{"id":"5f9703d4-e905-426a-884d-a6d0af3151a0","arxiv_id":"2506.05781","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RPG replaces autoregressive semantic ID generation with parallel multi-token prediction plus graph-constrained decoding, improving NDCG@10 by about 12.6% over generative baselines while keeping inference cost independent of catalog size.","lead":"RPG is a recommendation method that predicts all digits of a long semantic item code in parallel, instead of generating them one by one. It reports better ranking accuracy and much lower inference cost than autoregressive semantic ID recommenders on four Amazon datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Eq. (1) conditional-independence factorization is asserted, not tested; any inter-token dependencies in user preferences would bias the parallel scoring and the graph-decoding similarity premise.","rationale":"The reader and I converge on the same load-bearing point: the conditional-independence factorization in Eq. (1). I have looked for other candidates—the absence of error bars, the encoder mismatch between RPG (text-embedding-3-large) and TIGER (sentence-t5-base) in the main table, the lack of an approximation guarantee for graph decoding—but each of those is either disclosed in the text (encoder choice is analyzed in Sec. 3.5.2) or is a known property of heuristic search (graph decoding is explicitly an ANN-style method, Sec. 2.4.1). The factorization, by contrast, is asserted in one sentence (Sec. 2.2.1) and never tested. It is the foundation for both the training objective and the 'similar IDs have similar logits' observation that justifies graph-constrained decoding. If real user preferences induce correlations across OPQ subvectors, the product-of-marginals score is misspecified and could rank items incorrectly in exactly the long-ID regime the paper advocates. The paper's Figure 2 provides only a within-model sanity check, not evidence of fidelity to the true conditional distribution. I therefore recommend no change to the reader's CONDITIONAL verdict; the paper should be asked to either test the independence assumption directly (e.g., via the synthetic experiment above) or add a reranker/energy term that captures dependencies, and report the resulting NDCG@10. The method is otherwise coherent and the efficiency claims are well-supported by the complexity analysis and Figure 3.","tokens_in":18489,"tokens_out":10988,"duration_ms":115703,"concrete_test":"Run a controlled synthetic experiment in which items are generated from latent classes with tunable inter-subvector correlations: each class has a mean vector; sample item subvectors as class mean plus noise with a covariance that can be off-diagonal. Generate user sequences from a Markov chain over latent classes. Train RPG and an oracle autoregressive model (e.g., TIGER with long IDs) on the same data. If RPG's NDCG@10 drops relative to the oracle as the inter-subvector correlation increases, the Eq. (1) independence assumption is a real limitation; if RPG tracks the oracle, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise of RPG is that the m tokens of the next item's semantic ID are conditionally independent given the user sequence, so the model factorizes P(c_1,...,c_m | s) as the product of per-codebook marginals (Eq. 1) and scores a candidate item by summing the independent token logits (Eq. 3). This assumption is not a benign modeling convenience: item features encoded by different OPQ subvectors are selected by the same user preference, so the true conditional distribution generally has positive correlations (e.g., a user who tends to click running sneakers should make 'sport' and 'shoe-type' tokens co-occur). A product distribution matches only the marginals, and ranking by the sum of log-marginals is a Naive-Bayes score with no approximation guarantee. The paper offers no direct validation. Figure 2 is sometimes cited as support, but it measures the factorized model's own score differences; under Eq. (3) it is nearly tautological that IDs differing in more digits have larger score gaps. It says nothing about whether the factorized score approximates the true conditional item probability. The graph-constrained decoding (Sec. 2.3) inherits this risk: it assumes 'similar IDs have similar logits' and builds edges from token-embedding similarity, then propagates a beam. If the true score landscape has sharp, non-smooth dependencies, the beam can miss the genuine top-K items. The paper also never reports the NDCG@10 that would be obtained by exhaustive scoring