Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Just Ask for Music (JAM): Multimodal and Personalized Natural Language Music Recommendation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A translation equation user + query = track personalizes music search

desk verdict Worth a look for the JAMSessions dataset; the translation-mechanism claim needs a concat/sum ablation before you believe it. read the letter →

arxiv 2507.15826 v1 pith:N6OJUY3C submitted 2025-07-21 cs.IR cs.LG

classification cs.IRcs.LG
keywords naturallanguagemusicrecommendationmultimodaltranslation-basedembeddingscross-attentionaggregationsparsemixtureofexpertspersonalizeduser-query-itemtriplesJAMSessionsdataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

JAM (Just Ask for Music) argues that natural-language music recommendation can be done by a lightweight retrieval model instead of an expensive large language model. The core move is to treat a query as a vector translation that moves a user's long-term taste embedding toward the right tracks: $\mathbf{u} + \mathbf{q} \approx \mathbf{t}$. On JAMSessions, a new dataset of 112,337 real user-query-item triples, the cross-attention variant reaches Recall@10 of 0.086 and NDCG@10 of 0.311, outperforming the TalkRec and TwoTower baselines. The payoff, if the approach holds, is that a streaming service can bolt conversational search onto an existing recommender stack with small encoders and no full retraining.

What carries the argument

The central object is the translation identity $\mathbf{u} + \mathbf{q} = \hat{\mathbf{t}}$, inherited from TransE knowledge-graph embeddings: $\mathbf{u}$ is the user's long-term collaborative-filtering embedding, $\mathbf{q}$ is a text-encoded query, and $\hat{\mathbf{t}}$ is the aggregated multimodal item embedding. JAM explores three aggregation mechanisms for the item side — plain averaging, cross-attention where the query weights each modality, and sparse mixture-of-experts with noisy top-k gating — and trains with a Bayesian Personalized Ranking loss over positive triples and sampled negative items. Cross-attention aggregates audio, lyrics, and collaborative-filtering signals dynamically, which is what lets the same query emphasize different facets of a track depending on the user.

What would settle it

Sample JAMSessions triples, have independent annotators judge whether each track actually satisfies the paired query; if agreement with the playlist-derived labels is low, the measured Recall and NDCG gains would not reflect true query-item relevance. A second check would randomize the user embeddings and see whether cross-attention's advantage survives, since that tests whether personalization is really doing the work.

Watch

Extended reading notes

Core claim

The paper claims that personalized natural-language music recommendation can be modeled as a translation in a shared latent space, in the style of knowledge-graph embedding methods such as TransE: the user embedding plus the query embedding should land on the relevant item embedding. JAM keeps the precomputed user, item, and query representations fixed and learns only small projection layers plus a modality-aggregation module, so the underlying encoders do not need to be retrained. Among the aggregation variants tested, query-guided cross-attention over audio, lyrics, and collaborative-filtering embeddings performs best and gives the learned space interpretable translation semantics: the same query "partying like crazy" moves different users to different musical regions, while the same user issuing different queries lands in correspondingly different genres.

Load-bearing premise

The ground-truth relevance labels are assumed: a user searching a query, landing on an editor-curated playlist, and listening for over ten minutes makes every track in that playlist relevant to the query, and the LLM-based query augmentation introduces further unvalidated variability.

