Pith. sign in

REVIEW 4 major objections 5 minor 72 references

Disentangling LLM item states into geographic and semantic slots before quantization yields interpretable Semantic IDs that rank better and collide far less for local-life recommendation.

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 · grok-4.5

2026-07-31 22:37 UTC pith:2OXA52RE

load-bearing objection Solid industrial SID pipeline with real collision wins; the causal story that disentanglement (not dual-stream RQ + extra labels) drives AUC is still under-shown. the 4 major comments →

arxiv 2607.27944 v1 pith:2OXA52RE submitted 2026-07-30 cs.IR cs.AI

Interpretable Representation via LLM-Driven Generative Disentanglement for Local-Life Service Recommendation

Long Zhang , Hao Jiang , Sheng Yu , Fei Pan , Peng Jiang , Kun Gai This is my paper
classification cs.IR cs.AI
keywords Local Services RecommendationLarge Language ModelSemantic IDGenerative DisentanglementGeographic hierarchyResidual QuantizationInterpretable representation
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.

Local-life recommenders need item identifiers that carry both where a place is and what it sells, but most Semantic ID methods squash geography, brand, and category into one vector and then quantize it. That mix loses fine location distinctions, makes codes hard to read, and causes many different places to share the same ID. This paper argues the fix is to change the order of work: jointly encode the item text with an LLM so geo and content can still interact, route the hidden states into fixed geographic and semantic slots, train those slots so each one can be decoded into its attribute and still separate similar items, and only then quantize the two streams separately. On industrial Kuaishou data and public Foursquare data the resulting codes improve ranking AUC across many backbones (up to about 5.4% relative), recover coarse location labels at over 99% accuracy, and cut full-ID collisions from roughly 97% to about 40%. A sympathetic reader cares because local services are reachability-first: if the ID itself is grounded in hierarchy and brand, retrieval and diagnosis stop fighting an opaque collision soup.

Core claim

The paper’s central claim is that Semantic IDs for local-life recommendation should be built by generative disentanglement before quantization—an encode → disentangle → align → quantize pipeline—rather than by quantizing a single mixed item representation. Joint LLM encoding keeps geo–content dependencies; attribute-aligned slots plus progressive decoding and structure-aware contrast make those slots verifiable and discriminative; dual-stream residual quantization then turns them into compact codes whose positions correspond to geography versus semantics, improving ranking and sharply reducing collisions.

What carries the argument

LGRID’s pipeline, especially the Structured Disentangled Block (semantic anchors, anchor-guided cross-attention, and structured causal routing that keeps geo and semantic slot groups separate) plus Synergistic Alignment Learning (progressive generative decoding of known attributes and hierarchical hard-negative contrast) and Dual-Stream Residual Quantization.

Load-bearing premise

That a fixed set of hierarchy-aligned slots trained to decode known admin, brand, and category labels really captures the geo–content factors that matter for ranking, rather than throwing away interactions that do not fit those labeled fields.

What would settle it

Train the same dual-stream quantizer on slots that are not supervised by progressive attribute decoding (or that use scrambled slot-to-field assignments) and check whether coarse geo decoding stays near 99%, full-SID collision stays near 40%, and the AUC lift over strong SID baselines on Kuaishou disappears.

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

If this is right

  • SID positions can be inspected and debugged as province/city/district/town versus brand/category instead of opaque token indices.
  • Fewer full-SID collisions mean more distinct nearby merchants keep distinct codes under the same brand.
  • The same precomputed SIDs plug into many ranking backbones without architecture changes and still lift AUC.
  • Coarse geographic structure in the codes emerges with relatively little POI text; fine town-level structure needs much more labeled coverage.
  • Quantizing separated geo and semantic streams avoids codebook competition that single-stream residual quantizers show on mixed vectors.

Where Pith is reading between the lines

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

  • Any domain where items have a hard hierarchy plus a soft content type (not only LBS) may benefit from joint encode-then-slot before quantization rather than field-wise separate encoders.
  • If production catalogs lack clean town/brand labels, the progressive decoding objective would need weaker or self-supervised targets or the collision and interpretability gains may shrink.
  • Serving systems could expose slot-level controls (e.g., freeze geo codes, refresh semantic codes) because the streams are already separated offline.

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

4 major / 5 minor

Summary. The paper argues that Semantic ID (SID) generation for local-life / LBS recommendation is limited by a single-representation-then-quantization paradigm that entangles geography with brand/category and yields black-box, collision-prone codes. It proposes LGRID: an encode→disentangle→align→quantize pipeline that jointly encodes POI text with an LLM, routes hidden states into hierarchy-aligned geographic and semantic slots via a Structured Disentangled Block (SAI, AGP, SCR with inter-group mask), trains those slots with Progressive Generative Disentanglement, Structure-Aware Contrastive Learning, and diversity regularization, then applies Dual-Stream Residual Quantization. On Kuaishou and Foursquare, LGRID reports consistent AUC gains across many recommenders (up to 5.44% relative on Kuaishou), >99% coarse geographic decoding accuracy, improved attribute-consistent retrieval, and a full-SID collision rate of 39.9% versus ~97% for LGSID, with only modest offline construction overhead.