of all items under Eq. (3), so the loss incurred by the graph decoder is unknown. Because the 12.6% improvement claim depends on both the factorized score and the graph search being faithful, the unvalidated independence assumption is the most load-bearing weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RPG, a semantic ID-based sequential recommendation framework. Items are tokenized into long, unordered semantic IDs using optimized product quantization (up to 64 tokens), and a Transformer decoder is trained with a multi-token prediction (MTP) objective that factorizes the probability of the next item's ID tokens as a product of per-codebook marginals. At inference, RPG computes per-token logits in parallel and decodes a top-K list using graph-constrained beam propagation over a graph that connects semantically similar IDs. Experiments on four Amazon categories compare RPG with item-ID and semantic-ID baselines, report ablations, analyze scaling with ID length, and measure inference memory and time as the item pool grows.","tokens_in":18796,"tokens_out":4671,"duration_ms":47146,"significance":"If the central claims hold, RPG is a useful contribution to generative semantic-ID recommendation: it replaces autoregressive token-by-token beam search with a single encoder pass and parallel token scoring, enabling much longer IDs (16-64 tokens) than the 4-token IDs used by TIGER. The paper includes informative ablations (random vs. OPQ IDs, RQ IDs, projection-head variants, with/without graph constraints), an efficiency comparison with dummy items, and a cold-start analysis, and the code is released. The main caveats are that the conditional-independence factorization in Eq. (1) is asserted rather than validated, that the abstract overstates the role of m=64, and that the graph decoder's approximation quality relative to exhaustive scoring is never quantified.","major_comments":[{"comment":"The factorization P(c_{t,1},...,c_{t,m}|s) = \\prod_j P^{(j)}(c_{t,j}|s) is the load-bearing modeling assumption for the parallel design, but the paper does not validate it. Figure 2 only shows that the factorized score from Eq. (3) changes monotonically with the number of differing digits; since that score is a sum of per-token marginal logits, the figure is nearly a restatement of the score definition and does not establish that the product approximates the true conditional distribution over token combinations. If user preferences induce correlations across OPQ subvector tokens (e.g., a user who prefers running shoes should jointly elevate both the 'sport' and 'shoe-type' tokens), the MTP model is optimizing a misspecified objective and rankings from Eq. (3) can be biased. Please provide a direct validation, for example by comparing Eq. (3) scores with an autoregressive or jointly trained model on the same OPQ IDs, or by measuring inter-token dependencies in the learned conditional distribution.","section":"Sec. 2.2.1, Eq. (1)"},{"comment":"The abstract states that scaling semantic ID length to 64 enables RPG to outperform generative baselines by an average of 12.6% on NDCG@10, but Table 6 reports optimal semantic ID lengths of 16, 32, 16, and 64 for Sports, Beauty, Toys, and CDs, respectively, and Figure 4 shows performance saturating or declining for the smaller datasets. The reported 12.6% average gain is therefore obtained under per-dataset optimal lengths, not under m=64 uniformly. Please rephrase the abstract and Section 3.2 to describe the per-dataset optimal lengths and the average gain under those settings, and soften the claim that m=64 is the source of the improvement for all datasets.","section":"Abstract and Sec. 3.5.1 (Table 6, Fig. 4)"},{"comment":"The paper claims that graph-constrained decoding has time and memory complexity independent of catalog size, but Table 6 shows that RPG visits 10.90%, 24.79%, 20.97%, and 23.28% of the item pool on the four datasets under the best hyperparameters. Because the decoder is approximate, the paper should report the NDCG@10 that would be obtained by exhaustive scoring of all candidate items using Eq. (3). At the catalog sizes used here exhaustive scoring is feasible and would give an upper bound on the model's ranking quality; without it, the reader cannot tell how much of RPG's effectiveness comes from the learned scores versus from the graph-propagation heuristic, and the efficiency/effectiveness trade-off cannot be properly assessed.","section":"Sec. 2.3.3 and Table 6"},{"comment":"The inference-efficiency comparison with TIGER is informative, but the headline