Pith. sign in

REVIEW 6 major objections 6 minor 62 references

BONSAI shows that decoding-trie shape—low branching, variable depth—is a primary driver of LLM recommendation accuracy, with relative gains up to 21.6%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 20:23 UTC pith:M4PTTIKH

load-bearing objection BONSAI's trie-optimization idea is worth engaging; the paper overclaims when it attributes most of the gain to trie structure rather than feature quality. the 6 major comments →

arxiv 2607.16633 v1 pith:M4PTTIKH submitted 2026-07-18 cs.IR

Beyond Fixed Depths and Widths: Optimizing Textual Decoding Tries in LLM-based Generative Recommendation

Jingzhe Liu , Hanbing Wang , Jiliang Tang , Liam Collins , Tong Zhao , Neil Shah , Mingxuan Ju This is my paper
classification cs.IR
keywords generative recommendationdecoding trieconstrained beam searchterm IDsminimum set covervariable-length identifiersbranching factorLLM-based recommendation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that in LLM-based generative recommendation, the shape of the decoding trie—the prefix tree that constrains beam search over item term IDs—is a first-order determinant of accuracy, not a byproduct of term semantics. It identifies two properties a performant trie should have: adaptive variable-length IDs, so items with richer metadata get deeper paths, and low branching factors near the root, so early decoding decisions are easy. BONSAI builds such a trie by recursively selecting a minimum set of shared keywords to split each node, using a greedy set-cover approximation. On three Amazon datasets, BONSAI reports 16.0–21.6% relative improvements over the strongest baselines, and it shows the same structural principles transfer to existing term-ID methods when their keywords are reordered. The contribution is a shift of focus: design the search structure jointly with the identifiers, not only the semantics.

Core claim

The central claim is that the success of constrained beam search in generative recommendation is governed by trie structure as much as by the semantic quality of term IDs. Concretely, the paper proves (under stated regularity conditions) that an optimal trie can be found by Bellman dynamic programming, that greedy minimum set cover approximates the optimal split at each node within a factor of 1/(rho·alpha), and that variable-depth tries strictly outperform any fixed-depth trie. BONSAI operationalizes this by representing each item as a set of extracted keywords, then recursively partitioning the item set with a minimum cover of shared features; item IDs are the root-to-leaf paths. The resul

What carries the argument

The load-bearing object is the decoding trie, the prefix tree that defines which token sequences count as valid items during constrained beam search; an item's term ID is exactly its root-to-leaf path. The argument runs through two mechanisms. First, the beam-survival factor phi_B(k) = min(1, B/k)—the probability that a beam of width B keeps the correct child when a node has k children—formalizes why early branching matters: each extra branch multiplies the chance of losing the target item. Second, the trie constructor is a recursive minimum set cover: at each node, BONSAI selects the smallest set of shared keywords that partitions the current item subset into children, greedily choosing the

Load-bearing premise

The load-bearing premise is that the measured gains come from the trie's low branching and variable depth rather than from the quality of the extracted keywords, since the full pipeline never holds the keyword set fixed while varying only the trie.

What would settle it

Take one dataset and one fixed set of extracted keywords; build the BONSAI minimum-cover trie and, as a control, a trie with the same keywords arranged in a fixed-depth, high-branching order; train the same LLM backbone on both with identical SFT and RL budgets. If Recall@5 differs by only a small fraction of the reported 16–21.6% gap, the claim that trie structure drives the gains is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Reordering existing keyword-based item IDs into a minimum-cover trie improves performance even when the keyword set is unchanged, so trie organization is itself a lever.
  • Items with richer metadata naturally get longer IDs and deeper paths, preserving information that fixed-length IDs discard.
  • Keeping branching low near the root makes constrained beam search more likely to keep the correct item in the beam.
  • The reinforcement-learning stage adds roughly 6–10% relative gains, mostly for items with long or multiple valid IDs, implying training should account for multiple valid root-to-leaf paths.
  • The design principles transfer to other term-ID methods, yielding 7.7–9.1% gains in the paper's reorganization experiment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If trie structure is the dominant factor, similar gains should appear for semantic-ID (non-textual) generative recommenders, whose tries currently come from codebook quantization rather than set cover; this is directly testable.
  • The minimum-set-cover view turns item ID construction into a combinatorial optimization problem; future work could use exact solvers at larger scale or learn cover selection, potentially beating the greedy approximation.
  • Feature filtering is a confound in the headline numbers; a controlled study with fixed features and varying tries would clarify how much of the 16–21.6% is due to structure alone.
  • Variable-depth tries also change cold-start behavior: items with sparse metadata can be attached near the root, which may explain the reported cold-start gains and suggests ID design can partially substitute for collaborative signal.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