Significance. If the causal story holds, the work is a solid systems contribution to industrial SID design for LBS: it reframes the problem as representation construction before quantization, couples joint geo–content encoding with explicit attribute structure, and backs the claim with multi-backbone ranking, collision/utilization, decoding, retrieval, and efficiency evidence on a large industrial corpus plus a public benchmark. The collision drop and hierarchical SID-prefix / t-SNE analyses are particularly useful for practitioners who need diagnosable codes. Strengths include broad backbone coverage (Table 1), joint-vs-field-wise encoding (Table 4), codebook utilization analysis (Table 5), and detailed appendices on prompts, masks, and negatives. The main open question is whether the named disentanglement machinery—not extra attribute supervision or dual-stream capacity alone—drives ranking and collision gains.

major comments (4)
  1. [§4.4, Tables 6–7; cf. Table 1, Table 5] Central claim under-identified on ranking/collision: Tables 6–7 and Fig. 4 ablate SCR, PGD, SACL, LDR, and anchors only on attribute decoding and attribute-consistent retrieval, not on downstream AUC (Table 1) or full-SID collision/utilization (Table 5). The paper credits encode→disentangle→align→quantize for up to +5.44% AUC and 39.9% vs 97% collision, but without AUC/collision for w/o-SCR, w/o-PGD, w/o-SACL, and especially a fused-representation + dual-stream RQ control, it remains unclear whether gains come from the SCR mask (Eq. 3), curriculum PGD (§3.5.1), or from (i) structured admin/brand/category supervision unavailable to content-only SID baselines in the same form and/or (ii) separated codebooks in DSRQ (§3.6). Please report recommendation AUC and collision for these ablated SID variants on at least one strong backbone (e.g., ETA/TWIN) and one dataset.
  2. [§3.5.1, Table 3, §4.2] Supervision asymmetry vs SID baselines needs explicit control. PGD trains slots against oracle geographic and semantic attribute targets (Eq. 6; Appendix B.3), so >99% coarse-geo decoding (Table 3) is largely on-distribution and expected. LGSID/RQ-VAE-style baselines are not shown to receive equivalent structured field labels as generative targets. A fair control would give a strong single-stream SID baseline the same attribute labels (e.g., multi-task attribute prediction or field-conditioned quantization) and/or strip LGRID of PGD labels while keeping dual-stream RQ. Without this, the paradigm-level claim over single-representation-then-quantization is only partly isolated from “more side information.”
  3. [Table 1, Appendix A.3] Effect sizes and significance on Foursquare are thin for several backbones (e.g., DIN +0.05%, SASRec +0.06% in Table 1), while Kuaishou gains are larger but uneven. The manuscript should report variance over seeds or bootstrap CIs for AUC, and clarify whether candidate construction (D=15 km on Kuaishou; full ranking on Foursquare, Appendix A.3) interacts with SID quality. Small absolute gains do not invalidate the method, but the abstract’s “consistently outperforms” framing should be tempered where improvements are near noise.
  4. [§3.4.3 Eq. (3), Table 4, §1] Inter-group independence in SCR (Eq. 3: no geo↔semantic attention) is an axiom of the design, yet the paper’s own motivation stresses that geography contextualizes semantics (e.g., “Apple” as retailer vs fruit). Table 4 shows joint encoding beats field-wise encoding, which is good, but does not test whether hard inter-group blocking discards residual interactions that ranking needs. A soft-mask or limited cross-group attention ablation with AUC would test whether the hard independence constraint is necessary or overly strong.
minor comments (5)
  1. [Abstract, Fig. 1, §3.2] Fig. 1 and the abstract use “Encode → Disentangle → Align → Quantize”; keep hyphenation and arrow notation consistent with the body (§3.2).
  2. [Table 1] Table 1 “Base [16]” cites Qwen3; clarify that Base is the recommender without SID (or with raw LLM features) so readers do not confuse it with an SID baseline.
  3. [Fig. 6] NMI values in Fig. 6 are compelling; state the clustering procedure (features used, number of clusters vs ground-truth labels) in the caption or Appendix C.
  4. [Abstract, Fig. 2, Algorithm 1] Typos/style: “single-representation-then-quantizationparadigm” missing space in abstract; “G FFN” in Fig. 2 is unexplained; Algorithm 1 step 3 “Align: optimize Z with …” is schematic—point to Eq. (12).
  5. [§2.2] Related work could briefly situate against other structured/disentangled recommendation embeddings beyond geo-SID (not only LGSID/OneLoc), to clarify novelty of training-time slot construction vs post-hoc probing.

