{"id":"09f51802-7735-45ba-994a-a9b4a6715d0c","arxiv_id":"2412.12770","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A review that taxonomizes sequential recommendation research by item property construction and surveys recent LLM, multimodal, generative, and ultra-long-sequence methods.","lead":"This paper is a survey of sequential recommendation, organized by how item information is constructed, from pure IDs through side information to multimodal and LLM-based methods. It maps hundreds of recent models into a taxonomy and lists open research directions.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed four-category property-construction taxonomy is not actually the organizing principle of the survey: generative SR, LLM-powered SR, ultra-long SR, and data-augmented SR are classified by different axes, so the central novelty claim is unsupported.","rationale":"The reader's conditional verdict is reasonable, but the identified weakest assumption, the comparability of Table 5, is not the most load-bearing threat to the paper's central claim. The central claim is the property-construction taxonomy as a new perspective and comprehensive roadmap. That claim fails, or at least is substantially overstated, if large parts of the survey are organized by technique, sequence length, or data augmentation rather than by item-property construction. The generative SR section is especially telling: semantic IDs are an item-construction scheme, but they are not one of the four categories defined in Section 3.3, and Eq. (20) is formally a different prediction framework from the four probability expressions given earlier. This is an internal consistency issue, not a matter of disagreeing with community consensus. The Table 5 concern is real and worth fixing, but it affects the empirical observations in Section 7.4 and some conclusions in Section 9; the taxonomy concern affects the paper's stated reason for existing. Since both issues are correctable in revision, the conditional verdict stands unchanged, but the revision conditions should include reconciling the taxonomy with the actual contents of Sections 6.2-6.5, not only re-running or removing the copied experimental results.","tokens_in":42262,"tokens_out":5205,"duration_ms":50818,"concrete_test":"Construct a complete mapping table from every model listed in Figure 5 to one of the four formal categories in Section 3.3. If the models in Sections 6.2-6.5 (e.g., TIGER, SpecGR, P5, Lite-LLM4Rec, CL4SRec, MIMN) cannot be assigned to exactly one of the four property-based categories without either splitting a single model across categories or introducing a separate technique/sequence/data axis, then the four-category taxonomy is not comprehensive. A related check is to redraw Figure 5 with explicit orthogonal axes: if the figure then requires more than one axis to classify the field, the abstract's claim of a single new perspective is overstated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that it studies sequential recommendation from a new perspective, namely the construction of an item's properties, and that it categorizes existing SR models into four categories based on that perspective (Section 3.3, contributions). The four formal categories are: pure ID-based P(i|S_u), ID + general features P(i|S_u,N...), modality-only P(i|F...), and ID + multimodal features P(i|S_u,F...). However, the survey itself does not consistently apply this taxonomy. Section 6.2 introduces generative recommendation, whose item construction is a semantic ID, a distinct fifth construction absent from the four categories; Eq. (20) formalizes it as P(i|S_u)=∏_j p(y_{i,j}|S_u,y_{i,<j}). Section 6.3 groups models by whether LLMs are used for recommendation, embeddings, or data generation, which is a technique axis, not an item-property axis. Sections 6.4 and 6.5 classify by sequence length and data augmentation, which are orthogonal to item properties. The taxonomy in Figure 5 also uses a different top level: 'Pure ID-based SR', 'SR with Side Information', and 'Recent SR Advancements', rather than the promised four property-based categories. Thus the 'new perspective' is applied only to Sections 4, 5, and 6.1, while the rest of the survey uses other organizing principles. The survey may still be useful as a roadmap, but the central claim of a novel, comprehensive property-construction taxonomy is not supported as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript is a survey of sequential recommendation (SR) that proposes to organize the field from the perspective of how an item's properties are constructed. It defines four formal categories (pure ID-based, ID plus general features, modality-only, and ID plus multimodal features) in Section 3.3, and then devotes separate sections to multi-modal SR, generative SR, LLM-powered SR, ultra-long sequence SR, and data-augmented SR. The survey also collects datasets, evaluation protocols, metrics, and a table of experimental results, and ends with future research directions. Its stated contributions are a comprehensive summary, a four-category taxonomy, a summary of the latest techniques, and empirical studies.","tokens_in":42567,"tokens_out":4396,"duration_ms":42497,"significance":"If the property-construction taxonomy were rigorously applied, the survey would offer a genuinely useful organizing scheme for the heterogeneous SR literature, and its coverage of recent LLM-powered and generative recommendation work is timely. The paper is a secondary source with no derivations or fitted models, so its value lies entirely