Paper deep dive
Revisiting Generative vs. Dense Retrieval: A Static Value Prior Reverses the Gap
Renbo Tu, Yang Liu, Shuying Sun
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 10/15/2026, 4:55:45 AM
Summary
This paper compares Generative Retrieval (GR) and Dense Retrieval (DR) for sequential recommendation in e-commerce. It finds that GR appears superior in 'best relevant item value' primarily due to popularity bias, while DR leads in exact-target recall and overall list value. The authors propose a training-free static item value prior that, when applied to DR, reverses GR's value advantage while maintaining recall superiority. Industrial evaluation with an LLM judge confirms DR outperforms a much larger GR model.
Entities (13)
Relation Signals (8)
Popularity Bias → causes → Generative Retrieval Value Lead
confidence 92% · popularity contributes the most to GR’s measured value lead... 49.8% of its value lead on 2014 and 65.4% on 2023
Static Item Value Prior → appliedto → Dense Retrieval
confidence 90% · Adding the same static item value to DR in one reranking step reverses this lead
Dense Retrieval → outperforms → Generative Retrieval
confidence 90% · DR leads on exact-target recall and relevant value across the ranked list
Static Item Value Prior → reverses → Generative Retrieval Value Lead
confidence 90% · Adding the same static item value to DR in one reranking step reverses this lead and keeps recall above GR.
GTE Encoder → usedby → Dense Retrieval
confidence 90% · DR uses a GTE encoder with about 300 million parameters.
Qwen3-14B → usedby → Generative Retrieval
confidence 90% · GR uses a Qwen3 14B model that generates semantic IDs
Dense Retrieval → outperforms → Qwen3-14B
confidence 85% · a 14B SID-LLM still underperforms a 47× smaller dense encoder.
Generative Retrieval → outperforms → Dense Retrieval
confidence 85% · GR leads on best relevant item value among users both methods cover
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Generative retrieval (GR) has emerged as a strong candidate generator for sequential recommendation as language models and semantic-item identifiers have improved. However, GR is not a uniformly stronger replacement of dense retrieval (DR) in online retail candidate generation. We examine their differences on two Amazon Toys datasets under different metrics. GR leads on best relevant item value among users both methods cover, while DR leads on exact-target recall and relevant value across the ranked list. With deep analysis, we find that popularity contributes the most to GR's measured value lead. Adding the same static item value to DR in one reranking step reverses this lead and keeps recall above GR. Finally, on an industrial-scale internal e-commerce dataset with LLM-judged relevance, we show that a 14B SID-LLM still underperforms a $47\times$ smaller dense encoder.
Tags
Links
- Source: https://arxiv.org/abs/2610.19065v1
- Canonical: https://arxiv.org/abs/2610.19065v1
Trouble viewing inline? Open PDF directly →
Full Text
35,276 characters extracted from source content.
Expand or collapse full text
Research ://anonymous.4open.science/r/gr-dr-release-BCE4 Revisiting Generative vs. Dense Retrieval: A Static Value Prior Reverses the Gap Conference: The 35th ACM International Conference on Information and Knowledge Management; November 7–11, 2026; Rome, Italy11DOI: X.X Renbo Tu Note: This work was completed at Shopify. Affiliation: Shopify Affiliation: University of Toronto , Yang Liu Affiliation: Shopify and Shuying Sun Note: Corresponding author: shuying.sun@shopify.com. Affiliation: Shopify 2026 Abstract. Generative retrieval (GR) has emerged as a strong candidate generator for sequential recommendation as language models and semantic-item identifiers have improved. However, GR is not a uniformly stronger replacement of dense retrieval (DR) in online retail candidate generation. We examine their differences on two Amazon Toys datasets under different metrics. GR leads on best relevant item value among users both methods cover, while DR leads on exact-target recall and relevant value across the ranked list. With deep analysis, we find that popularity contributes the most to GR’s measured value lead. Adding the same static item value to DR in one reranking step reverses this lead and keeps recall above GR. Finally, on an industrial-scale internal e-commerce dataset with LLM-judged relevance, we show that a 14B SID-LLM still underperforms a 47×47× smaller dense encoder. Our code and artifacts are open-sourced at