Circularity Check

1 steps flagged

Empirical SID systems paper: no derivation-by-construction; mild on-objective decoding metric only.

specific steps
  1. fitted input called prediction [§3.5.1 PGD / Table 3 attribute decoding]
    "PGD makes each slot LLM-decodable through coarse-to-fine attribute supervision. ... LGRID recovers available coarse spatial attributes with over 99% accuracy on both datasets"

    Slots are trained with attribute-level cross-entropy on province/city/district/town/brand/category targets, then the same decoding task is reported as interpretability evidence. High coarse-geo accuracy is largely the training objective succeeding on held-out POIs, not an independent prediction forced from other premises. This is standard supervised eval and does not collapse AUC or collision claims; only a minor on-objective circularity of presentation.

full rationale

LGRID is an encode→disentangle→align→quantize engineering pipeline evaluated on external benchmarks (AUC vs. SID baselines on Kuaishou/Foursquare, full-SID collision/utilization, attribute-consistent retrieval). Those metrics are not algebraic restatements of the training losses or of fitted free parameters renamed as predictions. Progressive Generative Disentanglement (PGD) does supervise slots to decode known admin/brand/category labels, so Table 3’s high coarse-geo decoding accuracy is largely on-objective supervised evaluation rather than an independent discovery—but the paper presents it as slot-decodability/interpretability evidence, not as a first-principles forecast, and collision/AUC remain separate. Self-citations (e.g., LGSID, REST) supply baselines and problem framing from the same group; they are not invoked as uniqueness theorems that force the method. No self-definitional identity, fitted-constant-as-prediction chain, or renamed known law appears in the claimed contribution path. Score 1 only for the mild, expected overlap between PGD training and decoding accuracy reporting.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

Load-bearing content is methodological and empirical: standard transformer/RQ machinery plus domain choices (admin hierarchy slots, dual streams, loss weights, curated POI subset). No physical constants; free parameters are training/design knobs. Invented entities are architectural modules, not new ontological objects in nature.

free parameters (6)
  • Loss weights λ1/λ2/λ3 (PGD coarse/fine, SACL, LDR) = coarse 1.6; fine 1.8; SACL→0.66; LDR 0.05
    Hand-set weights (appendix: coarse 1.6, fine 1.8, contrastive warmup to 0.66, LDR 0.05) that shape the learned slots and thus SIDs.
  • Slot count and geo/semantic split (N=8, k=4) = 8 slots; 4 geo + 4 semantic
    Design choice mapping province/city/district/town vs brand/category/detail; central to claimed attribute correspondence.
  • DSRQ depth and codebook configuration (D=2 per stream) = D=2 per stream → 4-token SID
    Chooses 4-token SIDs; utilization and collision metrics depend on codebook sizes and residual depth.
  • InfoNCE temperature and hard-negative mining (K=6) = τ init 0.07, clip [0.05,0.20]; K=6
    Temperature clipped to [0.05,0.20]; structured negatives and fallbacks affect discriminability.
  • Curated 120k POI-text SID training subset = 120k train / 5k val POIs
    Balanced, cleaned subset of industrial POIs used to train disentanglement; selection can affect reported decoding and collision stats.
  • Offline spatial candidate radius D on Kuaishou = D=15 km
    Evaluation candidate set uses fixed 15 km threshold; ranking metrics depend on this protocol.
axioms (5)
  • domain assumption LLM hidden states over joint POI text encode extractable geo–content dependencies that survive attention-based routing into slots.
    Stated motivation in §1–3.3 citing probing/steering literature; required for Encode→Disentangle to beat field-wise encoding.
  • domain assumption Administrative hierarchy (province→city→district→town) plus brand/category/detail is the right attribute factorization for LBS SIDs.
    Hard-wired by SAI anchors, SCR groups, and PGD task schema (§3.4–3.5, Appendix B).
  • ad hoc to paper Inter-group independence mask (no geo↔semantic attention in SCR) improves SIDs more than it harms needed residual interactions.
    Design constraint in Eq. (3)/(16); ablations support it but it is a paper-specific inductive bias.
  • domain assumption Residual quantization of separated streams yields discrete codes suitable as additive features beside item IDs in standard rankers.
    Deployment protocol Eq. (14); common SID practice extended to dual streams.
  • standard math Standard optimization and attention machinery (AdamW, LoRA, cross-attention, InfoNCE, RQ commitment loss) behave as usual.
    Background ML tools used throughout §3 without new theory.
invented entities (3)
  • Structured Disentangled Block (SAI + AGP + SCR) no independent evidence
    purpose: Route entangled LLM states into attribute-aligned geographic and semantic slot tokens before quantization.
    Core architectural invention of the paper; evaluated via ablations, not an external physical entity.
  • Synergistic Alignment Learning (PGD + SACL + LDR) no independent evidence
    purpose: Make slots generatively decodable, discriminative, and non-redundant.
    Training objective bundle specific to LGRID’s slots.
  • Dual-Stream Residual Quantization (DSRQ) SIDs no independent evidence
    purpose: Separately discretize geo and semantic slots into a 4-token hierarchical SID with attribute correspondence.
    Output representation claimed to reduce collisions vs single-stream SIDs.

