Pith. sign in

REVIEW 3 major objections 5 minor 40 references

The paper argues that natural-language audience tags, converted into a collaborative factor query, can restore the collaborative signal missing from semantic-ID recommendation without retraining.

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 03:23 UTC pith:H4FWBVEN

load-bearing objection Useful training-free collaborative reranking for frozen SID recommenders, with a solid main result and an under-supported HPF-specificity claim. the 3 major comments →

arxiv 2607.27682 v1 pith:H4FWBVEN submitted 2026-07-30 cs.IR

Restoring Collaborative Signals in Semantic-ID Generative Recommendation via Personalized Natural Language

Changjiang Han , Qingyang Li , Yaqiang Zang , Jikun Kang , Pinghua Gong , Xue Liu , Bowei He This is my paper
classification cs.IR
keywords generative recommendationsemantic IDscollaborative filteringhierarchical Poisson factorizationinference-time rerankingnatural languageLLMhit rate
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.

Semantic-ID (SID) recommenders represent each item as a short code tuple and generate the next item's codes autoregressively; the paper argues that such content-built codes cannot carry collaborative (co-consumption) structure and content simultaneously, so collaboration loses — and because a mis-predicted SID is a wrong recommendation, this directly caps accuracy. The paper tries to establish that the missing collaborative signal can be restored purely at inference time, on a frozen generator, without retraining the tokens: factorise a second-order item–item co-view matrix into per-item collaborative factors, aggregate them onto SID prefixes, and learn a bridge that turns an LLM's summary of a user's history into the same factor space as a query. That query then re-ranks the model's own candidate codes level by level — expanding coarse codes, reranking mid-level codes, and constraining the final level to valid items — without widening the beam. Across two open backbones the paper reports consistent hit-rate gains (hit@10 11.14→15.50 on the larger model), attributes the gain to the mid-level decision, and shows the gain depends on the specific collaborative representation rather than an arbitrary low-dimensional one, while enabling native reasoning lowers accuracy. The language-reconstructed query preserves the utility of a history-derived one and separates users' audiences at the ranking level, but the effect is modest, which the paper traces to the profile's granularity rather than the channel.

Core claim

The central claim is that the collaboration deficit of content-built semantic IDs can be repaired without retraining the tokenizer or the generator: factorising a second-order item–item co-view matrix gives per-item collaborative factors that, aggregated onto SID prefixes, yield a content-agnostic signature addressable by exactly the codes the generator emits. A learned bridge translates an LLM's natural-language audience tags into this factor space, producing a user query that is faithful to the behavioural one and leakage-free. At inference the query adds a collaborative residual at the coarse and mid levels of the frozen backbone's own beam search — candidate expansion at the first code l

What carries the argument

The load-bearing object is the second-order item–item co-view matrix, counting how many users consumed each unordered pair of items, factorised by hierarchical Poisson factorization (a Gamma–Poisson probabilistic model) into non-negative per-item latent vectors. These are averaged onto SID prefixes to give signatures indexed by the very codes the generator emits. A learned bridge maps natural-language audience tags into this factor space to build a user query. The inference-time rule adds a collaborative residual to the backbone's own candidate scores level by level under a fixed beam.

Load-bearing premise

The load-bearing premise is that the second-order co-view matrix encodes genuine audience structure rather than popularity or content similarity already present in the SID; if it mostly reflects frequency, the mid-level reranking gain would collapse to a popularity artifact.

What would settle it

Construct a perturbed version of the benchmark where the item–item co-view matrix is rebuilt after randomly shuffling a user's history items while keeping the same marginals, then run the method on the frozen backbone. If hit@10 on the larger model still rises near 15.50, the reported gain comes from frequency rather than audience structure; if the gain vanishes, the collaborative interpretation is confirmed.