Editorial extensions

If this is right

  • All JAM variants beat the baselines, suggesting natural-language interfaces can be added to existing stacks without full retraining.
  • Cross-attention is the strongest aggregator because it reweights modalities per query; a query about an upbeat motif can lean on audio while a query about love songs can lean on lyrics.
  • The 112,337-triple JAMSessions dataset lets other researchers train and compare translation-based recommenders with both queries and long-term preferences.
  • The same query applied to different users lands in different parts of the latent space, giving practitioners an interpretable reason for each recommendation.
  • Sparsifying modalities with mixture-of-experts hurts accuracy, so all available modalities carry useful signal; dropping to one modality reduces performance sharply.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • This suggests the same translation recipe could transfer to other domains with precomputed embeddings, such as video or podcast search.
  • Because user and item collaborative-filtering embeddings are precomputed together, the observed dominance of the CF modality in attention weights may overstate collaborative filtering's true value; a cold-start evaluation with items lacking CF history would test this.
  • The paper's noted failure on artist-name queries suggests a direct extension: include an artist embedding as a fourth modality, which should anchor explicit artist requests to the right neighborhood.
  • A practical follow-up would freeze JAM's learned projections and re-evaluate on later time windows to measure how quickly the translation space drifts as the catalog and tastes change.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript introduces JAM, a framework for personalized natural-language music recommendation. The core modeling idea is to treat a query q as a translation vector in a shared latent space so that u + q ≈ t for a user u and item t, following the TransE knowledge-graph embedding paradigm. Item representations are aggregated from three modalities (audio, lyrics, and collaborative-filtering embeddings) using either averaging, query-conditioned cross-attention, or sparse mixture-of-experts. The model is trained with a BPR-style pairwise loss that maximizes dot-product similarity for positive triples over sampled negatives. The authors also release JAMSessions, a dataset of 112,337 user–query–item triples collected from Deezer search logs over one week in March 2025. In experiments, the CrossMixing variant achieves the best reported Recall@10 and NDCG@10 values, and qualitative examples illustrate how the same query leads to different recommendations for different users.

Significance. JAM addresses an industrially relevant deployment scenario: it keeps pretrained user, query, and item encoders fixed and learns only lightweight projection and aggregation layers, so it can be integrated into existing recommender stacks without full retraining. The release of JAMSessions, with over 100k triples and precomputed embeddings, is a potentially valuable resource for the community. The paper reports three-seed means and standard deviations, uses a chronological split, and provides a code/data link. The main strengths are the clean formulation and the concrete dataset artifact. However, the experimental design does not isolate the translation inductive bias from the more generic effect of conditioning on both user and query, and the ground-truth construction relies on an unvalidated playlist-landing proxy. These issues prevent the reported numbers from being read as direct evidence for the specific u+q=t mechanism, although the approach remains plausible and worth further investigation.

major comments (3)
  1. [Section 3 and Section 5, Table 2] The comparison does not support the central claim that the additive translation formulation drives accuracy. The two main baselines each drop one input: TalkRec uses the query but not the user, and TwoTower uses the user but not the query. JAM uses both. The observed gains are therefore consistent with any joint user+query encoder, not specifically with the u+q=t inductive bias. Please add an ablation that combines user and query through a non-translation operator (e.g., concatenation followed by a linear projection, or a gated sum) with the same item aggregation and the same loss, and report this variant in Table 2. Without such a baseline, the qualitative TSNE and top-3 examples cannot distinguish the translation mechanism from a generic additive composition.
  2. [Section 2 and Section 4] The relevance labels are constructed from a user entering a query, landing on an editor-curated playlist, and listening for over 10 minutes; every track in that playlist is then treated as relevant to the query. This is a strong proxy that needs validation. Please report an internal agreement study or a small human evaluation on a sample of triples, and state how tracks in eclectic or multi-genre playlists are handled. The paper also notes that LLM-augmented queries include erroneous generations; please either filter these or analyze their effect on the reported metrics, since label noise of this kind directly affects the validity of the accuracy numbers.
  3. [Section 4] The candidate set and negative sampling procedure are not specified. For a fair comparison and reproducibility, state whether negatives are drawn uniformly from the full catalog or from a candidate set, how many negatives are sampled per positive, and whether in-batch negatives are used. In addition, clarify whether the precomputed user/item CF embeddings are built from an interaction matrix that includes the test week; if so, user embeddings on the test set contain future information and the chronological split does not prevent leakage. This point is load-bearing for interpreting all reported metrics.
minor comments (5)
  1. [Appendix references] The text defers to an appendix ('Full prompt available in the Appendix', 'Details are provided in the Appendix'), but no appendix is present in the reviewed version. Please ensure the camera-ready includes the appendix with the two-shot prompt, hyperparameter search ranges, and the quality-check details.
  2. [Section 3, MoE equation] The MoE equation contains a typo ('Sotfmax') and inconsistent use of tildes on input representations; please proofread the formulas in this section.
  3. [Table 1] The column header 'Queries 7 Users Tracks' appears corrupted, and the MPD row reports 1,000,000 queries even though MPD is primarily a playlist dataset; please clarify the semantics of each column in the caption.
  4. [Section 5, CF modality discussion] The statement that 'CF signals contribute most' is based on attention/gating weights, but no modality ablation is reported; please either add such an ablation (e.g., running CrossMixing with subsets of modalities) or soften the claim, especially since the paper itself acknowledges the possible initialization bias.
  5. [Section 3, training objective] The phrase 'positive triplet satisfying u+q=t' is misleading: the equation is the modeling assumption, not a constraint that the data satisfies. Please rephrase to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: JAM's accuracy claims are supported by a standard train/test comparison, and the qualitative translation analysis illustrates the trained objective rather than serving as an independent prediction.