pith-pipeline@v1.2.0-daily-grok45 · 33110 in / 3911 out tokens · 75041 ms · 2026-07-31T22:37:06.204122+00:00 · methodology

0 comments
read the original abstract

While large language models (LLMs) have advanced ID-based recommendation through Semantic ID (SID) modeling, existing SID generation frameworks largely follow a single-representation-then-quantization paradigm. This design faces two bottlenecks: semantic entanglement mixes heterogeneous attributes, such as geography, brand, and category, causing information loss during quantization, low-quality SIDs, and severe collisions; moreover, black-box representation learning provides neither explicit attribute semantics nor clear geographic or semantic meanings for SID positions. These limitations weaken both retrieval reliability and the ability to diagnose or control SID generation. We propose Interpretable Representation via LLM-Driven Generative Disentanglement for Local-Life Service Recommendation (LGRID). LGRID introduces a generative disentanglement paradigm through an Encode -> Disentangle -> Align -> Quantize pipeline. It first uses joint LLM encoding to preserve cross-attribute geographic-semantic dependencies, rather than encoding fields independently. A Structured Disentangled Block then routes hidden states into attribute-aligned slots for geographic and semantic factors. Synergistic Alignment Learning makes these slots both generatively decodable and discriminative for retrieval, while Dual-Stream Residual Quantization separately discretizes the two streams into compact SIDs with explicit attribute correspondence. This design yields interpretable SIDs with positions grounded in item attributes and local-service semantics. Experiments on Kuaishou and Foursquare show that LGRID consistently outperforms strong SID baselines, achieving up to a 5.44 percent relative AUC gain. It also achieves over 99 percent attribute-decoding accuracy for coarse geographic fields and reduces the full-SID collision rate to 39.9 percent, compared with 97.0 percent for LGSID.

Figures

Figures reproduced from arXiv: 2607.27944 by Fei Pan, Hao Jiang, Kun Gai, Long Zhang, Peng Jiang, Sheng Yu.

Figure 1
Figure 1. Figure 1: Comparison between (a) the single-representation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the LGRID framework. The architecture integrates the SD Block to disentangle features, employs [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Scaling Law Analysis. Performance trajectories on [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Hierarchical city and district frequency distribution of LGRID for different SID prefixes [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: T-SNE visualization of items around cluster cen [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Brand- and category-level distributions of top-frequency SID prefixes. [PITH_FULL_IMAGE:figures/full_fig_p016_7.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

72 extracted references · 22 linked inside Pith

  1. [1]

    Spatiotemporal-enhanced network for click-through rate prediction in location-based services.arXiv preprint arXiv:2209.09427, 2022

    Shaochuan Lin, Yicong Yu, Xiyu Ji, Taotao Zhou, Hengxu He, Zisen Sang, Jia Jia, Guodong Cao, and Ning Hu. Spatiotemporal-enhanced network for click-through rate prediction in location-based services.arXiv preprint arXiv:2209.09427, 2022

    Pith/arXiv arXiv 2022

  2. [2]

    Spatial-temporal knowledge distillation for takeaway recommendation

    Shuyuan Zhao, Wei Chen, Boyan Shi, Liyong Zhou, Shuohao Lin, and Huaiyu Wan. Spatial-temporal knowledge distillation for takeaway recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 13365–13373, 2025

    2025

  3. [3]

    Rest: A plug-and-play spatially- constrained representation enhancement framework for local-life recommenda- tion

    Hao Jiang, Long Zhang, Guoquan Wang, Sheng Yu, Yang Zeng, Wencong Zeng, Fei Pan, Peng Jiang, and Guorui Zhou. Rest: A plug-and-play spatially- constrained representation enhancement framework for local-life recommenda- tion. InProceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 4682–4687, 2026

    2026

  4. [4]

    Localgpt: Bench- marking and advancing large language models for local life services in meituan

    Xiaochong Lan, Jie Feng, Jiahuan Lei, Xinlei Shi, and Yong Li. Localgpt: Bench- marking and advancing large language models for local life services in meituan. 2025

    2025

  5. [5]

    Fragment and integrate network (fin): A novel spatial- temporal modeling based on long sequential behavior for online food ordering click-through rate prediction

    Jun Li and Ge Zhang. Fragment and integrate network (fin): A novel spatial- temporal modeling based on long sequential behavior for online food ordering click-through rate prediction. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 4688–4694, 2023

    2023

  6. [6]

    Fim: Frequency-aware multi-view interest modeling for local-life service recommendation

    Guoquan Wang, Qiang Luo, Weisong Hu, Pengfei Yao, Wencong Zeng, Guorui Zhou, and Kun Gai. Fim: Frequency-aware multi-view interest modeling for local-life service recommendation. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1748–1757, 2025

    2025

  7. [7]

    Llm-aligned geographic item tokenization for local-life recommendation.arXiv preprint arXiv:2511.14221, 2025