6 major / 6 minor

Summary. The paper argues that the decoding trie used in constrained beam search for LLM-based generative recommendation is not merely an implementation detail but a first-class design object. It identifies two desirable properties—adaptive/variable ID length and constrained branching, especially near the root—provides a formal dynamic-programming framework and greedy minimum-set-cover construction (BONSAI), and reports large gains over existing baselines on three Amazon subsets. The central empirical claim is that co-designing textual term IDs and the decoding trie via minimum set cover yields up to 21.6% relative improvement over the strongest baseline, and that this improvement is primarily attributable to trie structure rather than to the quality of the underlying extracted features.

Significance. If the attribution claim were fully established, the paper would make a useful and somewhat novel contribution: it shifts the focus of term-ID design from semantic content alone to the interaction between term IDs and constrained decoding, and it provides a practical construction algorithm plus a formalization of the design space. The transfer experiment (Table 2) is a valuable idea, and the release of code is a strength. However, the headline causal claim is under-supported: the only experiment that isolates trie reorganization from feature quality explains roughly 8–9% of the observed 16–22% gain, and the theoretical results are largely definitional or assume what they set out to prove. The paper is promising but needs additional controlled experiments and more careful statistical reporting before the central claim can be accepted.

major comments (6)
  1. [§5.1 and Table 2] The paper's central attribution claim is not established. Table 1 shows BONSAI improves over GRLM by 16.4–21.6%, but Table 2, which reorganizes GRLM's original keywords using the minimum-cover strategy while keeping the feature set fixed, yields only 7.7–9.1% relative gains. Thus at least half of the headline gain is unaccounted for by trie reorganization. Section 5.1 states that using extracted rather than enriched terms 'helps isolate the source of BONSAI's gains: they arise primarily from a better-structured decoding trie,' but no experiment holds the feature set fixed while varying only trie construction in the full BONSAI pipeline. The feature extractor is Qwen3-32B while the backbone is Qwen3-1.7B, so the filter plausibly injects substantial extra signal. A controlled comparison is needed: e.g., BONSAI with GRLM-style keywords, or GRLM with BONSAI's extracted features, over the ful
  2. [Table 1] The statistical claim is under-reported. The table marks BONSAI with '*' and states 't-test with p<0.05' against the best baseline, but no variances, seeds, run counts, or effect-size details are given. With no error bars it is impossible to verify that the improvements are significant, especially for the smaller gaps (e.g., Sports R@10: 0.0524 vs 0.0449). The paper should report means and standard deviations over multiple seeds (or at least confidence intervals), specify the number of runs, and clarify whether the test is paired and how multiple comparisons across metrics are handled.
  3. [§6.3, Figures 7–8] The ablations do not cleanly isolate the two proposed properties. In Figure 7 the branching-factor comparison fixes depth at 4 and varies branching from 3 to 7, but the datasets have 12k+ items, so such shallow tries cannot uniquely identify items without appending unique suffixes at the final layer. This conflates the effect of branching with the effect of a synthetic suffix-based disambiguation mechanism that is not used in the original BONSAI trie. Similarly, Figure 8 compares the original variable-depth trie with truncated shallower structures, but the truncation inevitably changes the representation of all deeper items; the text does not specify how collisions are resolved in the truncated variants. These results are consistent with the paper's hypothesis but do not prove that branching and depth per se drive the gains.
  4. [§4, Theorems 1–3] The theoretical analysis is weaker than presented. Theorem 1 is the Bellman equation for the paper's own objective, so it does not independently establish that the proposed greedy construction is globally good. Theorem 2's approximation bound contains an unspecified problem-dependent constant ρ and an α_U that is essentially a tautological ratio; the bound 1/(ρ α_U) could be arbitrarily small and is not instantiated empirically. Theorem 3's second part relies on Assumption 3, which already asserts that forced extensions multiply success probability by at most η<1—i.e., the variable-depth conclusion is built into the assumption. The paper should either present these results as a formalization of the objective and search space, or supply conditions under which the greedy construction has a meaningful, non-vacuous approximation guarantee with respect to the true beam-search objective.
  5. [§6.4, Table 2] The transfer experiment is presented as evidence that 'the decoding trie is a major performance bottleneck,' but the table only reports percentage improvement and branching factors, not the absolute Recall/NDCG values for the O and M conditions. The text says the reorganization 'leads to consistent performance gains over the original results reported in Table 1,' yet the 7.7–9.1% numbers in Table 2 should be compared against the particular baselines in Table 1 with the same experimental setting. Please include the actual metric values for the reorganized GRLM tries.
  6. [§6.2, Figure 6] The RL contribution analysis in Figure 6 reports percentages without absolute values or error bars. The claim that RL gains are concentrated in long/multi-representation items is interesting, but the overlap between 'Top-50% Long Repr' and 'Multi-Repr' is acknowledged; a clearer separation of the two effects would strengthen the conclusion that RL helps specifically because of variable-length, multi-path term IDs.
minor comments (6)
  1. [§6.4, Table 2] The table caption uses 'O' and 'M' without defining them in the caption; the definitions appear only in the text. Please define them in the caption.
  2. [Appendix C, Table 5] The 'Title-10 token' row is not described in the text. Please explain how this representation is constructed and why it is included.
  3. [Appendix A] The feature-filtering prompt section contains a stray sentence: 'Let ⟨r_j⟩ denote the trie path used to represent item i_j.' This appears to be a leftover from the sequential recommendation prompt and should be removed or moved.
  4. [§6.3] The sentence 'Performance generally improves as the trie becomes deeper, but the original trie consistently outperforms all truncated variants' would be easier to evaluate if the y-axis in Figure 8 explicitly showed the 'Original' value and if the truncated depths were described in the text rather than only in the caption.
  5. [§5.2 and Algorithm 1] The algorithm handles the case where the set cover is empty (Line 5), but it is unclear what happens when an item is not covered by any shared feature and also lacks a unique feature. The text says 'every item also has its own unique feature' in Algorithm 1, but the feature-filtering stage does not guarantee this. Please clarify the handling of items with empty feature sets.
  6. [References] Several references are arXiv-only preprints from 2026; please check the final version for any published venue updates, particularly [53] and [41], which are heavily used as baselines.