https://anonymous.4open.science/r/gr-dr-release-BCE4. Keywords: generative retrieval, dense retrieval, sequential recommendation, semantic IDs, popularity bias, recommendation evaluation 1. Introduction Dense retrieval (DR) has long been a standard candidate-generation approach for recommendation. It encodes a user’s interaction history and catalog items in a shared embedding space, then retrieves nearby items (Covington et al., 2016; Yi et al., 2019). Advances in language models and semantic IDs have made generative retrieval (GR) an increasingly viable alternative. GR represents items with semantic IDs and generates the next item’s ID one token at a time (De Cao et al., 2021; Tay et al., 2022; Wang et al., 2022; Lee et al., 2022; Rajput et al., 2023). Recent work has demonstrated GR at industrial scale (Zheng et al., 2024; Singh et al., 2024; Zhou and others, 2025a; Zhou and others, 2025b), but its use does not establish that it consistently outperforms or can replace DR in candidate generation. As shown in Fig. 1, normally DR retrieves the held out next item more often, while GR’s best relevant item has higher value on average for users both methods cover. Figure 1. GR and DR use the same user history but generate candidates through semantic ID decoding and dense search. GR leads on best item value, while DR leads on exact target retrieval and full list value. A static prior reranks DR candidates and moves the potential valuable item into the top k. The apparent winner depends on how retrieval is evaluated. This metric choice is particularly crucial in e-commerce candidate generation, where retrieved products differ not only in relevance to the user but also its commercial value. To understand why, we need to know whether GR’s value lead comes from the user’s history or from generally valuable items. Recall and target NDCG show whether the held out item is retrieved and where it ranks. They do not account for other relevant items, their ranks, or their value. To understand the disagreement, in this paper, we measure both the best relevant item and relevant value across the full list. We apply these metrics in controlled GR and DR comparisons on Amazon Toys 2014 and 2023 (McAuley et al., 2015; He and McAuley, 2016; Hou et al., 2024; Rajput et al., 2023; Yang et al., 2024). The metric disagreement persists on both public e-commerce datasets. From analysis, popularity is the largest component of GR’s measured value lead. Among users where GR wins, popularity contributes 49.8% of its value lead on 2014 and 65.4% on 2023. Reranking DR with the static item value score reverses the best item value gap while preserving a recall lead over GR. Further checks with cold targets, substitute credit, a random control, and graded industrial relevance do not reveal a consistent GR advantage. Our main contributions are: • We define complementary metrics for sequential recommendation in e-commerce scenario and show that they favor different methods in the GR and DR comparison. • We decompose GR’s measured value lead and find that popularity is its largest component on both public datasets. • We propose a training-free value prior, applied to DR as a single rerank of its retrieval results. This reverses GR’s lead in best relevant item value while preserving DR’s recall lead. 2. Related Work The Generative Turn for Retrieval Systems Prior to decoder-only approaches, sequential recommendation models for retrieval have been prevalent in production (Covington et al., 2016; Yi et al., 2019). The canonical dense retrieval (DR) is therefore defined as bidirectional sequential models for next-item prediction (Hidasi et al., 2016; Kang and McAuley, 2018; Sun et al., 2019). Generative retrieval (GR) enters the foreground as a constrained autoregressive decoding mechanism to generative identifier tokens, semantic IDs (SIDs), mappable to an underlying item, whose codebook construction follows RQ-VAE (De Cao et al., 2021; Tay et al., 2022; Wang et al., 2022; Lee et al., 2022). SID generative retrieval benefits from the rise of LLMs and demonstrates generalization, establishing itself for industry applications (Rajput et al., 2023; Zheng et al., 2024; Singh et al., 2024; Zhou and others, 2025a; Zhou and others, 2025b). Note that GR umbrellas non-SID approaches such as P5 and HSTU (Geng et al., 2022; Zhai et al., 2024; Liu et al., 2026). Our controlled DR-GR comparison is grounded on LIGER, while research on GR’s scaling difficulties and DR’s capacity limits bounds our conclusions without determining them (Yang et al., 2024; Pradeep et al., 