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

If this is right

  • Enabling a frozen backbone's native reasoning lowers full-SID accuracy at every cutoff, so the text-interface route cannot supply the missing collaborative signal.
  • The gain is carried by the mid-level reranking: adding coarse-code expansion alone is within noise, and the mid-level residual produces the significant jump.
  • The benefit is representation-specific: HPF co-watch factors improve over TF-IDF+SVD and K-means co-watch representations, both of which fall below the frozen baseline.
  • The language-reconstructed query matches the history-derived query in downstream utility, so collaborative signal survives the language bottleneck; it also separates users at the ranking level.
  • Because the channel acts only on the collaborative factor space and never on the SID embeddings, it in principle extends to other semantic-ID generators, including encoder-decoder ones.

Where Pith is reading between the lines

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

  • If the content–collaborative exclusivity diagnosis is right, then any content-only SID tokenizer caps the accuracy of frozen generators at roughly the prefix-reachability ceiling, making external collaborative injection a necessary complement rather than a luxury.
  • The modest user differentiation suggests the profile, not the channel, is the bottleneck; a per-user memory that retrieves tags conditioned on the current need — which the authors flag as future work — is a testable extension that should sharpen the query and enlarge the effect.
  • The method's ceiling is the reachability of the target prefix in the candidate pool; a natural stress test is to measure how the gain varies with the candidate budgets, since expansion beyond the pool requires either a wider beam or a hybrid retrieve-and-generate scheme.
  • The same ladder — factor query, language bridge, level-wise residual — could be reused outside full-SID recommendation, e.g., for controlled generation with other hierarchical discrete codes, where the language-conditioned factor query acts as a steerable knob.

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

3 major / 5 minor

Summary. The paper proposes a training-free, inference-time method for injecting collaborative signals into a frozen semantic-ID (SID) generative recommender. It factorizes a second-order item–item co-view matrix with hierarchical Poisson factorization (HPF), summarizes each user's history into natural-language audience tags, learns a bridge P(f|t) mapping tags to the HPF factor space, and reranks the frozen backbone's own SID candidates level-by-level (coarse-code expansion at s_a, residual reranking at s_b, and an inventory mask at s_c). Experiments on two OneRec backbones (1.7B and 8B) on the RecIF benchmark report consistent full-SID hit-rate gains (e.g., 8B hit@10 from 11.14 to 15.50), with level-wise ablations attributing most of the gain to the s_b reranking step. Controls include a popularity reranker, a semantic-similarity reranker, a history-derived factor query, and a user-specificity analysis showing a modest but statistically significant audience-differentiation effect.

Significance. If the findings hold, the paper offers a practical recipe for supplying collaborative signal to SID generators without retraining the backbone or the semantic IDs, which is appealing because SID codebooks are expensive to relearn. The empirical study is careful in several respects: paired-bootstrap confidence intervals are reported, the ablation in Table 2 isolates the contribution of each level, and the limitations section honestly states the single-benchmark and single-backbone-family scope. The paper also includes useful controls (popularity, text-interface, history-derived query). However, the central claim of HPF-specificity—that the gain depends on this particular collaborative representation rather than any low-dimensional embedding—is not yet adequately supported by the evidence in Table 3, and the headline number mixes a non-collaborative inventory constraint with the collaborative reranking effect. These issues need to be resolved before the paper's main claims can be accepted as stated.