memory/time savings depend on treating graph construction as outside the online inference budget. Table 8 shows that RPG has total storage O(N(m+k)) and that inference fetches only O(bqkm) storage, but the one-time cost of building the decoding graph from pairwise token-embedding similarities is neither reported nor discussed in the complexity analysis. Please state explicitly whether graph construction is included in the experimental memory/time measurements, and quantify its one-time cost, so that the 'independent of the number of items' claim is scoped to online inference rather than the full system.","section":"Sec. 3.3 and Table 8"}],"minor_comments":[{"comment":"The evaluation metrics paragraph contains a typo: 'NDCD@K' should be 'NDCG@K'.","section":"Sec. 3.1"},{"comment":"Equation (1) defines L as a loss, but Section 2.2.2 says 'we compute L in Equation (1) as the logits.' Using the same symbol for the loss and the negative log-likelihood score is confusing; please introduce a separate notation for the inference score.","section":"Sec. 2.2.1 and 2.2.2"},{"comment":"The row labeled 'sentence-t5-base 16 ~ 64' is ambiguous: it is unclear whether this denotes the best length in a range, a single tuned length, or aggregation over several lengths. Please specify the exact setting.","section":"Table 4"},{"comment":"The y-axis label '|logits|' should be expanded to 'absolute difference in factorized sequence scores' and the plot should state which dataset, model checkpoint, and evaluation split are used.","section":"Figure 2"},{"comment":"The statement that RPG reduces sequence encoder forward passes 'from approximately O(bm) to just O(1)' should clarify that the O(1) refers to encoder forward passes only, not to the total decoding operations, which include graph propagation steps of cost O(bqkm).","section":"Sec. 2.4.2"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Eq. (1) is legitimate and should be addressed head-on. I would like to see an exhaustive-scoring upper bound and a corrected abstract before considering the paper for acceptance; the core idea is promising and the experimental setup is generally sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Genuinely new, and likely worth citing: RPG predicts all tokens of a long product-quantized semantic ID in parallel, trains with a multi-token-prediction loss, and decodes over a graph of similar IDs so per-user cost does not scale with catalog size. No one has put that combination together before, and the experimental story is coherent. The catch is that the abstract's \"scaling up to 64\" framing oversells things: optimal ID lengths are 16, 32, 16, and 64 on Sports, Beauty, Toys, and CDs, so only CDs actually wants 64 tokens.\n\nWhat the paper does well: the ablations are the best part. Random tokens kill performance, RQ tokens hurt relative to OPQ, beam search produces exactly 0.0000 NDCG@10 on all four datasets (as it should, since unordered PQ IDs have no sequential structure), and removing the graph drops NDCG@10 by more than half everywhere. The dummy-item efficiency experiment cleanly shows inference memory and time staying flat as the pool grows, and Table 8 is honest that total storage still grows with N—only per-user access does not.\n\nSoft spots, in rough order. First, the conditional-independence assumption in Eq. (1) is asserted, not validated, and Figure 2 is near-tautological under the factorized score in Eq. (3): sums over more differing terms differ more, by construction. The paper never reports the NDCG@10 from exhaustively scoring all items under Eq. (3), so the cost of the factorization and the cost of the graph search are never separated. That one missing table is what I would ask a revision for. I do not think this assumption is disqualifying—additive scoring is standard in this field, and RPG beats VQ-Rec, which uses the same OPQ tokens but retrieves by vector distance, so the MTP-trained score is carrying real weight. Second, the Appendix C experiment comparing TIGER at longer ID lengths is confounded: TIGER uses RQ-VAE, and extra residual levels degrade RQ code quality independent of inference cost, so that comparison does not isolate the efficiency argument. Third, the abstract's m=64 framing should be corrected to match Table 6.