2023; Weller et al., 2026). Difficulty of Retrieval Evaluation Offline recommendation evaluation suffers from sensitivity to missing feedback, candidate sampling and temporal splitting decisions (Steck, 2010; Cremonesi et al., 2010; Krichene and Rendle, 2020; Cañamares and Castells, 2020; Dallmann et al., 2021; Ji et al., 2023). Audits on these decisions reveal that they can change recommendation rankings, while recent studies identify GR artifacts from SID collision and benchmark shortcuts(Ferrari Dacrema et al., 2019; Hidasi and Czapp, 2023; Sun, 2023; Zhang et al., 2026; Han et al., 2026). Our contribution identifies substitute non-identifiability under exact-target metrics and show how it worsens with catalog scale. Popularity Bias is a Static Prior In recommendation, popularity of certain products are deeply associated with benchmark accuracy (Steck, 2011; Jannach et al., 2015). However, popularity also presents itself as a bias that degrades exposure and fairness for non-popular items (Abdollahpouri et al., 2019; Klimashevskaia et al., 2024). Popularity bias in GR becomes prominent due to inherent bias within LLMs (Zhang et al., 2023; Yin et al., 2026; Fan et al., 2026). While these works mitigate popularity bias, our result show that under a same model GR-DR comparison, the advantage from popularity is replaceable by a user-independent prior that can be transferred to DR without decoding. 3. Controlled Comparison Setup Problem Definition In the e-commerce setting, for each user u, let hu=(i1,…,it)h_u=(i_1,…,i_t) denote the observed user product interaction history and yu=it+1y_u=i_t+1 the held out next item. A candidate generator x uses huh_u to produce the ranked list Lxk(u)=(ix,u(1),…,ix,u(k)),L_x^k(u)= (i_x,u^(1),…,i_x,u^(k) ), where ix,u(j)i_x,u^(j) is the item at rank j. Models We compare GR and DR using the TIGER and LIGER framework (Rajput et al., 2023; Yang et al., 2024). Both arms use the same interaction data, item catalog, and jointly trained T5 encoder decoder. GR decodes valid semantic IDs through prefix constrained beam search, while DR ranks items by cosine similarity using the encoder representation of the same history. Thus, the arms differ mainly in their retrieval readout. Datasets Our public evaluation uses the 2014 and 2023 releases of Amazon Toys. The 2014 dataset contains 11,924 items and 19,412 user histories under a leave last out split. Among their held out targets, 19,116 were seen during training and 296 were absent from training. Our results reproduce the LIGER Recall@10 values, while our GR recall exceeds the published TIGER result. The 2023 catalog is 13.6×13.6× larger, which tests whether the observed pattern and the proposed intervention survive a substantial increase in catalog scale (Hou et al., 2024). We transfer the value weights and reranking coefficient from 2014 to 2023 without change unless stated otherwise. 4. Evaluation Metrics We compare GR and DR by asking which retrieved items are relevant and how much value they carry. A relevance labeling rule qu(i)∈[0,1]q_u(i)∈[0,1] assigns evaluation credit after ranking and is not a model score. It may credit only the held-out item, include substitute candidates, or assign partial credit through graded judgments. In contrast, vi∈[0,1]v_i∈[0,1] measures a static property of item i. For any rule q, the relevant set with k candidates is defined as Rxk(u,q)=i∈Lxk(u)∣qu(i)>0.R_x^k(u;q)=\i∈ L_x^k(u) q_u(i)>0\. This definition keeps relevance and item value separate across evaluation protocols. In this section, We first define exact-target and category-based relevance, then introduce static catalog value, and finally combine them at the best-item and whole-list levels. 4.1. Exact-Target Retrieval The conventional next item protocol assigns credit only to the held out item, as defined by quexact(i)=[i=yu]q_u^exact(i)= 1[i=y_u]. Recall@k measures whether this item appears in the top-k results. Recallx@k=1||∑u∈[yu∈Lxk(u)].Recall_x@k= 1|U| _u 1[y_u∈ L_x^k(u)]. NDCG@k evaluates the same target with a rank discount (Järvelin and Kekäläinen, 2002). Both metrics measure recovery of the observed next interaction and give no credit to other potentially useful items. 