full rationale

The paper's central claims are empirical: JAM is trained with a BPR loss on user-query-item triples from JAMSessions, and its accuracy is compared against TalkRec, TwoTower, and simple baselines on a chronologically split test set. The translation formulation u + q ≈ t is not derived from the evaluation results; it is the model definition, and Table 2 reports measured retrieval metrics that are not equivalent to the training objective by construction. The qualitative analysis in Section 5 shows that the trained model places u + q near recommended items, which is a direct consequence of the loss, but the paper presents this as an illustration of the learned space rather than as evidence that the translation mechanism outperforms alternatives. The absence of a concat/sum ablation is a legitimate experimental limitation, but it does not make the reported gains logically forced by the inputs. The JAMSessions dataset is built from search logs with a clearly stated relevance assumption, and the precomputed CF embeddings and their potential influence on modality attention are acknowledged in the paper. Self-citations and co-authored references are used for contextual or methodological support, not as a uniqueness theorem or as the sole justification of the main result. Therefore, no step in the derivation chain reduces to its own inputs or to a self-citation chain.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim depends on five assumptions: validity of playlist-based relevance labels, faithfulness of precomputed embeddings, LLM augmentation preserving intent, the translation modeling assumption, and split independence. Free parameters are the embedding dimension, learning rate, MoE K, and negative count. No invented entities are introduced.

free parameters (4)
  • Shared latent dimension d = Tuned, reported in appendix
    Dimensionality of the shared user/query/item space; affects capacity of the translation model and is tuned for all baselines.
  • Learning rate = Tuned, reported in appendix
    Tuned jointly with the embedding dimension; a standard training hyperparameter that influences convergence and final accuracy.
  • MoE Top-K gating K = 2, with 1 also tested
    Number of modalities allowed to contribute in sparse mixture-of-experts; K=2 is used for main results and K=1 performs considerably worse.
  • Number of negative items per positive triple = 4
    Fixed negative sampling rate in the BPR-style loss; affects contrastive learning and ranking behavior.
assumptions (5)
  • domain assumption Relevant items are tracks from an editor-curated playlist the user landed on and listened to for over 10 minutes after submitting a search query.
    Defines ground truth for both training and evaluation; assumed to reflect query relevance and user intent.
  • domain assumption Precomputed CF, audio, and lyrics embeddings faithfully represent users and items in the three modalities.
    JAM keeps these embeddings fixed; if they are weak, query translation inherits their weaknesses.
  • ad hoc to paper LLM-augmented queries preserve the user's original intent and add useful variability.
    DeepSeek-generated rewrites with two-shot prompting are used without automated validation; internal checks found erroneous generations.
  • ad hoc to paper The user-query-item relation can be modeled as vector translation u + q = t.
    TransE-inspired geometric assumption; central modeling choice not derived from data.
  • domain assumption Chronological splitting prevents leakage even though the same users can appear in train and test.
    Test queries come from the last day and training from earlier days, but the paper does not analyze user overlap or session dependence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Just Ask for Music (JAM): Multimodal and Personalized Natural Language Music Recommendation." pith.science (2026). https://pith.science/paper/N6OJUY3C