\n\nAnyone working on semantic-ID or generative recommenders should read this; I would take it to a reading group. The code is public, so verification is cheap. My referee call: conditional accept, with the exhaustive-scoring number and a re-framed abstract as the asks.","headline":"RPG is genuinely new—parallel generation of long OPQ semantic IDs with a multi-token-prediction loss plus graph-constrained decoding—and the experiments mostly hold up, but the abstract oversells m=64 and the paper never quantifies the cost of its factorized scoring.","tokens_in":19424,"tokens_out":10269,"would_cite":true,"duration_ms":97315,"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":"By predicting every token of a long semantic ID in parallel, a recommender beats autoregressive generative baselines by 12.6% on NDCG@10 while keeping inference cost independent of catalog size.","keywords":["sequential recommendation","semantic IDs","product quantization","multi-token prediction","parallel decoding","graph-constrained decoding","generative recommendation","inference efficiency"],"falsifier":"Run a held-out test in which two code positions are deliberately made informative only through their co-occurrence, such as a brand subvector and a category subvector that jointly identify a niche product class, and compare the factorized scorer against a scorer that uses the full token combination. If the joint scorer improves NDCG@10, or if brute-force scoring of every valid ID reorders RPG's graph-decoded top-$K$, the central independence and parallel-decoding claim is refuted.","tokens_in":18285,"feed_emoji":"⚡","tokens_out":9915,"duration_ms":103614,"temperature":0.7,"pith_summary":"Generative recommenders that tokenize items into semantic IDs have been limited to very short IDs (typically four tokens) because each token is produced by an autoregressive beam search, whose cost grows with ID length. This paper argues that the length limit is self-imposed: if the ID is built with product quantization so its tokens live in separate codebooks, the model can be trained to predict all tokens at once with a multi-token prediction loss, and a candidate item's score becomes the sum of independent per-token log probabilities. Decoding then needs only one encoder forward pass, followed by a fixed number of graph-propagation steps over a prebuilt graph of similar valid IDs. The paper reports that scaling semantic IDs to 64 tokens lets this method outperform the best generative baseline by an average of 12.6% on NDCG@10 across four datasets, while using roughly 25x less inference memory and running about 15x faster. If the result holds, expressive long IDs and fast inference are not in conflict for semantic-ID recommenders.","feed_headline":"Semantic IDs of length 64 beat generative recommenders by 12.6%","feed_subtitle":"Predicting all ID tokens in one pass means 15x faster inference, 25x lower memory, and better ranking.","key_machinery":"The load-bearing object is the factorized logit sum: for a candidate ID $(c_1,\\ldots,c_m)$, the recommendation score is $\\sum_{j=1}^m \\log p^{(j)}_{c_j}$, where $p^{(j)}$ is the cached softmax vector over the $j$-th codebook after one sequence encoder pass. This identity turns ranking into a nearest-neighbor-style search over a product space and motivates the decoding graph, whose nodes are valid IDs and whose edges connect IDs with high summed token-embedding similarity. Iterative graph propagation with a beam of size $b$ and $q$ steps explores only $O(bqk)$ candidate IDs, making inference complexity independent of the item count.","core_discovery":"RPG's core claim is that next-item prediction can be done directly in the space of long, unordered semantic IDs. Each item is encoded by optimized product quantization into $m$ tokens drawn from $m$ separate codebooks; conditional on the user sequence, the model assumes the tokens are independent, so the probability of a candidate ID factors as $P(c_1,\\dots,c_m|s)=\\prod_{j=1}^m P^{(j)}(c_j|s)$. The training loss is the negative sum of per-token log probabilities, and at inference the score of each candidate ID is the sum of cached per-token logits. Because decoding is parallel, ID length can grow to 64 without adding autoregressive steps, and the model avoids invalid token combinations by propagating a small beam through a graph whose edges connect IDs that differ in few tokens. The paper's empirical claim is that this design outranks all compared baselines on 11 of 12 metric-dataset combinations, with an average 12.6% NDCG@10 improvement over the best generative baseline, while keeping inference cost independent of catalog size.","pith_inferences":["A direct test of the independence assumption would be to compare RPG's factorized score with a jointly trained scorer that sees full token combinations; if the joint scorer wins on held-out