4.2. Beyond Exact Match: Substitute Blindness Exact target evaluation cannot distinguish a substitute from an irrelevant item because both receive zero credit. We refer to this limitation as substitute blindness. Therefore we define a coarse proxy for substitute-aware relevance using the shared leaf category as qucat(i)=[leaf(i)=leaf(yu)],S(u)=i≠yu:qucat(i)=1.q_u^cat(i)= 1[leaf(i)=leaf(y_u)], S(u)=\i≠ y_u:q_u^cat(i)=1\. where leaf(i)leaf(i) and leaf(yu)leaf(y_u) identify the leaf categories of candidate item i and the held-out target. An item receives qucat(i)=1q_u^cat(i)=1 when their categories match. A larger substitute set |S(u)||S(u)| makes an exact miss less informative, but category overlap alone does not establish that two items are substitutes. We therefore stress-test this proxy using co-purchase credit, a mass-matched random-credit control, and LLM-judged graded relevance in Section 6. 4.3. Static Item Value Relevance determines whether an item is useful for a given user. In e-commerce, products also differ in static catalog attributes that may matter to the platform. We summarize these attributes with a transparent item-value proxy (1) vi()=wppopi+wrratei+wbbrandi,wp+wr+wb=1.v_i(w)=w_p\,pop_i+w_r\,rate_i+w_b\,brand_i, w_p+w_r+w_b=1. Here, popipop_i, rateirate_i, and brandibrand_i are the catalog-percentile ranks of log popularity, the Wilson lower bound on star rating, and brand footprint, respectively (Wilson, 1927). All weights are nonnegative. Unless otherwise stated, we use the default weights =(0.25,0.25,0.50)w=(0.25,0.25,0.50). This score is a transparent proxy for item value, not a universal measure of recommendation utility. In the sensitivity analysis, we vary w over nonnegative mixtures whose components sum to one. Unlike qu(i)q_u(i), which varies across users and relevance labeling rules, vi()v_i(w) is computed once for each catalog item. 4.4. Best-Item Value vs. Whole-List Value Having defined relevance q and item value v separately, we evaluate each ranked list using two complementary metrics. Conditional best-item value (BV) BV asks how valuable the best relevant item returned by method x is for user u. BVxk(u,q,)=maxi∈Rxk(u,q)vi().BV_x^k(u;q,w)= _i∈ R_x^k(u;q)v_i(w). In our main analysis, BV uses the binary category rule qcatq^cat from Sec 4.2 to select relevant items but does not weight them by relevance magnitude. Let x,akC_x,a^k denote the users for whom both method x and comparator a retrieve at least one relevant item. We report their average gap as ΔBVx−a@k=1|x,ak|∑u∈x,ak[BVxk(u)−BVak(u)], _x-a@k= 1|C_x,a^k| _u _x,a^k [BV_x^k(u)-BV_a^k(u) ], where q and w are omitted for readability. We compare the BV scores of methods x and a for each user u. A difference above τ=0.05τ=0.05 is a win for x, a difference below −τ-τ is a win for a, and all other differences are ties. Whole list Value. BV focuses on the best relevant item. Its comparison includes only users with a relevant item from both methods. To score the full top k list for every user, we use NDCG with item value weights (vNDCGvNDCG). Each item contributes its relevance times its value, with less credit at lower ranks. For method x and user u, let ij=ix,u(j)i_j=i_x,u^(j) be the item at rank j. The list score is vDCGxk(u)=∑j=1kqu(ij)vij()log2(j+1).vDCG_x^k(u)= _j=1^k q_u(i_j)v_i_j(w) _2(j+1). For each user, we pool the items returned by all methods and rank them by relevance times value, qu(i)vi()q_u(i)v_i(w). The discounted score of the top k items is the ideal score IDCGk(u)IDCG^k(u). We divide each method’s score by this shared score, then average across users vNDCGx@k=1||∑u∈vDCGxk(u)IDCGk(u).vNDCG_x@k= 1|U| _u vDCG_x^k(u)IDCG^k(u). Table 1. Amazon Toys results. Positive Δ @50 values favor the listed method over GR. Best values are bold. Recall@50 NDCG@50 Δ @50 vNDCG@50vNDCG@50 Method 2014 2023 2014 2023 2014 2023 2014 2023 GR 0.167 0.052 0.056 0.017 — — 0.215 0.160 DR 0.228 0.084 0.077 0.029 −0.020-0.020 −0.016-0.016 0.222 0.168 DR+prior 0.210 0.083 0.072 0.028 +0.023+0.023 +0.006+0.006 0.240 0.175 If the ideal score is zero for a user, their vNDCGvNDCG score is zero. All methods use the same ideal score for each user, so their scores can be compared within the fixed pool. Absolute scores from different method pools should not be compared directly. To check whether coverage alone explains the metric gap, we also compare BV over all users, performing imputation by assigning zero when a method retrieves no relevant item. 5. From Metric Dependency to a Static Prior We now apply these complementary metrics in a three-step analysis. We first show that GR leads on BV while DR leads on vNDCGvNDCG. We then measure how popularity, rating, and brand contribute to GR’s value advantage. Finally, we rerank DR with the same static item value and test whether the gap reverses. 