Circularity Check

1 steps flagged

Variable-depth theorem is definitional (assumption restated), but benchmark improvements are empirically self-contained.

specific steps
  1. self definitional [Section 4, Theorem 3; Appendix B.1, Assumption 3; Appendix B.4 proof]
    "Theorem 3 ... if a feasible tree T has leaves of different depths, then any equal-depth tree obtained by extending the shallower leaves of T satisfies R(eT)<R(T). ... Assumption 3 (Nontrivial extension): Whenever a shallower leaf is extended only to match the depth of a deeper leaf, at least one added level has branching factor k>B. Equivalently, there exists η<1 such that every such forced extension multiplies the corresponding leaf success probability by at most η."

    The theorem's strict conclusion R(eT)<R(T) is just Assumption 3's 'every such forced extension multiplies the corresponding leaf success probability by at most η' with η<1, restated at the level of the tree objective. The non-strict half is immediate from set inclusion Teq ⊆ T, so it only says the optimum over all trees is at least as good as the optimum over equal-depth trees. Thus the paper's 'theoretical validation' of adaptive ID length assumes the harmful-extension conclusion rather than deriving it from beam-search mechanics; the variable-depth property is built into the assumption.

full rationale

The empirical core of BONSAI is not circular. No parameter is fitted to the test labels; headline gains (16.4–21.6% over GRLM) are measured on held-out data, and Table 2 reorganizes GRLM's own keywords with the minimum-cover ordering and reports 7.7–9.1% gains, so the trie-reorganization effect has independent evidence. The self-references [20], [23], and [30] are background/related-work citations and are not used as a uniqueness theorem or as the justification for the algorithm; they do not make the argument circular. The feature-filter confound (Qwen3-32B extractor vs. Qwen3-1.7B backbone) is a real external-validity concern for the claim that gains come 'primarily' from the trie, but it is an unsupported-attribution issue rather than a derivation that reduces to its inputs, so I do not count it as circularity here. The one definitional step is Theorem 3: its variable-depth-preference conclusion is either set inclusion or a restatement of Assumption 3, making the theoretical guarantee for property (1) tautological. The Bellman recursion in Theorem 1 is also essentially the definition of the paper's own optimal-value function, but as a standard optimality identity it is not separately scored. Because the benchmark improvement claims remain empirically self-contained, overall circularity is modest.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The theoretical guarantees in the paper depend on four unverified assumptions, two of which (Assumptions 1 and the uniform phi model) are contradicted by measurements reported in the same paper. The method itself introduces no new physical entities; the free parameters are design choices in the trie-construction pipeline and the RL training stage. The most honesty-relevant observation is that the paper does not control for the feature-extraction model, which is a significant source of potential confound.