@misc{pith2026250715826,
  author       = {Pith},
  title        = {Pith review of: Just Ask for Music (JAM): Multimodal and Personalized Natural Language Music Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N6OJUY3C}},
  note         = {Machine review of arXiv:2507.15826}
}
read the original abstract

Natural language interfaces offer a compelling approach for music recommendation, enabling users to express complex preferences conversationally. While Large Language Models (LLMs) show promise in this direction, their scalability in recommender systems is limited by high costs and latency. Retrieval-based approaches using smaller language models mitigate these issues but often rely on single-modal item representations, overlook long-term user preferences, and require full model retraining, posing challenges for real-world deployment. In this paper, we present JAM (Just Ask for Music), a lightweight and intuitive framework for natural language music recommendation. JAM models user-query-item interactions as vector translations in a shared latent space, inspired by knowledge graph embedding methods like TransE. To capture the complexity of music and user intent, JAM aggregates multimodal item features via cross-attention and sparse mixture-of-experts. We also introduce JAMSessions, a new dataset of over 100k user-query-item triples with anonymized user/item embeddings, uniquely combining conversational queries and user long-term preferences. Our results show that JAM provides accurate recommendations, produces intuitive representations suitable for practical use cases, and can be easily integrated with existing music recommendation stacks.

Figures

Figures reproduced from arXiv: 2507.15826 by the authors.

Figure 1
Figure 1. JAM (Just Ask for Music) framework outline. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. TSNE of item embeddings, two users (u’, u"), their [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 27 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report

  2. [2]

    Antoine Bordes, Nicolas Usunier, Alberto Garcia-Durán, Jason Weston, and Ok- sana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. In Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2 (Lake Tahoe, Nevada) (NIPS’13). Curran Associates Inc., Red Hook, NY, USA, 2787–2795

  3. [3]

    Arun Tejasvi Chaganty, Megan Leszczynski, Shu Zhang, Ravi Ganti, Krisztian Balog, and Filip Radlinski. 2023. Beyond single items: Exploring user preferences in item sets with the conversational playlist curation dataset. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2754–2764

  4. [4]

    Arun Tejasvi Chaganty, Megan Leszczynski, Shu Zhang, Ravi Ganti, Krisztian Balog, and Filip Radlinski. 2023. Beyond Single Items: Exploring User Preferences in Item Sets with the Conversational Playlist Curation Dataset. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2023, Taipei, ...

  5. [5]

    Ching-Wei Chen, Paul Lamere, Markus Schedl, and Hamed Zamani. 2018. Recsys challenge 2018: automatic music playlist continuation. In Proceedings of the 12th ACM Conference on Recommender Systems, RecSys 2018, Vancouver, BC, Canada, October 2-7, 2018. ACM, 527–528. doi:10.1145/3240323.3240342

  6. [6]

    Jin Chen, Zheng Liu, Xu Huang, Chenwang Wu, Qi Liu, Gangwei Jiang, Yuanhao Pu, Yuxuan Lei, Xiaolong Chen, Xingmei Wang, et al. 2024. When large language models meet personalization: Perspectives of challenges and opportunities.World Wide Web 27, 4 (2024), 42

  7. [7]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for YouTube Recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems (Boston, Massachusetts, USA) (RecSys ’16). Association for Computing Machinery, New York, NY, USA, 191–198. doi:10.1145/2959100. 2959190

  8. [8]

    Marco De Nadai, Francesco Fabbri, Paul Gigioli, Alice Wang, Ang Li, Fabrizio Silvestri, Laura Kim, Shawn Lin, Vladan Radosavljevic, Sandeep Ghael, et al

Show all 49 references
  1. [9]

    Mathieu Delcluze, Antoine Khoury, Clémence Vast, Valerio Arnaudo, Léa Briand, Walid Bendada, and Thomas Bouabça. 2025. Text2Playlist: Generating Personal- ized Playlists from Text on Deezer. InThe 47th European Conference on Information Retrieval (ECIR 2025)