5.1. The Metric-Dependent GR–DR Gap We begin with the empirical disagreement. GR leads on BV, while raw DR leads on vNDCG@50vNDCG@50, Recall@50, and target NDCG@50 on both datasets as shown in Table 1. GR wins the paired BV comparison more than twice as often as raw DR (Figure 2(a)), and its lead remains when users without a relevant retrieved item receive zero. DR’s vNDCG@50vNDCG@50 lead is 2.9% in 2014 and 5.1% in 2023, with confidence intervals above zero (Figure 2(b)). Value weighting compresses DR’s relevance advantage by an order of magnitude. We conduct a full sweep on value weights to demonstrate robustness. Across the 231-point weight simplex shown in Fig. 3, the sign of the vNDCGvNDCG gap holds on 223 points on 2014, failing only in popularity-dominated edge cases, and on all 231 points on 2023. Figure 2. Metric choice reverses the result. (a) GR wins more often under BV@50BV@50 in both datasets. (b) Under vNDCG@50vNDCG@50, DR scores higher than GR and DR+prior scores highest. Error bars show standard errors for adjacent differences. 5.2. Diagnosing GR’s Value Advantage Among users where GR wins BVBV, we use Equation 1 to break its value lead over DR into popularity, rating, and brand. Popularity contributes for 49.8% of the lead on 2014 and 65.4% on 2023. It is the largest component on both datasets, and its share is higher on the larger 2023 catalog. Rating and brand contribute the remainder, but neither provides the same cross-dataset dominance. This pattern narrows the interpretation of GR’s value edge. Each value component is computed once per item and does not depend on the current history. The measured lift is therefore largely captured by static item-level properties, especially popularity. Repeated exposure in training offers a possible explanation for GR’s value lead. The released 2014 popularity score correlates with how often held out targets appear in training histories (Spearman ρ=0.582ρ=0.582). At least ten training histories contain 50.3% of held out targets in 2014 and 75.1% in 2023. GR may learn a preference for frequently seen items while predicting their semantic IDs. This could contribute to its lead because popularity is part of the item value score. That statement yields a falsifiable implication. If static item properties account for the operational advantage, then explicitly adding the same information to DR should reproduce it without autoregressive decoding. 5.3. Transferring the Static Prior to DR To test whether static item value can close GR’s lead in BVBV, we rerank DR’s top 200 candidates for each user. We call this method DR+prior. DR+prior scales the DR similarity scores to 0,10,1 within each pool, then average each score with the item’s value vi()v_i(w) using the default weights from BVBV in Sec 4.3. Because BVBV uses the same item value, this test checks whether an explicit catalog prior can reproduce GR’s measured lead. This requires no retraining or decoder at serving time. We score each candidate by σu(i)=(1−α)s~u(i)+αvi,α=0.5,i∈Pu _u(i)\;=\;(1-α)\, s_u(i)\;+\;α\,v_i, α=0.5, i∈ P_u As results, adding the prior turns DR’s BV@50BV@50 deficit against GR into a lead. On 2014, DR moves from trailing by 0.0200.020 to leading by 0.0230.023. On 2023, it moves from trailing by 0.0160.016 to leading by 0.0060.006. As shown in Table 1, DR+prior also has the highest vNDCG@50vNDCG@50 on both datasets, at 0.240 and 0.175. Its Recall@50 declines from raw DR but stays above GR. Figure 3. Sign of the raw DR-vs-GR vNDCG@50vNDCG@50 gap over the static value on Amazon Toys datasets. Grid step is 0.05. 6. Robustness Beyond Exact-Match Evaluation The public results show that the item prior closes GR’s value gap. We now test whether GR does better on cold targets or when useful alternatives to the held out target receive credit. For the latter, we use public substitute evidence and graded judgments on an industrial recommendation dataset. 6.1. Cold Items We test whether generation helps retrieve targets absent from interaction training. GR finds none of 296 cold targets on 2014 and 5 of 1,899 on 2023. DR finds 47 and 27, while DR+prior finds 36 and 21, respectively. Both DR methods find more cold targets than GR on each dataset. 