free parameters (4)
  • LLM feature-extraction prompt and model (Qwen3-32B) = N/A (not a learned scalar)
    The set of keywords per item is produced by an open-ended LLM prompt; the quality of this extraction is a design choice that heavily influences trie structure and downstream performance, yet it is not controlled against baselines.
  • Per-category partitioning before trie construction = N/A
    Items are first partitioned by metadata category (§6 Implementation Details), which fixes the root-level structure and prevents cross-category feature sharing; this choice is made by hand.
  • Collision-resolution heuristic (append first word of title) = N/A
    Used to disambiguate items with identical IDs (§6 Implementation Details); this changes path lengths and deep-level branching and is chosen by hand.
  • RL rollout beam size (10) and RL epoch count (1) = 10, 1
    Training hyperparameters in the GRPO stage (§6 Implementation Details) that affect final performance; no sensitivity analysis is reported.
axioms (4)
  • domain assumption Assumption 1 (Beam-limited regime): every feasible split P satisfies |P| >= B
    Used in the proof of Theorem 2 to write phi_B(|P|)=B/|P|. Table 2 of the paper itself reports branching factors 1.02–4.55 at deeper levels with B=10, so this assumption is false for the constructed tries; the guarantee does not apply.
  • domain assumption Assumption 2 (Bounded continuation heterogeneity): W(U,P) <= rho W(U,Q) for all feasible splits P,Q
    Required for the greedy set-cover approximation bound in Theorem 2. This is a strong Lipschitz-style condition on the continuation values of every possible split; the paper provides no empirical check.
  • domain assumption Assumption 3 (Nontrivial extension): extending a shallow leaf to match a deeper leaf adds at least one level with branching factor k > B
    Used in Theorem 3 to claim fixed-depth extension strictly hurts. Whether this holds for the actual trie is not verified; if the added levels have branching <=B, extending is harmless.
  • domain assumption Beam survival factor model phi_B(k)=min(1,B/k)
    The theoretical objective R(T) assumes that the probability the beam keeps the correct path at a node with k children is B/k. With the realized Level-1 branching factors of 375–508, this gives per-step survival probabilities below 3%, predicting a success rate an order of magnitude smaller than the measured Recall@5 (§6.1), so the model is not descriptive of the actual LLM decoding behavior.

pith-pipeline@v1.3.0-alltime-deepseek · 20028 in / 20298 out tokens · 179850 ms · 2026-08-01T20:23:08.917713+00:00 · methodology

0 comments
read the original abstract

Generative recommendation (GR) is an increasingly popular paradigm in recommender systems, with a prominent line of work using LLMs as autoregressive backbones to predict the next item's term IDs (e.g., titles or keywords). The success of autoregressive generation hinges on constrained beam search over a decoding trie to ensure that generated outputs correspond to valid items. However, current research predominantly focuses on generating more comprehensive term IDs to describe items, while largely neglecting the structural design of the decoding trie formed by these terms. This can lead to a trie that is poorly suited to beam search, which degrades performance. To address this, we examine the effectiveness of term IDs from the perspective of decoding trie optimization. Through empirical and theoretical analyses, we identify two desirable properties for a highly performant trie: (1) adaptive and variable ID length, enabling items with varying semantic richness to be represented by IDs of appropriate lengths, and (2) constrained branching factors, especially at shallow levels, which drastically improves the success rate of constrained beam search. Motivated by these properties, we introduce BONSAI: Branching-Optimized Node Structure for Adaptive Identifiers, a novel framework that co-designs textual term IDs and their underlying decoding trie. BONSAI extracts recommendation-informative words from item metadata and employs a minimum set cover formulation to recursively build a trie that satisfies the above properties. Experiments reveal that BONSAI achieves up to a 21.6% relative improvement over state-of-the-art baselines. Further analyses confirm the crucial role of our proposed properties, and demonstrate their generalizability to be applied to enhance the performance of other term ID methods.