ranking, some of RPG's signal is being left on the table.","The same parallel-scoring scheme should transfer to other generative retrieval settings where items are represented by product-quantized codes, such as document or product search, provided a valid-ID graph can be built.","Because the graph is static once built, RPG would need an incremental graph update to stay correct in a dynamic catalog where items are added or re-embedded; the paper does not address that serving scenario.","The reported 12.6% gain bundles the effect of the longer ID with the new objective and decoder; a length-controlled comparison that varies only the ID length while holding everything else fixed would separate the two contributions."],"forward_implications":["A single sequence encoder forward pass supplies all token logits, so the number of decoding steps does not grow with semantic ID length; long IDs are therefore no longer a latency burden.","The decoding graph can be precomputed once after training, so serving-time memory and latency stay flat as the catalog grows, unlike retrieval-based recommenders that must scan item embeddings.","The multi-token prediction objective lets the model learn from sub-item semantic structure directly, which is what the paper credits for its gains on cold-start items with few interactions.","Because candidate scoring is parallel, the same architecture can consume richer item encoders and longer IDs without changing the decoder, which the paper shows translates into better NDCG@10 on larger datasets."],"supporting_citations":[{"why":"The autoregressive generative baseline whose 4-token RQ-VAE IDs and beam-search decoding restrict ID length; supplies the benchmark numbers RPG must beat.","marker":"[39]"},{"why":"The retrieval-based semantic ID baseline whose OPQ tokenization and pooled token-embedding item representation RPG borrows.","marker":"[14]"},{"why":"Product quantization and its optimized variant, the techniques that build long unordered semantic IDs with one codebook per digit.","marker":"[8, 19]"},{"why":"Multi-token prediction objective that RPG adapts to predict all tokens of the next semantic ID in parallel.","marker":"[10]"},{"why":"Constrained beam search, the tree-structured constrained decoding idea that RPG extends to a graph over unordered semantic IDs.","marker":"[13]"},{"why":"Analysis cited for why residual-quantized IDs concentrate information unevenly and scale poorly, motivating OPQ for long IDs.","marker":"[57]"},{"why":"The large user-review collection that provides interaction sequences for all benchmark experiments.","marker":"[34]"}],"fun_headline_variants":["Parallel semantic IDs improve ranking by 12.6% with 15x speedup","RPG: 64-token IDs predicted in parallel, +12.6% NDCG@10","Long semantic IDs decoded in parallel, cutting beam search cost","Generate 64 semantic tokens in one pass: 12.6% better NDCG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, once the user's history is fixed, the separate parts of the next item's code do not influence each other, so the model can rank by adding up independent per-part scores; if combinations of code parts carry extra meaning about what a user wants, the ranking objective is misspecified.","fun_headline_variants_meta":{"raw":{"variants":["Parallel semantic IDs improve ranking by 12.6% with 15x speedup","RPG: 64-token IDs predicted in parallel, +12.6% NDCG@10","Long semantic IDs decoded in parallel, cutting beam search cost","Generate 64 semantic tokens in one pass: 12.6% better NDCG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3244,"prompt_tokens":982,"completion_tokens":2262,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":2171}},"tokens_in":598,"tokens_out":2262,"duration_ms":16827,"temperature":1.0,"reasoning_tokens":2171,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:14:02.915569+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a held-out test in which two code positions are deliberately made informative only through their co-occurrence, such as a brand subvector and a category subvector that jointly identify a niche product class, and compare the factorized scorer against a scorer that uses the full token combination. If the joint scorer improves NDCG@10, or if brute-force scoring of every valid ID reorders RPG's graph-decoded top-$K$, the central independence and parallel-decoding claim is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Constrained beam search, the tree-structured constrained decoding idea that RPG extends to a graph over unordered semantic IDs."}],"review_version":1}