in the accuracy, completeness, and consistency of its synthesis. The descriptive sections are generally consistent with the cited literature and the reference list is broad, which makes the survey a potentially useful entry point. However, the central novelty claim is only partially realized, and the empirical comparisons in Section 7.4 rest on numbers copied from different papers that are not comparable under a common protocol. These issues do not require new experiments to fix, but they do require reworking the taxonomy and the empirical claims, so the current version needs revision before it can be judged as a reliable roadmap.","major_comments":[{"comment":"The claimed four-category taxonomy based on \"construction of item properties\" is not the organizing principle of the survey. Section 3.3 defines the four categories P(i|S_u), P(i|S_u,N), P(i|F), and P(i|S_u,F), and the abstract and contribution list say the survey studies SR from this new perspective. Yet Section 6.2 introduces generative recommendation with semantic IDs, formalized in Eq. (20) as P(i|S_u)=prod_j p(y_{i,j}|S_u,y_{i,<j}); this is a fifth type of item construction that is absent from the four formal categories. Section 6.3 groups models by whether LLMs are used for direct recommendation, semantic embeddings, or data generation, which is a technique axis; Section 6.4 groups by sequence length; and Section 6.5 groups by data augmentation. Figure 5 itself uses a different top-level taxonomy ('Pure ID-based SR', 'SR with Side Information', 'Recent SR Advancements'). Therefore the paper's central claim of a comprehensive property-construction-based categorization is not supported as stated. Please either restructure the survey so that all covered directions are derived from the item-property axis, or narrow the novelty claim to the categories where the axis is actually used.","section":"Section 3.3, Section 3.4, and Figure 5"},{"comment":"The comparative conclusions in Section 7.4 rest on numbers copied from three different papers ([138], [147], [149]), as the text acknowledges. These results were obtained under potentially different preprocessing, candidate sampling, negative sampling, hyperparameter settings, and data splits; the survey also explicitly says the data in each domain is processed according to the corresponding original papers. The bullet claims (e.g., \"LLM-powered SR models show better recommendation performance than traditional SR models\" and \"SR models with side information achieve better performance than pure ID-based SR ones\") treat these non-comparable numbers as though they were produced in a single evaluation. Table 5 also reports no variance or significance measures. As a survey, it is acceptable to quote previously published results, but the authors should either label them strictly as case-by-case reported figures with no cross-paper comparability, or carry out (or cite) a uniform reimplementation. The current wording overstates what the evidence supports.","section":"Section 7.4, Table 5"}],"minor_comments":[{"comment":"The phrase \"entity IDs ... are quietly different\" should read \"quite different\".","section":"Section 3.2"},{"comment":"In Table 5, the row header \"FDSR\" appears inconsistent with the text of Section 7.4, which names the model as \"FDSA\" (reference [121]). Please make the naming consistent.","section":"Section 7.4, Table 5"},{"comment":"In the sentence \"In order words, the representations of item i are obtained...\", the phrase should be \"In other words\".","section":"Section 6.1"},{"comment":"The text says \"Compared with random sampling, this split method is more stable,\" but the method being described is a candidate-sampling procedure, not a data-split method. Please rephrase to avoid the inaccurate use of \"split method.\"","section":"Section 7.2"},{"comment":"The text inside Figure 5 appears to contain several OCR-like typographical errors (e.g., \"Pur e TD-base d SR\", \"Neu r al Neiwo r ks\"). Please provide a clean version of the figure.","section":"Figure 5"},{"comment":"Reference [160] has a typo in its title: \"Leveraing\" should be \"Leveraging\".","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a broad survey that could serve a useful purpose, but its central novelty claim and its headline empirical observations both need substantial reworking. The taxonomy issue is internal to the paper: the survey asserts a four-way item-property partition and then organizes a large fraction of its content by different axes, so the contribution bullet in the abstract is currently misleading. The Section 7.4 comparison is the riskiest part of the paper, since it makes strong comparative claims from numbers that the authors themselves say are copied from different papers. Both issues are fixable within the scope of a survey: the authors can either restructure the text and figure to honor the property-based taxonomy or explicitly reframe the contribution as a roadmap with a three-part organization, and they can reword the empirical bullets as non-comparable reported results or add a uniform evaluation. I do not see evidence of fabrication or circularity, but the current version would not give readers a reliable basis for the comparative conclusions the abstract and Section 7.4 advertise."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent, useful survey of sequential recommendation up to early 2025, but its headline novelty — a four-way taxonomy based on item-property construction — only covers the first half of the paper. Sections 6.2–6.5 group work by generative paradigm, LLM usage, sequence length, and data augmentation, which are different axes. Section 7.4 also overreaches by comparing copied numbers from three different papers as if they were one benchmark.