generative recommenders need not trade speed for quality: a 10-item slate can be produced in roughly 10 planning steps plus fast parallel item decodes instead of 30 sequential token steps. The paper reports offline gains of over 10% against state-of-the-art baselines, a 5x inference speedup, and online A/B improvements of 1.22% in watch time and 1.73% in video plays on a large commercial platform.","feed_headline":"Hierarchical planning lifts slate accuracy 10% at 5x speed","feed_subtitle":"Coarse-to-fine decoding over structured semantic IDs lifts watch time 1.22% and plays 1.73% in live A/B tests.","key_machinery":"The central objects are (a) the Contrastive Residual Quantized VAE (CRQ-VAE), which adds temperature-scaled contrastive losses to the first D-1 layers of a residual quantizer so that similar items cluster under shared prefixes, plus a global quantization loss that prevents residual vanishing; (b) the preference embedding, defined as the sum of an item's SID embeddings, which serves as the unit of planning; (c) the Hierarchical Slate Decoder with a deep coarse-grained planner and a shallow, parameter-shared fine-grained item generator, using cross-attention to a user-context embedding; and (d) the Greedy-Slate Beam-Item inference strategy, which decodes items independently given planner-produ","core_discovery":"The paper introduces HiGR, a hierarchical generative slate recommendation framework. Its first component, CRQ-VAE, applies prefix-level contrastive constraints to residual quantization so that an item's high-level ID codes (prefixes) encode shared semantics, while the final code retains item-level discrimination. Its second component, the Hierarchical Slate Decoder, decouples generation into a coarse-grained slate planner that autoregressively predicts preference embeddings (sums of SID embeddings) for each slot, and a shared fine-grained item generator that independently decodes each predicted preference into an item's SID sequence. Its third component uses ORPO, a reference-model-free pref","pith_inferences":["An untested consequence is that the preference-embedding space, being decoupled from individual item IDs, may be transferable across domains or time periods; a slate planner trained in one content domain could potentially plan in another with only the item generator retrained.","The prefix-contrastive constraint is not specific to recommendation: any discrete tokenization that needs hierarchical semantic structure (e.g., code generation, molecular design, or hierarchical text representation) could benefit from the same contrastive prefix alignment.","If prefixes are truly stable semantic anchors, one could precompute a small set of 'preference archetypes' and constrain the planner to emit only those embeddings, enabling explicit user control or explainable slate composition without retraining.","The efficiency gain opens a new trade-off: the saved latency budget could be reinvested into larger beam widths or a deeper planner on the same hardware, potentially pushing quality further within fixed response-time limits."],"forward_implications":["If the design works as reported, generative slate recommenders can satisfy sub-100ms latency budgets: the paper reports a 5x inference speedup over the strongest autoregressive baseline.","Structured, prefix-aligned semantic IDs let the system impose diversity and relevance constraints directly on ID prefixes during decoding, rather than measuring them after the fact.","Listwise ORPO alignment with three complementary objectives (ranking fidelity, genuine interest, diversity) improves all offline metrics over no alignment and over DPO/SimPO alternatives.","The framework exhibits a power-law scaling relationship between model size and recommendation quality, suggesting that predictable gains come from simply increasing capacity.","Online A/B tests on a commercial platform serving hundreds of millions of users show a 1.22% increase in average watch time and a 1.73% increase in average video plays, indicating the approach translates to live business metrics."],"fun_headline_variants":["Hierarchical slate rec: 5x faster, 10% better, live on Tencent","Coarse-to-fine decoding lifts slate quality 10%, speeds 5x","HiGR: 5x speedup and 10% accuracy gain in slate recommendation","Tencent's HiGR: hierarchical planning boosts slate rec 10% at 5x speed","Slate rec via hierarchical SIDs: 5x faster, 10% better watch time"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The approach assumes that the learned ID prefixes are genuinely semantic anchors, so that the sum of the code embeddings acts as a reliable 'preference embedding' for the planner; if the prefixes are not truly consistent, the two-stage design will propagate and amplify errors.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical slate rec: 5x faster, 10% better, live on Tencent","Coarse-to-fine decoding lifts slate quality 10%, speeds 5x","HiGR: 5x speedup and 10% accuracy gain in slate recommendation","Tencent's HiGR: hierarchical planning boosts slate rec 10% at 5x speed","Slate rec via hierarchical SIDs: 5x faster, 10% better watch time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1445,"prompt_tokens":840,"completion_tokens":605,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":499}},"tokens_in":584,"tokens_out":605,"duration_ms":5679,"temperature":1.0,"reasoning_tokens":499,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T13:14:59.009532+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test of prefix semantic consistency: take a held-out set of items, group them by their first-level code, and ask human raters or an external semantic-similarity benchmark whether items sharing a prefix are more similar than items with different prefixes. A null result would mean the planner's preference embeddings are not semantically grounded. A second test would be to ablate the contrastive loss and show that planner quality and prefix consistency do not degrade, which would contradict the paper's mechanism.","supporting_citations":[],"review_version":1}