    Hao Jiang, Guoquan Wang, Donglin Zhou, Sheng Yu, Yang Zeng, Wencong Zeng, Kun Gai, and Guorui Zhou. Llm-aligned geographic item tokenization for local-life recommendation.arXiv preprint arXiv:2511.14221, 2025

    arXiv 2025

  8. [8]

    Oneloc: Geo-aware generative recommender systems for local life service.arXiv preprint arXiv:2508.14646, 2025

    Zhipeng Wei, Kuo Cai, Junda She, Jie Chen, Minghao Chen, Yang Zeng, Qiang Luo, Wencong Zeng, Ruiming Tang, Kun Gai, et al. Oneloc: Geo-aware generative recommender systems for local life service.arXiv preprint arXiv:2508.14646, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Reasoning over space: Enabling geographic reasoning for llm-based generative next poi recommendation.arXiv preprint arXiv:2601.04562, 2026

    Dongyi Lv, Qiuyu Ding, Heng-Da Xu, Zhaoxu Sun, Zhi Wang, Feng Xiong, and Mu Xu. Reasoning over space: Enabling geographic reasoning for llm-based generative next poi recommendation.arXiv preprint arXiv:2601.04562, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    Deep interest evolution network for click-through rate prediction

    Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. Deep interest evolution network for click-through rate prediction. InProceedings of the AAAI conference on artificial intelligence, volume 33, pages 5941–5948, 2019

    2019

  11. [11]

    Dynamic forgetting and spatio- temporal periodic interest modeling for local-life service recommendation.arXiv preprint arXiv:2508.02451, 2025

    Zhaoyu Hu, Jianyang Wang, Hao Guo, Yuan Tian, Erpeng Xue, Xianyang Qi, Hongxiang Lin, Lei Wang, and Sheng Chen. Dynamic forgetting and spatio- temporal periodic interest modeling for local-life service recommendation.arXiv preprint arXiv:2508.02451, 2025

    arXiv 2025

  12. [12]

    Next-poi recommendation via spatial-temporal knowledge graph contrastive learning and trajectory prompt.IEEE Transactions on Knowl- edge and Data Engineering, 2025

    Wei Chen, Haoyu Huang, Zhiyu Zhang, Tianyi Wang, Youfang Lin, Liang Chang, and Huaiyu Wan. Next-poi recommendation via spatial-temporal knowledge graph contrastive learning and trajectory prompt.IEEE Transactions on Knowl- edge and Data Engineering, 2025

    2025

  13. [13]

    Next point-of-interest (poi) recommendation model based on multi-modal spatio-temporal context feature embedding.arXiv preprint arXiv:2509.22661, 2025

    Lingyu Zhang, Guobin Wu, Yan Wang, Pengfei Xu, Jian Liang, Xuan Song, and Yunhai Wang. Next point-of-interest (poi) recommendation model based on multi-modal spatio-temporal context feature embedding.arXiv preprint arXiv:2509.22661, 2025

    arXiv 2025

  14. [14]

    Integrating personalized spatio-temporal clustering for next poi recommendation

    Chao Song, Zheng Ren, and Li Lu. Integrating personalized spatio-temporal clustering for next poi recommendation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 12550–12558, 2025

    2025

  15. [15]

    Video corpus moment retrieval with query-specific context learning and progressive localization.IEEE Transactions on Circuits and Systems for Video Technology, 2025

    Long Zhang, Peipei Song, Zhangling Duan, Shuo Wang, Xiaojun Chang, and Xun Yang. Video corpus moment retrieval with query-specific context learning and progressive localization.IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  16. [16]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

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

    Pith/arXiv arXiv 2025

  17. [17]

    Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Recommender systems with generative retrieval.Advances in Neural Information Processing Systems, 36:10299–10315, 2023

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al. Recommender systems with generative retrieval.Advances in Neural Information Processing Systems, 36:10299–10315, 2023

    2023

  19. [19]

    Adapting large language models by integrating collaborative semantics for recommendation

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. Adapting large language models by integrating collaborative semantics for recommendation. In2024 IEEE 40th International Conference on Data Engineering (ICDE), pages 1435–1448. IEEE, 2024

    2024

  20. [20]

    Qarm: Quantitative alignment multi-modal recommendation at kuaishou

    Xinchen Luo, Jiangxia Cao, Tianyu Sun, Jinkai Yu, Rui Huang, Wei Yuan, Hezheng Lin, Yichen Zheng, Shiyao Wang, Qigen Hu, et al. Qarm: Quantitative alignment multi-modal recommendation at kuaishou. InProceedings of the 34th ACM International Conference on Information and Knowledge Management, pages 5915– 5922, 2025

    2025

  21. [21]

    Generative next poi recommendation with semantic id

    Dongsheng Wang, Yuxi Huang, Shen Gao, Yifan Wang, Chengrui Huang, and Shuo Shang. Generative next poi recommendation with semantic id. InPro- ceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, pages 2904–2914, 2025

    2025

  22. [22]