6.2. Substitute Credit and Graded Relevance An exact target miss does not tell us whether a list contains a useful alternative. For the median user, the held out target has 157 same-leaf alternatives on 2014 and 1,046 on 2023, a 6.7×6.7× increase. Among DR’s 8,635 exact target misses on 2014, 58.2% include an item from that category in the top 50 while the target ranks beyond 200. In another 25.7%, the target ranks between 50 and 200. GR finds the exact target in only 748 of these DR misses, so the misses alone do not show an advantage for GR. We give binary credit to items linked to the target in Amazon co purchase data, which is reported in Table 2. In this test, a 0.0447 lead for DR under exact Recall@50 becomes a 0.0183 lead for GR. For the binary test, randomly chosen items given the same amount of credit move the gap at least as far toward GR. With graded NDCG, DR leads again by 0.0153. The binary reversal therefore does not establish that GR retrieves better alternatives. Leaf category overlap and co-purchase links are imperfect evidence of usefulness. We next test graded relevance on the industrial dataset. Table 2. Substitute credit on 2014. The last column shows the extra shift toward GR from co purchase credit compared with random credit of equal total weight. DR–GR Real–random credit Recall@50, exact only +0.0447+0.0447 — ++ co-purchase credit −0.0183-0.0183 −0.0190-0.0190 ++ graded NDCG@50 +0.0153+0.0153 +0.0111+0.0111 6.3. Industrial LLM-Judged Relevance The industrial evaluation uses real e-commerce interaction histories and a product catalog. GR uses a Qwen3 14B model that generates semantic IDs, while DR uses a GTE encoder with about 300 million parameters. GR has roughly 47×47× as many parameters. A substitute-aware LLM judge calibrated with 250 human labels scores each candidate’s relevance from zero to one, allowing partial credit for useful alternatives. Because the models differ in backbone and size, this evaluation tests whether the public result also holds in this setting. It cannot isolate the effect of retrieval architecture. Table 3 shows that giving useful alternatives graded credit preserves the order GR << DR << DR+prior on the industrial dataset. The prior raises DR’s NDCG@50 by 11.8%. Table 3. Industrial dataset evaluation under an LLM-judge. Population (NDCG@50) GR DR DR+prior Full evaluation set (n=5,998n=5,998) 0.453 0.474 0.530 7. Conclusion In this paper, we revisited the GR and DR in a controlled e-commerce comparison and we find two distinct parts to GR’s measured advantage. First, metric choice changes the apparent winner. GR leads on the value of the best relevant item, while DR leads on exact target recall and relevant value across the ranked list on both public datasets. Second, popularity makes the largest contribution to GR’s value lead. Adding the same static item value to DR in one reranking step reverses this gap while preserving DR’s recall lead. Cold target and substitute credit checks do not reveal a broader GR advantage. Graded judgments on an industrial dataset place DR above a larger GR model and DR with the prior above both. These results make DR with a static value prior a useful baseline for future GR systems. References Abdollahpouri et al. (2019) H. Abdollahpouri, M. Mansoury, R. Burke, and B. Mobasher The unfairness of popularity bias in recommendation. arXiv preprint arXiv:1907.13286. External Links: 1907.13286, Link Cited by: §2. Cañamares and Castells (2020) R. Cañamares and P. Castells On target item sampling in offline recommender system evaluation. In Proceedings of the 14th ACM Conference on Recommender Systems, p. 259–268. External Links: Document Cited by: §2. Covington et al. (2016) P. Covington, J. Adams, and E. Sargin Deep neural networks for YouTube recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems, p. 191–198. External Links: Document Cited by: §1, §2. Cremonesi et al. (2010) P. Cremonesi, Y. Koren, and R. Turrin Performance of recommender algorithms on top-N recommendation tasks. In Proceedings of the 4th ACM Conference on Recommender Systems, p. 39–46. External Links: Document Cited by: §2. Dallmann et al. (2021) A. Dallmann, D. Zoller, and A. Hotho A case study on sampling strategies for evaluating neural sequential item recommendation models. In Proceedings of the 15th ACM Conference on Recommender Systems, p. 505–514. External Links: Document Cited by: §2. De Cao et al. (2021) N. De Cao, G. Izacard, S. Riedel, and F. Petroni Autoregressive entity retrieval. In 9th International Conference on Learning Representations (ICLR), External Links: 2010.00904, Link Cited by: §1, §2. Fan et al. (2026) Z. Fan, Z. Chen, L. Ma, J. Huang, L. Morishetti, K. Nag, S. Kumar, and K. Achan CRAB: codebook rebalancing for bias mitigation in generative recommendation. arXiv preprint arXiv:2604.05113. External Links: 2604.05113, Link Cited by: §2. Ferrari Dacrema et al. (2019) M. Ferrari Dacrema, P. Cremonesi, and D. Jannach Are we really