  2. [10]

    SeungHeon Doh, Keunwoo Choi, Daeyong Kwon, Taesu Kim, and Juhan Nam

  3. [11]

    Seungheon Doh, Keunwoo Choi, and Juhan Nam. 2025. TALKPLAY: Multi- modal Music Recommendation with Large Language Models. arXiv preprint arXiv:2502.13713 (2025)

  4. [12]

    Elena V Epure, Gabriel Meseguer-Brocal, Darius Afchar, and Romain Hennequin

  5. [13]

    Large Language Models

    Music Discovery Dialogue Generation Using Human Intent Analysis and 10https://research.deezer.com RecSys ’25, September 22–26, 2025, Prague, Czech Republic Melchiorre et al. Large Language Models. In Proceedings of the 25th International Society for Music Information Retrieval...

  6. [14]

    Andres Ferraro, Sergio Oramas, Massimo Quadrana, and Xavier Serra. 2020. Maximizing the engagement: exploring new signals of implicit feedback in music recommendations. In Proceedings of the Workshops on Recommendation in Complex Scenarios and the Impact of Recommender Systems...

  7. [15]

    Giovanni Gabbolini, Romain Hennequin, and Elena Epure. 2022. Data-Efficient Playlist Captioning With Musical and Linguistic Knowledge. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, Abu Dhabi...

  8. [16]

    In Proceedings of the 3rd Workshop on NLP for Music and Audio (NLP4MusA)

    Harnessing High-Level Song Descriptors towards Natural Language-Based Music Recommendation. In Proceedings of the 3rd Workshop on NLP for Music and Audio (NLP4MusA). 17–24

  9. [17]

    Andres Ferraro, Yuntae Kim, Soohyeon Lee, Biho Kim, Namjun Jo, Semi Lim, Suyon Lim, Jungtaek Jang, Sehwan Kim, Xavier Serra, and Dmitry Bogdanov

  10. [18]

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. 1991. Adaptive mixtures of local experts. Neural computation 3, 1 (1991), 79–87

  11. [19]

    Dietmar Jannach, Ahtsham Manzoor, Wanling Cai, and Li Chen. 2021. A survey on conversational recommender systems. ACM Computing Surveys (CSUR) 54, 5 (2021), 1–36

  12. [20]

    Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, et al . 2025. How can recommender systems benefit from large language models: A survey. ACM Transactions on Information Systems 43, 2 (2025), 1–47

  13. [21]

    Christian Ganhör, Marta Moscati, Anna Hausberger, Shah Nawaz, and Markus Schedl. 2024. A Multimodal Single-Branch Embedding Network for Recommenda- tion in Cold-Start and Missing Modality Scenarios. InProceedings of the 18th ACM Conference on Recommender Systems (Bari, Italy)(...

  14. [22]

    Ruining He, Wang-Cheng Kang, and Julian McAuley. 2017. Translation-based recommendation. In Proceedings of the eleventh ACM conference on recommender systems. 161–169

  15. [23]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In Proc. ICLR

  16. [24]

    Ilaria Manco, Emmanouil Benetos, Elio Quinton, and George Fazekas. 2022. Con- trastive Audio-Language Learning for Music. In Ismir 2022 Hybrid Conference

  17. [25]

    Lilian Marey, Bruno Sguerra, and Manuel Moussallam. 2024. Modeling Activity- Driven Music Listening with PACE. In Proceedings of the 2024 Conference on Human Information Interaction and Retrieval . 346–351

  18. [26]

    Adam J Lonsdale and Adrian C North. 2011. Why do we listen to music? A uses and gratifications analysis. British Journal of Psychology 102, 1 (2011), 108–134

  19. [27]

    Ilya Loshchilov and Frank Hutter. 2016. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)

  20. [28]

    Sergio Oramas, Andres Ferraro, Alvaro Sarasua, and Fabien Gouyon. 2024. Talking to Your Recs: Multimodal Embeddings For Recommendation and Retrieval. In Proceedings of the 2nd Music Recommender Systems Workshop 2024 co-located with the 18th ACM Conference on Recommender System...