major comments (3)
  1. [Methodology, Eq. (3)] The central bridge equation P(f|t) is typeset in a way that is not interpretable: the fraction as rendered has no clear denominator or normalization. Since this equation defines the natural-language bridge that is a core contribution, the method is underspecified. Please rewrite Eq. (3) with explicit terms, including the normalization over f and the role of the smoothing constant μ.
  2. [Experiments, Table 3 and 'Dependence on the collaborative representation'] The claim that the gain depends on HPF rather than an arbitrary low-dimensional representation is supported only by comparisons against TF-IDF+SVD and K-means. These are not standard collaborative-filtering baselines: TF-IDF is an unusual normalization for a co-view count matrix, and K-means discards fine-grained similarity. A generic low-rank embedding such as truncated SVD on the raw counts, or a standard MF baseline (ALS or logistic MF), should be tested. If a generic SVD or MF also yields large gains, the HPF-specificity claim collapses to 'any collaborative vector helps'. This is load-bearing because the contributions explicitly argue for HPF specificity. Please add such baselines or substantially soften the claim.
  3. [Experiments, Table 2 vs. Table 1 and Abstract/Contributions] The baseline Base in Table 2 appears not to include the s_c inventory constraint, while Ours does. In the cumulative ablation, the inventory constraint adds a non-collaborative gain (e.g., 8B: +0.85pp on hit@10, from 14.65 to 15.50). The abstract and contributions quote the full 11.14→15.50 improvement without decomposing it. This conflates a trivial inventory mask with the collaborative channel. Please report a baseline that already includes the inventory constraint (so the comparison isolates the collaborative signal), and in the abstract/contributions either quote the collaborative-only gain or explicitly state that the headline number includes the inventory mask.
minor comments (5)
  1. [Abstract and Contributions] The phrase 'the channel yields consistent hit-rate gains' should be qualified: the gain decomposes into the s_b collaborative reranking and the s_c inventory mask. Consider reporting both numbers in the abstract.
  2. [Experimental Setup] The LLM prompt used to summarize user histories into audience tags is not described. The tag ontology is also only given by two examples. For reproducibility, please provide the full prompt template and a list of canonical tags in an appendix.
  3. [Figure 2] The PCA plot is stated to be 'quantified in full embedding dimension', but the quantitative separation percentage (75–79%) is only reported for the PCA-projected visualization. Give the actual overlap measure in the embedding space.
  4. [Notation in Eq. (5)] The subscript of the softmax in 'softmax_sa' is ambiguous; it should be clear whether the softmax is over all s_a codes or over the candidate set. This affects the scale of the collaborative residual at the coarse level.
  5. [Limitations] The limitations paragraph on reachability is helpful, but it could be more explicit that the reported (s_a,s_b)@20 reachability improves precisely because the collaborative reranking changes which prefixes are in the beam, not because the beam width grows.

Circularity Check

1 steps flagged

Minor definitional circularity in the natural-language query bridge; the central held-out hit-rate result is externally grounded.

specific steps
  1. self definitional [Methodology, Natural-Language Bridge to the Factor Space, Eqs. (3)-(4)]
    "Writing q^F_u for a training user’s behavioural factor query (the row-normalized mean of their history-item factors φ_i)... we accumulate each user’s factors, split evenly across their tags, and smooth toward the global factor prior π: P(f|t) = Σ_{u:t∈T_u} 1/|T_u| q^F_u + µπ ... (3) ... the natural-language query is the normalized sum of the corresponding tag distributions, q_u = normalize(Σ_{t∈T_u} P(f|t)) (4), which lies in the same factor space as the behavioural query."

    The 'natural-language' query q_u is not an independent language-derived signal: Eq. (3) defines each tag distribution P(f|t) as a smoothed average of training users' behavioural factor queries q^F_u, and Eq. (4) makes q_u a normalized sum of those same training-user behavioural queries. Therefore the claim that the bridge is 'faithful (the language-reconstructed query matches the behavioural one)' is true by construction rather than by independent measurement. The paper's downstream hit-rate gains, however, are evaluated on held-out next-item targets and do not reduce to this definition, so the circularity is partial and confined to the 'language channel preserves the behavioural query' claim.