\n\nThe reference work is broad and mostly accurate. The paper does a solid job on pure ID-based models and side-information fusion, and the diagrams for fusion strategies (Figures 7 and 8) are genuinely helpful. The recent-advancement sections are a good index for newcomers, and the future-directions list is reasonable. The formalization of generative recommendation via the chain rule (Eq. 20) is a nice touch. This is not a research contribution, but it is a fair survey of the field as of late 2024.\n\nSoft spots, in order of severity. First, the claimed four-category taxonomy is not the organizing principle of the whole paper. Section 3.4 itself presents a three-aspect structure, and the recent-advancement sections use orthogonal axes. The novelty claim should be rewritten to say something like: \"we organize the classic SR literature by item-property construction, then survey recent advances by technique and problem setting.\" Second, Section 7.4 compares HR@10 and NDCG@10 numbers taken from three different papers without a shared protocol. The observations that follow — especially \"LLM-powered SR models show better performance than traditional SR models\" — are not supported. The paper should either re-run the models under one protocol or replace those conclusions with carefully hedged statements. These are fixable issues, but they are exactly what a referee should catch.\n\nWho is this for? An incoming grad student wanting a roadmap, or a practitioner checking what has been tried. It does not advance the science, but it is a legitimate survey reference. I would send it to a serious referee, with the expectation that the taxonomy framing and the empirical conclusions get corrected before publication.","headline":"Useful survey of sequential recommendation with a taxonomy claim that only covers half the paper and comparative numbers that do not support the conclusions.","tokens_in":43042,"tokens_out":2312,"would_cite":true,"duration_ms":22494,"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":"This survey claims that sequential recommendation research is best organized by how an item's properties are constructed—unique ID, side information, multi-modal features, or semantic and generative representations.","keywords":["sequential recommendation","survey","taxonomy","item properties","side information","multi-modal recommendation","generative recommendation","large language models"],"falsifier":"Re-run the representative models—SASRec, BERT4Rec, FDSA, S3-Rec, UniSRec, MISSRec, TIGER, SpecGR, P5, and Lite-LLM4Rec—on the same Amazon Office, Game, and Toy splits with identical preprocessing, candidate sets, and hyperparameter budgets; if the ranking implied by Table 5 does not reproduce, the survey's comparative conclusions would be refuted.","tokens_in":42065,"feed_emoji":"🧭","tokens_out":6632,"duration_ms":58011,"temperature":0.7,"pith_summary":"This paper is a survey that tries to establish a useful roadmap for sequential recommendation by classifying models according to how an item's properties are constructed. It groups the literature into pure ID-based models, models that add side information, multi-modal models, and recent paradigms including generative, LLM-powered, ultra-long-sequence, and data-augmented approaches. A sympathetic reader would care because the taxonomy gives a single axis for comparing efficiency, transferability, and cold-start behavior across a rapidly growing literature, and it identifies open problems such as open-domain, continuous, cloud-edge collaborative, and explainable SR.","feed_headline":"New survey organizes sequential recommendation by item construction","feed_subtitle":"From pure IDs to LLM and generative models, one taxonomy maps transferability, cold-start, and cost.","key_machinery":"The organizing machinery is the axis of item-property construction. An item in SR can be represented by a unique ID, an ID enriched with categorical, numerical, or graph features, modality embeddings from text, image, or video encoders, or semantic tokens from quantized generative tokenizers; the survey uses this spectrum as a lens through which all models are compared. This axis carries the argument because it converts scattered architectural differences into a single question: what is the item, and what information does that representation carry for transfer, sparsity, and cold-start performance.","core_discovery":"The survey's central claim is that sequential recommendation (SR) research is most clearly organized by asking how an item's properties are constructed. On that axis it distinguishes models that use only a unique item ID; models that pair the ID with categorical, numerical, knowledge-graph, or social features; models that replace or enrich the ID with multi-modal features such as text, images, and video; and the newest generation that builds semantic IDs, calls on large language models, retrieves from ultra-long sequences, or augments the data itself. The authors argue that this framing exposes the key trade-offs: ID-only models are efficient but cannot transfer across platforms or cope with cold-start, side information and multi-modal features improve sparsity and transferability at higher cost, and generative and LLM-based models point toward open-domain, explainable, and data-centric recommendation.","pith_inferences":["Inference: If the