  21. [29]

    Enrico Palumbo, Gustavo Penha, Andreas Damianou, José Luis Redondo Gar- cía, Timothy Christopher Heath, Alice Wang, Hugues Bouchard, and Mounia Lalmas. 2024. Text2Tracks: Generative Track Retrieval for Prompt-based Music Recommendation. In The 1st Workshop on Risks, Opportunit...

  22. [30]

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

  23. [31]

    Brian McFee, Thierry Bertin-Mahieux, Daniel PW Ellis, and Gert RG Lanckriet

  24. [32]

    Noah Tekle, Alline Ayala, Jonathan Haile, Abdulla Alshabanah, Corey Baker, and Murali Annavaram. 2024. Music Recommendation through LLM Song Summary. In The 1st Workshop on Risks, Opportunities, and Evaluation of Generative Models in Recommender Systems (ROEGEN@RECSYS’24)

  25. [33]

    Gabriel Meseguer-Brocal, Dorian Desblancs, and Romain Hennequin. 2024. An experimental comparison of multi-view self-supervised methods for music tag- ging. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1141–1145

  26. [34]

    Viet-Anh Tran, Guillaume Salha-Galvan, Romain Hennequin, and Manuel Mous- sallam. 2021. Hierarchical latent relation modeling for collaborative metric learning. In Proceedings of the 15th ACM Conference on Recommender Systems . 302–309

  27. [35]

    Viet-Anh Tran, Guillaume Salha-Galvan, Bruno Sguerra, and Romain Hennequin

  28. [36]

    Roberto Turrin, Massimo Quadrana, Andrea Condorelli, Roberto Pagano, and Paolo Cremonesi. 2015. 30Music Listening and Playlists Dataset. In Poster Pro- ceedings of the 9th ACM Conference on Recommender Systems, RecSys 2015, Vienna, Austria, September 16, 2015 (CEUR Workshop Pr...

  29. [37]

    arXiv preprint arXiv:1205.2618 (2012)

    BPR: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618 (2012)

  30. [38]

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538 (2017)

  31. [39]

    Benjamin Warner, Antoine Chaffin, Benjamin Clavié, Orion Weller, Oskar Hall- ström, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, et al. 2024. Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long con...

  32. [40]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

  33. [41]

    Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, et al. 2024. Recommender systems in the era of large language models (llms). IEEE Transactions on Knowledge and Data Engineering (2024)

  34. [43]

    In Proceedings of the 18th ACM Conference on Recommender Systems

    Transformers Meet ACT-R: Repeat-Aware and Sequential Listening Session Recommendation. In Proceedings of the 18th ACM Conference on Recommender Systems. 486–496

  35. [45]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017)

  36. [46]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672 (2024)

  37. [48]

    Lanling Xu, Junjie Zhang, Bingqian Li, Jinpeng Wang, Mingchen Cai, Wayne Xin Zhao, and Ji-Rong Wen. 2024. Prompting large language models for recommender systems: A comprehensive framework and empirical analysis. arXiv preprint arXiv:2401.04997 (2024)

  38. [2012]

    In Proceedings of the 21st International Conference on World Wide Web

    The million song dataset challenge. In Proceedings of the 21st International Conference on World Wide Web. 909–916

  39. [2021]

    In International Conference on Acoustics, Speech and Signal Processing (ICASSP 2021)

    Melon Playlist Dataset: a public dataset for audio-based playlist generation and music tagging. In International Conference on Acoustics, Speech and Signal Processing (ICASSP 2021)

  40. [2024]

    In Companion Proceedings of the ACM Web Conference 2024

    Personalized audiobook recommendations at spotify through graph neural networks. In Companion Proceedings of the ACM Web Conference 2024 . 403–412

  41. [2764]

    doi:10.1145/3539618.3591881

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.