Figures

Figures reproduced from arXiv: 2607.16633 by Hanbing Wang, Jiliang Tang, Jingzhe Liu, Liam Collins, Mingxuan Ju, Neil Shah, Tong Zhao.

Figure 1
Figure 1. Figure 1: A segment of the trie for variable-length term IDs. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The length distribution of item metadata in Amazon [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of BONSAI and its four main components: (a) feature filtering, which extracts recommendation-relevant [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The improvements from RL over SFT. Most improve [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Impacts of varying branching factor. The perfor [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Impacts of trie depth. We observe consistent gaps [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Experiments on model and data scaling. BONSAI [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 23 linked inside Pith

  1. [1]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Xinyue Huo, Chong Chen, and Fuli Feng

  2. [2]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He

  3. [3]

    2012.Dynamic programming and optimal control: Volume I

    Dimitri Bertsekas. 2012.Dynamic programming and optimal control: Volume I. Vol. 4. Athena scientific

    2012

  4. [4]

    Runjin Chen, Mingxuan Ju, Ngoc Bui, Dimosthenis Antypas, Stanley Cai, Xi- aopeng Wu, Leonardo Neves, Zhangyang Wang, Neil Shah, and Tong Zhao

  5. [5]

    Weixin Chen, Yuhan Zhao, Jingyuan Huang, Zihe Ye, Clark Mingxuan Ju, Tong Zhao, Neil Shah, Li Chen, and Yongfeng Zhang. 2026. MemRec: Col- laborative Memory-Augmented Agentic Recommender System.arXiv preprint arXiv:2601.08816(2026)

    Pith/arXiv arXiv 2026

  6. [6]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. InProceedings of the 10th ACM conference on recommender systems. 191–198

    2016

  7. [7]

    Enhancing item tokenization for generative recommendation through self-improvement.arXiv preprint arXiv:2412.17171(2024)

    Pith/arXiv arXiv 2024

  8. [8]

    Qin Ding, Kevin Course, Linjian Ma, Jianhui Sun, Rouchen Liu, Zhao Zhu, Chunx- ing Yin, Wei Li, Dai Li, Yu Shi, et al. 2026. Bending the Scaling Law Curve in Large-Scale Recommendation Systems.arXiv preprint arXiv:2602.16986(2026)

    arXiv 2026

  9. [9]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5). InRecSys

    2022

  10. [10]

    Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment.arXiv preprint arXiv:2502.18965 (2025)

    Pith/arXiv arXiv 2025

  11. [11]

    Ruining He, Lukasz Heldt, Lichan Hong, Raghunandan Keshavan, Shifan Mao, Nikhil Mehta, Zhengyang Su, Alicia Tsai, Yueqi Wang, Shao-Chuan Wang, et al

  12. [12]

    Minjie Hong, Zetong Zhou, Zirun Guo, Ziang Zhang, Ruofan Hu, Weinan Gan, Jieming Zhu, and Zhou Zhao. 2025. Generative Reasoning Recommendation via LLMs.arXiv preprint arXiv:2510.20815(2025)

    arXiv 2025

  13. [13]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. 2025. Deepseek-r1: Incen- tivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025)

    Pith/arXiv arXiv 2025

  14. [14]

    Peiyu Hu, Wayne Lu, and Jia Wang. 2026. From ids to semantics: A generative framework for cross-domain recommendation with adaptive semantic tokeniza- tion. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 14874–14882

    2026

  15. [15]

    Wenyue Hua, Shuyuan Xu, Yingqiang Ge, and Yongfeng Zhang. 2023. How to Index Item IDs for Recommendation Foundation Models. InSIGIR-AP, Qingyao Ai, Yiqin Liu, Alistair Moffat, Xuanjing Huang, Tetsuya Sakai, and Justin Zobel (Eds.)