    Forge: Forming semantic identifiers for generative retrieval in industrial datasets.arXiv preprint arXiv:2509.20904, 2025

    Kairui Fu, Tao Zhang, Shuwen Xiao, Ziyang Wang, Xinming Zhang, Chenchi Zhang, Yuliang Yan, Junjun Zheng, Yu Li, Zhihong Chen, et al. Forge: Forming semantic identifiers for generative retrieval in industrial datasets.arXiv preprint arXiv:2509.20904, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Mmq: Multimodal mixture-of- quantization tokenization for semantic id generation and user behavioral adap- tation.arXiv preprint arXiv:2508.15281, 2025

    Yi Xu, Moyu Zhang, Chenxuan Li, Zhihao Liao, Haibo Xing, Hao Deng, Jinxin Hu, Yu Zhang, Xiaoyi Zeng, and Jing Zhang. Mmq: Multimodal mixture-of- quantization tokenization for semantic id generation and user behavioral adap- tation.arXiv preprint arXiv:2508.15281, 2025

    arXiv 2025

  24. [24]

    Enhancing partially relevant video retrieval with robust alignment learning.arXiv preprint arXiv:2509.01383, 2025

    Long Zhang, Peipei Song, Jianfeng Dong, Kun Li, and Xun Yang. Enhancing partially relevant video retrieval with robust alignment learning.arXiv preprint arXiv:2509.01383, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Representation engineering: A top-down approach to ai transparency.arXiv preprint arXiv:2310.01405, 2023

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, et al. Representation engineering: A top-down approach to ai transparency.arXiv preprint arXiv:2310.01405, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Enhancing multiple dimensions of trustworthiness in llms via sparse activation control.Advances in Neural Information Processing Systems, 37:15730–15764, 2024

    Yuxin Xiao, Wan Chaoqun, Yonggang Zhang, Wenxiao Wang, Binbin Lin, Xiaofei He, Xu Shen, and Jieping Ye. Enhancing multiple dimensions of trustworthiness in llms via sparse activation control.Advances in Neural Information Processing Systems, 37:15730–15764, 2024

    2024

  27. [27]

    Steering knowledge selection behaviours in llms via sae-based representation engineering

    Yu Zhao, Alessio Devoto, Giwon Hong, Xiaotang Du, Aryo Pradipta Gema, Hongru Wang, Xuanli He, Kam-Fai Wong, and Pasquale Minervini. Steering knowledge selection behaviours in llms via sae-based representation engineering. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Lan...

    2025

  28. [28]

    Episodic memory repre- sentation for long-form video understanding.arXiv preprint arXiv:2508.09486, 2025

    Yun Wang, Long Zhang, Jingren Liu, Jiaqi Yan, Zhanjie Zhang, Jiahao Zheng, Xun Yang, Dapeng Wu, Xiangyu Chen, and Xuelong Li. Episodic memory repre- sentation for long-form video understanding.arXiv preprint arXiv:2508.09486, 2025

    arXiv 2025

  29. [29]

    Towards efficient partially relevant video retrieval with active moment discovering.arXiv preprint arXiv:2504.10920, 2025

    Peipei Song, Long Zhang, Long Lan, Weidong Chen, Dan Guo, Xun Yang, and Meng Wang. Towards efficient partially relevant video retrieval with active moment discovering.arXiv preprint arXiv:2504.10920, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Autodis: Automatic discretization for embedding numerical features in ctr prediction

    Huifeng Guo, Bo Chen, Ruiming Tang, Zhenguo Li, and Xiuqiang He. Autodis: Automatic discretization for embedding numerical features in ctr prediction. arXiv preprint arXiv:2012.08986, 2020

    Pith/arXiv arXiv 2012

  31. [31]

    Deep & cross network for ad click predictions

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. Deep & cross network for ad click predictions. InProceedings of the ADKDD’17, pages 1–7. 2017

    2017

  32. [32]

    Deep interest network for click-through rate prediction

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. Deep interest network for click-through rate prediction. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1059–1068, 2018

    2018

  33. [33]

    Learnable item tokenization for generative recommendation

    Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. Learnable item tokenization for generative recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 2400–2409, 2024

    2024

  34. [34]

    Eager: Two-stream generative recommender with behavior-semantic collaboration

    Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, et al. Eager: Two-stream generative recommender with behavior-semantic collaboration. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3245–3254, 2024

    2024

  35. [35]

    Onerec technical report.arXiv preprint arXiv:2506.13695, 2025

    Guorui Zhou, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Qiang Luo, Qianqian Wang, Qigen Hu, Rui Huang, Shiyao Wang, et al. Onerec technical report.arXiv preprint arXiv:2506.13695, 2025

    arXiv 2025

  36. [36]

    Onerec-v2 technical report

    Guorui Zhou, Hengrui Hu, Hongtao Cheng, Huanjie Wang, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Lu Ren, Liao Yu, et al. Onerec-v2 technical report. arXiv preprint arXiv:2508.20900, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Recbase: Generative foundation model pretraining for zero-shot recommendation