making much progress? a worrying analysis of recent neural recommendation approaches. In Proceedings of the 13th ACM Conference on Recommender Systems, p. 101–109. External Links: Document Cited by: §2. Geng et al. (2022) S. Geng, S. Liu, Z. Fu, Y. Ge, and Y. Zhang Recommendation as language processing (RLP): a unified pretrain, personalized prompt & predict paradigm (P5). In Proceedings of the 16th ACM Conference on Recommender Systems, p. 299–315. External Links: Document Cited by: §2. Han et al. (2026) H. Han, L. Ma, H. Wang, B. Li, D. Zha, C. H. Tan, H. Gao, X. Liu, S. Moyerman, S. Katariya, H. Liu, and J. Tang An embarrassingly simple graph heuristic reveals shortcut-solvable benchmarks for sequential recommendation. arXiv preprint arXiv:2605.07125. External Links: 2605.07125, Link Cited by: §2. He and McAuley (2016) R. He and J. McAuley Ups and downs: modeling the visual evolution of fashion trends with one-class collaborative filtering. In Proceedings of the 25th International Conference on World Wide Web, p. 507–517. External Links: Document Cited by: §1. Hidasi and Czapp (2023) B. Hidasi and Á. T. Czapp Widespread flaws in offline evaluation of recommender systems. In Proceedings of the 17th ACM Conference on Recommender Systems, p. 848–855. External Links: Document Cited by: §2. Hidasi et al. (2016) B. Hidasi, A. Karatzoglou, L. Baltrunas, and D. Tikk Session-based recommendations with recurrent neural networks. In 4th International Conference on Learning Representations (ICLR), External Links: 1511.06939, Link Cited by: §2. Hou et al. (2024) Y. Hou, J. Li, X. Fu, Z. He, A. Yan, X. Chen, and J. McAuley Bridging language and items for retrieval and recommendation: benchmarking LLMs as semantic encoders. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 3251–3265. External Links: Document, 2403.03952, Link Cited by: §1, §3. Jannach et al. (2015) D. Jannach, L. Lerche, I. Kamehkhosh, and M. Jugovac What recommenders recommend: an analysis of recommendation biases and possible countermeasures. User Modeling and User-Adapted Interaction 25 (5), p. 427–491. External Links: Document Cited by: §2. Järvelin and Kekäläinen (2002) K. Järvelin and J. Kekäläinen Cumulated gain-based evaluation of IR techniques. ACM Transactions on Information Systems 20 (4), p. 422–446. External Links: Document Cited by: §4.1. Ji et al. (2023) Y. Ji, A. Sun, J. Zhang, and C. Li A critical study on data leakage in recommender system offline evaluation. ACM Transactions on Information Systems 41 (3), p. 1–27. External Links: Document Cited by: §2. Kang and McAuley (2018) W. Kang and J. McAuley Self-attentive sequential recommendation. In 2018 IEEE International Conference on Data Mining (ICDM), p. 197–206. External Links: Document Cited by: §2. Klimashevskaia et al. (2024) A. Klimashevskaia, D. Jannach, M. Elahi, and C. Trattner A survey on popularity bias in recommender systems. User Modeling and User-Adapted Interaction 34 (5), p. 1777–1834. External Links: Document Cited by: §2. Krichene and Rendle (2020) W. Krichene and S. Rendle On sampled metrics for item recommendation. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, p. 1748–1757. External Links: Document Cited by: §2. Lee et al. (2022) D. Lee, C. Kim, S. Kim, M. Cho, and W. Han Autoregressive image generation using residual quantization. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 11513–11522. External Links: Document Cited by: §1, §2. Liu et al. (2026) Y. Liu, N. Loo, A. Khanafer, S. Sun, A. Soni, Z. Wu, and L. Yang T-RoPE: time-aware rotary position embedding for sequential recommendation. arXiv preprint arXiv:2609.30576. External Links: 2609.30576, Link Cited by: §2. McAuley et al. (2015) J. McAuley, C. Targett, Q. Shi, and A. van den Hengel Image-based recommendations on styles and substitutes. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 43–52. External Links: Document Cited by: §1. Pradeep et al. (2023) R. Pradeep, K. Hui, J. Gupta, A. D. Lelkes, H. Zhuang, J. Lin, D. Metzler, and V. Q. Tran How does generative retrieval scale to millions of passages?. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, p. 1305–1321. External Links: Document, 2305.11841, Link Cited by: §2. Rajput et al. (2023) S. Rajput, N. Mehta, A. Singh, R. H. Keshavan, T. Vu, L. Heldt, L. Hong, Y. Tay, V. Q. Tran, J. Samost, M. Kula, E. H. Chi, and M. Sathiamoorthy Recommender systems with generative retrieval. In Advances in Neural Information Processing Systems, Vol. 36, p. 10299–10315. External Links: Document, 2305.05065, Link Cited by: §1, §1, §2, §3. Singh et al. (2024) A. Singh, T. Vu, N. Mehta, R. Keshavan, M. Sathiamoorthy, Y. Zheng, L. Hong, L. Heldt, L. Wei, D. Tandon, E. Chi, and X. Yi Better generalization with semantic IDs: a case study in ranking for recommendations. In Proceedings of the 18th ACM Conference on Recommender Systems, p. 1039–1044. External Links: Document, 2306.08121, Link Cited by: §1, §2. Steck (2010) H. Steck Training and testing of recommender systems on data missing not at random. In Proceedings of the 16th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, p. 713–722. External Links: Document Cited by: §2. Steck (2011) H. Steck Item popularity and recommendation accuracy. In Proceedings of the 5th ACM Conference on Recommender Systems, p. 125–132. External Links: Document Cited by: §2. Sun (2023) A. Sun Take a fresh look at recommender systems from an evaluation standpoint. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 2629–2638. External Links: Document Cited by: §2. Sun et al. (2019) F. Sun, J. Liu, J. Wu, C. Pei, X. Lin, W. Ou, and P. Jiang BERT4Rec: sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, p. 1441–1450. External Links: Document Cited by: §2. Tay et al. (2022) Y. Tay, V. Tran, M. Dehghani, J. Ni, D. Bahri, H. Mehta, Z. Qin, K. Hui, Z. Zhao, J. Gupta, T. Schuster, W. Cohen, and D. Metzler Transformer memory as a differentiable search index. In Advances in Neural Information Processing Systems, Vol. 35, p. 21831–21843. External Links: Document Cited by: §1, §2. Wang et al. (2022) Y. Wang, Y. Hou, H. Wang, Z. Miao, S. Wu, Q. Chen, Y. Xia, C. Chi, G. Zhao, Z. Liu, X. Xie, H. Sun, W. Deng, Q. Zhang, and M. Yang A neural corpus indexer for document retrieval. In Advances in Neural Information Processing Systems, Vol. 35, p. 25600–25614. External Links: Document Cited by: §1, §2. Weller et al. (2026) O. Weller, M. Boratko, I. Naim, and J. Lee On the theoretical limitations of embedding-based retrieval. In 14th International Conference on Learning Representations (ICLR), External Links: 2508.21038, Link Cited by: §2. Wilson (1927) E. B. Wilson Probable inference, the law of succession, and statistical inference. Journal of the American Statistical Association 22 (158), p. 209–212. External Links: Document Cited by: §4.3. Yang et al. (2024) L. Yang, F. Paischer, K. Hassani, J. Li, S. Shao, Z. G. Li, Y. He, X. Feng, N. Noorshams, S. Park, B. Long, R. D. Nowak, X. Gao, and H. Eghbalzadeh Unifying generative and dense retrieval for sequential recommendation. arXiv preprint arXiv:2411.18814. External Links: 2411.18814, Link Cited by: §1, §2, §3. Yi et al. (2019) X. Yi, J. Yang, L. Hong, D. Z. Cheng, L. Heldt, A. Kumthekar, Z. Zhao, L. Wei, and E. Chi Sampling-bias-corrected neural modeling for large corpus item recommendations. In Proceedings of the 13th ACM Conference on Recommender Systems, p. 269–277. External Links: Document Cited by: §1, §2. Yin et al. (2026) J. Yin, B. Zhu, P. Huo, R. Liu, H. Chen, S. Wang, S. Pan, and C. Zhang Echoes in filter bubble: diagnosing and curing popularity bias in generative recommenders. arXiv preprint arXiv:2605.16825. External Links: 2605.16825, Link Cited by: §2. Zhai et al. (2024) J. Zhai, L. Liao, X. Liu, Y. Wang, R. Li, X. Cao, L. Gao, Z. Gong, F. Gu, J. He, Y. Lu, and Y. Shi Actions speak louder than words: trillion-parameter sequential transducers for generative recommendations. In Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 235, p. 58484–58509. External Links: Link Cited by: §2. Zhang et al. (2023) J. Zhang, K. Bao, Y. Zhang, W. Wang, F. Feng, and X. He Is ChatGPT fair for recommendation? evaluating fairness in large language model recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems, p. 993–999. External Links: Document Cited by: §2. Zhang et al. (2026) Q. Zhang, L. Szymanski, H. Zhang, and J. D. Deng How reliable are semantic-ID tokenizer comparisons in generative recommendation?. arXiv preprint arXiv:2605.25330. External Links: 2605.25330, Link Cited by: §2. Zheng et al. (2024) B. Zheng, Y. Hou, H. Lu, Y. Chen, W. X. Zhao, M. Chen, and J. Wen Adapting large language models by integrating collaborative semantics for recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), p. 1435–1448. External Links: Document Cited by: §1, §2. Zhou et al. (2025a) G. Zhou et al. OneRec technical report. arXiv preprint arXiv:2506.13695. External Links: 2506.13695, Link Cited by: §1, §2. Zhou et al. (2025b) G. Zhou et al. OneRec-V2 technical report. arXiv preprint arXiv:2508.20900. External Links: 2508.20900, Link Cited by: §1, §2.