    2023

  16. [16]

    Jui-Ting Huang, Ashish Sharma, Shuying Sun, Li Xia, David Zhang, Philip Pronin, Janani Padmanabhan, Giuseppe Ottaviano, and Linjun Yang. 2020. Embedding- based retrieval in facebook search. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2553–2561

    2020

  17. [17]

    Yupeng Hou, Jianmo Ni, Zhankui He, Noveen Sachdeva, Wang-Cheng Kang, Ed H Chi, Julian McAuley, and Derek Zhiyuan Cheng. 2025. ActionPiece: Contextually Tokenizing Action Sequences for Generative Recommendation.arXiv preprint arXiv:2502.13581(2025)

    Pith/arXiv arXiv 2025

  18. [18]

    Jianchao Ji, Zelong Li, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Juntao Tan, and Yongfeng Zhang. 2024. Genrec: Large language model for generative recommen- dation. InEuropean Conference on Information Retrieval. Springer, 494–502

    2024

  19. [19]

    Wei Jin, Haitao Mao, Zheng Li, Haoming Jiang, Chen Luo, Hongzhi Wen, Haoyu Han, Hanqing Lu, Zhengyang Wang, Ruirui Li, et al . 2023. Amazon-m2: A multilingual multi-locale shopping session dataset for recommendation and text generation.Advances in Neural Information Processing Systems36 (2023), 8006– 8026

    2023

  20. [20]

    Clark Mingxuan Ju, Liam Collins, Leonardo Neves, Bhuvesh Kumar, Louis Yufeng Wang, Tong Zhao, and Neil Shah. 2025. Generative Recommendation with Semantic IDs: A Practitioner’s Handbook.arXiv preprint arXiv:2507.22224(2025)

    Pith/arXiv arXiv 2025

  21. [21]

    Dietmar Jannach and Malte Ludewig. 2017. When Recurrent Neural Networks meet the Neighborhood for Session-Based Recommendation. InRecSys, Paolo Cremonesi, Francesco Ricci, Shlomo Berkovsky, and Alexander Tuzhilin (Eds.). ACM

    2017

  22. [22]

    Clark Mingxuan Ju, Leonardo Neves, Bhuvesh Kumar, Liam Collins, Tong Zhao, Yuwei Qiu, Qing Dou, Yang Zhou, Sohail Nizam, Rengim Ozturk, et al . 2025. Learning Universal User Representations Leveraging Cross-domain User Intent at Snapchat.arXiv preprint arXiv:2504.21838(2025)

    Pith/arXiv arXiv 2025

  23. [23]

    Clark Mingxuan Ju, Tong Zhao, Leonardo Neves, Liam Collins, Bhuvesh Kumar, Jiwen Ren, Lili Zhang, Wenfeng Zhuo, Vincent Zhang, Xiao Bai, et al. 2026. Seman- tic IDs for Recommender Systems at Snapchat: Use Cases, Technical Challenges, and Design Choices.arXiv preprint arXiv:2604.03949(2026)

    Pith/arXiv arXiv 2026

  24. [24]

    Mingxuan Ju, William Shiao, Zhichun Guo, Yanfang Ye, Yozen Liu, Neil Shah, and Tong Zhao. 2024. How does message passing improve collaborative filtering? Advances in Neural Information Processing Systems37 (2024), 8760–8784

    2024

  25. [25]

    Clark Mingxuan Ju, Leonardo Neves, Bhuvesh Kumar, Liam Collins, Tong Zhao, Yuwei Qiu, Qing Dou, Sohail Nizam, Sen Yang, and Neil Shah. 2025. Revisiting Self-attention for Cross-domain Sequential Recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 1094–1105

    2025

  26. [26]

    Xiaoyu Kong, Leheng Sheng, Junfei Tan, Yuxin Chen, Jiancan Wu, An Zhang, Xiang Wang, and Xiangnan He. 2025. Minionerec: An open-source framework for scaling generative recommendation.arXiv preprint arXiv:2510.24431(2025)

    arXiv 2025

  27. [27]

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. 2022. Autoregressive image generation using residual quantization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11523–11532

    2022

  28. [28]

    Lei Li, Yongfeng Zhang, Dugang Liu, and Li Chen. 2023. Large Language Models for Generative Recommendation: A Survey and Visionary Discussions.CoRR abs/2309.01157 (2023)

    Pith/arXiv arXiv 2023

  29. [29]

    Wang-Cheng Kang and Julian J. McAuley. 2018. Self-Attentive Sequential Rec- ommendation. InICDM