    Sashuai Zhou, Weinan Gan, Qijiong Liu, Ke Lei, Jieming Zhu, Hai Huang, Yan Xia, Ruiming Tang, Zhenhua Dong, and Zhou Zhao. Recbase: Generative foundation model pretraining for zero-shot recommendation. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 15598– 15610, 2025

    2025

  38. [38]

    Sparse meets dense: Unified generative recommendations with cascaded sparse-dense representations.arXiv preprint arXiv:2503.02453, 2025

    Yuhao Yang, Zhi Ji, Zhaopeng Li, Yi Li, Zhonglin Mo, Yue Ding, Kai Chen, Zijian Zhang, Jie Li, Shuanglong Li, et al. Sparse meets dense: Unified generative recommendations with cascaded sparse-dense representations.arXiv preprint arXiv:2503.02453, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Gflowgr: Fine-tuning generative recommendation frameworks with generative flow networks.arXiv preprint arXiv:2506.16114, 2025

    Yejing Wang, Shengyu Zhou, Jinyu Lu, Qidong Liu, Xinhang Li, Wenlin Zhang, Feng Li, Pengjie Wang, Jian Xu, Bo Zheng, et al. Gflowgr: Fine-tuning generative recommendation frameworks with generative flow networks.arXiv preprint arXiv:2506.16114, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    Unified semantic and id representation learning for deep recommenders.arXiv preprint arXiv:2502.16474, 2025

    Guanyu Lin, Zhigang Hua, Tao Feng, Shuang Yang, Bo Long, and Jiaxuan You. Unified semantic and id representation learning for deep recommenders.arXiv preprint arXiv:2502.16474, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    Carolina Zheng, Minhui Huang, Dmitrii Pedchenko, Kaushik Rangadurai, Siyu Wang, Gaby Nahum, Jie Lei, Yang Yang, Tao Liu, Zutian Luo, et al. Enhancing embedding representation stability in recommendation systems with semantic Long Zhang, Hao Jiang, Sheng Yu, Fei Pan, Peng Jiang, and Kun Gai id.arXiv preprint arXiv:2504.02137, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    Language models as semantic indexers

    Bowen Jin, Hansi Zeng, Guoyin Wang, Xiusi Chen, Tianxin Wei, Ruirui Li, Zhengyang Wang, Zheng Li, Yang Li, Hanqing Lu, et al. Language models as semantic indexers. InProceedings of the 41st International Conference on Machine Learning, pages 22244–22259, 2024

    2024

  43. [43]

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

    Ruining He, Lukasz Heldt, Lichan Hong, Raghunandan Keshavan, Shifan Mao, Nikhil Mehta, Zhengyang Su, Alicia Tsai, Yueqi Wang, Shao-Chuan Wang, et al. Plum: Adapting pre-trained language models for industrial-scale generative recommendations.arXiv preprint arXiv:2510.07784, 2025

    arXiv 2025

  44. [44]

    Personalized prompt learning for explain- able recommendation.ACM Transactions on Information Systems, 41(4):1–26, 2023

    Lei Li, Yongfeng Zhang, and Li Chen. Personalized prompt learning for explain- able recommendation.ACM Transactions on Information Systems, 41(4):1–26, 2023

    2023

  45. [45]

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

    Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, et al. Onerec-think: In-text reasoning for generative recommendation.arXiv preprint arXiv:2510.11639, 2025

    arXiv 2025

  46. [46]

    Probing classifiers: Promises, shortcomings, and advances

    Yonatan Belinkov. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207–219, 2022

    2022

  47. [47]

    Inference-time intervention: Eliciting truthful answers from a language model.Advances in Neural Information Processing Systems, 36:41451–41530, 2023

    Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Watten- berg. Inference-time intervention: Eliciting truthful answers from a language model.Advances in Neural Information Processing Systems, 36:41451–41530, 2023

    2023

  48. [48]

    The linear representation hy- pothesis and the geometry of large language models

    Kiho Park, Yo Joong Choe, and Victor Veitch. The linear representation hy- pothesis and the geometry of large language models. InProceedings of the 41st International Conference on Machine Learning, pages 39643–39666, 2024

    2024

  49. [49]

    Patchscopes: a unifying framework for inspecting hidden representations of language models

    Asma Ghandeharioun, Avi Caciularu, Adam Pearce, Lucas Dixon, and Mor Geva. Patchscopes: a unifying framework for inspecting hidden representations of language models. InProceedings of the 41st International Conference on Machine Learning, pages 15466–15490, 2024

    2024

  50. [50]

    Activation scaling for steering and interpreting language models

    Niklas Stoehr, Kevin Du, Vésteinn Snæbjarnarson, Robert West, Ryan Cotterell, and Aaron Schein. Activation scaling for steering and interpreting language models. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 8189–8200, 2024

    2024

  51. [51]