property-construction axis is adopted as a design principle, it suggests combining modality-based representations for cold-start and transfer with item IDs in warm-start regimes, since the survey notes IDs capture collaborative signals that modality features miss.","Inference: The Table 5 comparisons are copied from three separate papers rather than re-run; a controlled re-benchmark under one preprocessing and candidate-sampling protocol would be the natural test of the survey's performance ordering.","Inference: The taxonomy could be extended to judge future work by asking whether a new model moves along the axis, for instance toward fully semantic or ID-free representations, rather than by which architecture it uses."],"forward_implications":["If the taxonomy is right, the next wave of SR designs will be chosen by where they sit on the ID-to-semantic spectrum rather than by backbone architecture alone.","Pure ID-based models remain the practical efficiency baseline; side-information and multi-modal variants buy cold-start and transferability improvements at higher compute and tuning cost.","Generative SR with semantic IDs can keep embedding tables from growing linearly with item count and shares knowledge between similar items.","LLM-powered SR models improve semantic understanding and enable data generation, but their recommendation gains are demonstrated only under the comparisons assembled in the survey.","Ultra-long sequences are handled by retrieving a few relevant items before attention, which keeps inference latency bounded."],"supporting_citations":[{"why":"SASRec is the pure ID-based Transformer baseline that the survey compares against across model categories.","marker":"[1]"},{"why":"UniSRec is the leading example of modality-based SR using text descriptions for transferable sequence representations.","marker":"[8]"},{"why":"BERT4Rec is the pure ID-based bidirectional Transformer baseline used in the empirical comparisons.","marker":"[36]"},{"why":"S3-Rec is a self-supervised side-information model that supports the survey's claim that side information helps sparsity and cold-start.","marker":"[120]"},{"why":"FDSA is a feature-level self-attention model combining item IDs with side information and appears in the comparative table.","marker":"[121]"},{"why":"MISSRec is a multi-modal pre-training and transfer model whose results are copied into Table 5.","marker":"[138]"},{"why":"TIGER is the first generative retrieval model using semantic IDs and grounds the generative SR section.","marker":"[141]"},{"why":"SpecGR is a generative SR model whose experimental results are copied into Table 5 for direct comparison.","marker":"[147]"},{"why":"P5 is a representative LLM-based recommendation model that directly generates recommendations from interaction sequences.","marker":"[148]"},{"why":"Lite-LLM4Rec is an LLM architecture for sequential recommendation whose results support the survey's observation on LLM-powered performance.","marker":"[149]"}],"fun_headline_variants":["Sequential recommendation survey: from IDs to LLMs","How item construction defines sequential recommendation","New survey: organizing sequential recommendation by item properties","Taxonomy of sequential recommendation: ID, side info, multimodal, LLM","Sequential recommendation: one lens to map models and trade-offs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the performance numbers copied into Table 5 from three different papers are directly comparable despite likely differences in preprocessing, candidate sampling, and hyperparameters; if those conditions differ, the observed rankings between model families do not stand.","fun_headline_variants_meta":{"raw":{"variants":["Sequential recommendation survey: from IDs to LLMs","How item construction defines sequential recommendation","New survey: organizing sequential recommendation by item properties","Taxonomy of sequential recommendation: ID, side info, multimodal, LLM","Sequential recommendation: one lens to map models and trade-offs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000154,"raw_usage":{"total_tokens":1168,"prompt_tokens":857,"completion_tokens":311,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":233}},"tokens_in":473,"tokens_out":311,"duration_ms":3660,"temperature":1.0,"reasoning_tokens":233,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:44:13.347776+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the representative models—SASRec, BERT4Rec, FDSA, S3-Rec, UniSRec, MISSRec, TIGER, SpecGR, P5, and Lite-LLM4Rec—on the same Amazon Office, Game, and Toy splits with identical preprocessing, candidate sets, and hyperparameter budgets; if the ranking implied by Table 5 does not reproduce, the survey's comparative conclusions would be refuted.","supporting_citations":[{"cited_title":"MISSRec: Pre-training and transferring multi-modal interest-aware sequence representation for recommendation","cited_arxiv_id":null,"evidence_quote":"MISSRec is a multi-modal pre-training and transfer model whose results are copied into Table 5."},{"cited_title":"Rec- ommender systems with generative retrieval","cited_arxiv_id":null,"evidence_quote":"TIGER is the first generative retrieval model using semantic IDs and grounds the generative SR section."},{"cited_title":"Recommendation as language processing (rlp): A unified pretrain, person- alized prompt & predict paradigm (P5)","cited_arxiv_id":null,"evidence_quote":"P5 is a representative LLM-based recommendation model that directly generates recommendations from interaction sequences."}],"review_version":1}