    2018

  30. [30]

    Jingzhe Liu, Liam Collins, Jiliang Tang, Tong Zhao, Neil Shah, and Clark Mingx- uan Ju. 2025. Understanding generative recommendation with semantic ids from a model-scaling view.arXiv preprint arXiv:2509.25522(2025)

    Pith/arXiv arXiv 2025

  31. [31]

    Qidong Liu, Xiangyu Zhao, Yejing Wang, Zijian Zhang, Howard Zhong, Chong Chen, Xiang Li, Wei Huang, and Feng Tian. 2025. Bridge the domains: Large language models enhanced cross-domain sequential recommendation. InProceed- ings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1582–1592

    2025

  32. [32]

    Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, et al

  33. [33]

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. LLaRA: Large Language-Recommendation Assistant. In SIGIR

    2024

  34. [34]

    Laurent Perron, Frédéric Didier, and Steven Gay. 2023. The CP-SAT-LP solver (in- vited talk). In29th International Conference on Principles and Practice of Constraint Programming (CP 2023). Schloss Dagstuhl–Leibniz-Zentrum für Informatik, 3–1

    2023

  35. [35]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Mahesh Sathiamoorthy. 2023. Recommender Systems with Generative Retrieval. InNeurIPS, Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine (Eds.)

    2023

  36. [36]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme

  37. [37]

    Onerec-think: In-text reasoning for generative recommendation.arXiv preprint arXiv:2510.11639(2025)

    arXiv 2025

  38. [38]

    OpenAI. 2023. GPT-4 Technical Report.CoRR(2023)

    2023

  39. [39]

    Junfei Tan, Yuxin Chen, An Zhang, Junguang Jiang, Bin Liu, Ziru Xu, Han Zhu, Jian Xu, Bo Zheng, and Xiang Wang. 2025. Reinforced preference optimization for recommendation.arXiv preprint arXiv:2510.12211(2025)

    arXiv 2025

  40. [40]

    Juntao Tan, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Zelong Li, and Yongfeng Zhang. 2024. Idgenrec: Llm-recsys alignment with textual id learning. InProceed- ings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 355–364

    2024

  41. [41]

    Zhouhang Xie, Bo Peng, Zhankui He, Ziqi Chen, Alice Han, Isabella Ye, Ben- jamin Coleman, Noveen Sachdeva, Fernando Pereira, Julian McAuley, et al. 2026. AgenticTagger: Structured Item Representation for Recommendation with LLM Agents.arXiv preprint arXiv:2602.05945(2026)

    arXiv 2026

  42. [42]

    Wujiang Xu, Yunxiao Shi, Zujie Liang, Xuying Ning, Kai Mei, Kun Wang, Xi Zhu, Min Xu, and Yongfeng Zhang. 2025. iagent: Llm agent as a shield between user and recommender systems. InFindings of the Association for Computational Linguistics: ACL 2025. 18056–18084

    2025

  43. [43]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

    Pith/arXiv arXiv 2024

  44. [44]

    Zhengyang Su, Isay Katsman, Yueqi Wang, Ruining He, Lukasz Heldt, Raghu- nandan Keshavan, Shao-Chuan Wang, Xinyang Yi, Mingyan Gao, Onkar Dalal, et al. 2026. Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators.arXiv preprint arXiv:2602.22647(2026)

    Pith/arXiv arXiv 2026

  45. [45]

    Xinyang Yi, Ji Yang, Lichan Hong, Derek Zhiyuan Cheng, Lukasz Heldt, Aditee Kumthekar, Zhe Zhao, Li Wei, and Ed Chi. 2019. Sampling-bias-corrected neural modeling for large corpus item recommendations. InProceedings of the 13th ACM RecSys ’26, September 27-October 02, 2026, Minneapolis, MN, USA Jingzhe Liu et al. conference on recommender systems. 269–277

    2019

  46. [46]

    Zhenrui Yue, Sara Rabhi, Gabriel de Souza Pereira Moreira, Dong Wang, and Even Oldridge. 2023. Llamarec: Two-stage recommendation using large language models for ranking.arXiv preprint arXiv:2311.02089(2023)

    Pith/arXiv arXiv 2023

  47. [47]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhao- jie Gong, Fangda Gu, Michael He, et al. 2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations.arXiv preprint arXiv:2402.17152(2024)