    Sharp: Steering hallucination in lvlms via representation engineering

    Junfei Wu, Yue Ding, Guofan Liu, Tianze Xia, Ziyue Huang, Dianbo Sui, Qiang Liu, Shu Wu, Liang Wang, and Tieniu Tan. Sharp: Steering hallucination in lvlms via representation engineering. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 14357–14372, 2025

    2025

  52. [52]

    Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769, 2024

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769, 2024

    Pith/arXiv arXiv 2024

  53. [53]

    Blip-2: bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: bootstrapping language-image pre-training with frozen image encoders and large language models. InProceedings of the 40th International Conference on Machine Learning, pages 19730–19742, 2023

    2023

  54. [54]

    Object-centric learning with slot attention.Advances in neural information processing systems, 33:11525–11538, 2020

    Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahen- dran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention.Advances in neural information processing systems, 33:11525–11538, 2020

    2020

  55. [55]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuropean conference on computer vision, pages 213–229. Springer, 2020

    2020

  56. [56]

    Attention is not all you need: Pure attention loses rank doubly exponentially with depth

    Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International conference on machine learning, pages 2793–2803. PMLR, 2021

    2021

  57. [57]

    The power of scale for parameter- efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter- efficient prompt tuning. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059, 2021

    2021

  58. [58]

    Rezero is all you need: Fast convergence at large depth

    Thomas Bachlechner, Bodhisattwa Prasad Majumder, Henry Mao, Gary Cottrell, and Julian McAuley. Rezero is all you need: Fast convergence at large depth. In Uncertainty in Artificial Intelligence, pages 1352–1361. PMLR, 2021

    2021

  59. [59]

    Highway networks.arXiv preprint arXiv:1505.00387, 2015

    Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber. Highway networks.arXiv preprint arXiv:1505.00387, 2015

    Pith/arXiv arXiv 2015

  60. [60]

    End-to-end user behavior retrieval in click-through rateprediction model.arXiv preprint arXiv:2108.04468, 2021

    Qiwei Chen, Changhua Pei, Shanshan Lv, Chao Li, Junfeng Ge, and Wenwu Ou. End-to-end user behavior retrieval in click-through rateprediction model.arXiv preprint arXiv:2108.04468, 2021

    Pith/arXiv arXiv 2021

  61. [61]

    Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction

    Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. InProceedings of the 29th ACM International Conference on Information & Knowledge Management, pages 2685– 2692, 2020

    2020

  62. [62]

    Twin v2: Scal- ing ultra-long user behavior sequence modeling for enhanced ctr prediction at kuaishou

    Zihua Si, Lin Guan, ZhongXiang Sun, Xiaoxue Zang, Jing Lu, Yiqun Hui, Xingchao Cao, Zeyu Yang, Yichen Zheng, Dewei Leng, et al. Twin v2: Scal- ing ultra-long user behavior sequence modeling for enhanced ctr prediction at kuaishou. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 4890–4897, 2024

    2024

  63. [63]

    Actions speak louder than words: Trillion- parameter sequential transducers for generative recommendations.arXiv preprint arXiv:2402.17152, 2024

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

    Pith/arXiv arXiv 2024

  64. [64]

    Rankmixer: Scaling up ranking models in industrial recommenders

    Jie Zhu, Zhifang Fan, Xiaoxie Zhu, Yuchen Jiang, Hangyu Wang, Xintian Han, Haoran Ding, Xinmin Wang, Wenlin Zhao, Zhen Gong, et al. Rankmixer: Scaling up ranking models in industrial recommenders. InProceedings of the 34th ACM International Conference on Information and Knowledge Management, pages 6309– 6316, 2025

    2025

  65. [65]

    Session-based recommendations with recurrent neural networks.arXiv preprint arXiv:1511.06939, 2015

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. Session-based recommendations with recurrent neural networks.arXiv preprint arXiv:1511.06939, 2015

    Pith/arXiv arXiv 2015

  66. [66]

    Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. InProceedings of the 28th ACM international conference on information and knowledge management, pages 1441–1450, 2019

    2019

  67. [67]

    Self-attentive sequential recommenda- tion

    Wang-Cheng Kang and Julian McAuley. Self-attentive sequential recommenda- tion. In2018 IEEE international conference on data mining (ICDM), pages 197–206. IEEE, 2018

    2018

  68. [68]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024.URL https://arxiv

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024.URL https://arxiv. org/abs/2402.03300, 2(3):5, 2024

    Pith/arXiv arXiv 2024

  69. [69]

    Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741, 2023

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741, 2023

    2023

  70. [70]

    Qwen3 technical report, 2025

    Qwen Team. Qwen3 technical report, 2025

    2025

  71. [71]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    2022

  72. [72]

    local specialty snacks

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. Interpretable Representation via LLM-Driven Generative Disentanglement for Local-Life Service Recommendation Appendix A REPRODUCIBILITY DETAILS All industrial data are anonymized and used only for offline evalua- tion; user identifiers and exact ...

    Pith/arXiv arXiv 2017