    Pith/arXiv arXiv 2024

  48. [48]

    An Zhang, Leheng Sheng, Yuxin Chen, Hao Li, Yang Deng, Xiang Wang, and Tat-Seng Chua. 2023. On Generative Agents in Recommendation.CoRR abs/2310.10108 (2023)

    Pith/arXiv arXiv 2023

  49. [49]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    Pith/arXiv arXiv 2025

  50. [50]

    Ji Yang, Xinyang Yi, Derek Zhiyuan Cheng, Lichan Hong, Yang Li, Simon Xiaom- ing Wang, Taibai Xu, and Ed H Chi. 2020. Mixed negative sampling for learning two-tower neural networks in recommendations. InCompanion proceedings of the web conference 2020. 441–447

    2020

  51. [51]

    Luankang Zhang, Yonghao Huang, Hang Lv, Mingjia Yin, Liangyue Li, Zulong Chen, Hao Wang, and Enhong Chen. 2026. Why Thinking Hurts? Diagnosing and Rectifying the Reasoning Shift in Foundation Recommender Models.arXiv preprint arXiv:2602.16587(2026)

    Pith/arXiv arXiv 2026

  52. [52]

    Taiyan Zhang, Hongtao Wang, Yunqian Fan, Kunda Yang, Jichuan Zeng, and Renchi Yang. [n. d.]. A Survey of Item Identifiers in Generative Recommendation: Construction, Alignment, and Generation. ([n. d.])

  53. [53]

    Zhiyang Zhang, Junda She, Kuo Cai, Bo Chen, Shiyao Wang, Xinchen Luo, Qiang Luo, Ruiming Tang, Han Li, Kun Gai, et al. 2026. Unleashing the Native Recom- mendation Potential: LLM-Based Generative Recommendation via Structured Term Identifiers.arXiv preprint arXiv:2601.06798(2026)

    arXiv 2026

  54. [54]

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2023. Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation.CoRRabs/2311.09049 (2023)

    Pith/arXiv arXiv 2023

  55. [55]

    McAuley, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen

    Junjie Zhang, Yupeng Hou, Ruobing Xie, Wenqi Sun, Julian J. McAuley, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2023. AgentCF: Collaborative Learning with Autonomous Language Agents for Recommender Systems.CoRRabs/2310.09233 (2023)

    Pith/arXiv arXiv 2023

  56. [56]

    Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2023. Recommendation as Instruction Following: A Large Language Model Empowered Recommendation Approach.CoRRabs/2305.07001 (2023)

    Pith/arXiv arXiv 2023

  57. [61]

    Guorui Zhou, Honghui Bao, Jiaming Huang, Jiaxin Deng, Jinghao Zhang, Junda She, Kuo Cai, Lejian Ren, Lu Ren, Qiang Luo, et al. 2025. OpenOneRec Technical Report.arXiv preprint arXiv:2512.24762(2025)

    arXiv 2025

  58. [62]

    Guorui Zhou, Hengrui Hu, Hongtao Cheng, Huanjie Wang, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Lu Ren, Liao Yu, et al. 2025. Onerec-v2 technical report.arXiv preprint arXiv:2508.20900(2025). A LLM Prompt Samples LLM Prompts for Feature Filtering You are an expert product keywords extractor. Your task is to extract informative keywords from this pr...

    Pith/arXiv arXiv 2025

  59. [2012]

    BPR: Bayesian Personalized Ranking from Implicit Feedback.CoRR abs/1205.2618 (2012)

    Pith/arXiv arXiv 2012

  60. [2023]

    InRecSys

    TALLRec: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation. InRecSys

  61. [2024]

    Optimizing Textual Decoding Tries in LLM-based Generative Recommendation RecSys ’26, September 27-October 02, 2026, Minneapolis, MN, USA

    Decoding matters: Addressing amplification bias and homogeneity issue for llm-based recommendation.arXiv preprint arXiv:2406.14900(2024). Optimizing Textual Decoding Tries in LLM-based Generative Recommendation RecSys ’26, September 27-October 02, 2026, Minneapolis, MN, USA

    Pith/arXiv arXiv 2024

  62. [2025]

    Plum: Adapting pre-trained language models for industrial-scale generative recommendations.arXiv preprint arXiv:2510.07784(2025)

    arXiv 2025