full rationale

The central empirical claim — that a frozen SID generator's hit-rate improves when a collaborative factor-space query is injected at inference — is tested on held-out next-item targets from the open RecIF/OpenOneRec benchmark, with a GT-disjoint evaluation set and a shared candidate budget. Hit@10 gains (e.g., OneRec-8B 11.14 to 15.50) are external outcomes, not consequences of fitting the reranking weight (λ=0.1 fixed) or of the bridge equations. The level-wise ablation (Table 2) and representation control (Table 3) are likewise empirical, even though the HPF-vs-SVD/K-means comparison is a weak control set; that is an evidence-strength concern, not circularity. Self-citations to OneRec Team 2026 and He et al. 2026 for 'reasoning hurts' are not load-bearing because the paper reproduces that result with its own experiments (8B hit@10 10.90→7.26). The only definitional reduction is the natural-language query: by Eqs. (3)-(4) it is a smoothed average of the very behavioural factor queries it is said to reconstruct, making the 'faithful bridge' claim partially self-fulfilling. This is a minor, non-central circularity; the main accuracy result stands on independent held-out evaluation.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central claim rests on a chain of modeling choices: SIDs under-express collaboration (cited from DiscRec), co-view counts carry audience structure, HPF factors aggregated to prefixes are the right carrier, LLM tags plus the bridge reconstruct behavior, and the frozen backbone's candidate pool is reachable. Each is plausible and partially controlled, but none is independently verified here.

free parameters (6)
  • HPF rank d = 64
    Dimensionality of the collaborative factor space; chosen by authors; controls expressiveness of signatures.
  • Reranking weight λ = 0.1
    Scalar strength of collaborative residual; fixed across all arms and reported as not tuned on test, but no validation procedure is described.
  • Bridge smoothing μ = not reported
    Smoothing constant toward the global factor prior in Eq. 3; value omitted.
  • Affinity smoothing ϵ = not reported
    Numerical floor inside r(·) = λ log(max(q^T sig,0)+ϵ); value omitted.
  • Beam budgets M_a, M_b = 20, 20 (top-40 prefixes)
    Inference-time candidate budgets; fixed across methods; affect the reachability ceiling.
  • HPF prior/VI hyperparameters = not reported
    Gamma priors and coordinate-ascent iterations for HPF-CA are not specified; needed for exact replication.
axioms (6)
  • domain assumption Compact content-derived SIDs under-express collaborative signal; content and collaborative objectives compete (DiscRec, Liu et al. 2025).
    Central motivation; cited, not re-derived; the entire method exists to supply this missing signal.
  • domain assumption The second-order item–item co-view matrix C (Eq. 1), counting unordered pairs within each user's full history, captures audience structure rather than popularity or session artifacts.
    If C is dominated by popularity, the HPF factors will encode a global frequency effect; the Pop control partially checks this, but the premise is not independently established.
  • domain assumption HPF Gamma–Poisson factorization of C yields item factors whose per-prefix averages (Eq. 2) are discriminative at the (s_a,s_b) level.
    The method's positive result is representation-specific: TF-IDF+SVD and K-means controls fail, so the gain depends on HPF being the right decomposition.
  • ad hoc to paper LLM-summarized audience tags plus the bridge P(f|t) (Eq. 3) reconstruct the user's behavioral factor query well enough to preserve accuracy.
    Core premise of the NL channel; the paper's own user-differentiation result is real but small (d≈0.25), so the premise holds only weakly.
  • domain assumption The frozen backbone's in-pool candidates contain the target prefix often enough that reranking can recover it.
    The method's ceiling is bounded by backbone reachability; the authors acknowledge this in Limitations, and it is not established for other backbones/datasets.
  • domain assumption The GT-disjoint split (826 of 1,000 RecIF cases) and training-only co-view statistics make evaluation leakage-free.
    Holding out cases whose target was a training prediction target reduces memorization; target items may still appear in training histories, so the co-view signal can legitimately inform the query.

pith-pipeline@v1.3.0-daily-deepseek · 11507 in / 19841 out tokens · 184115 ms · 2026-08-01T03:23:46.018622+00:00 · methodology

0 comments
read the original abstract

Making LLM-based generative recommendation models stronger and more personalized through natural language and explicit reasoning is a widely anticipated yet still unsolved goal. Such models cast recommendation as autoregressively generating an item's semantic-ID (SID), a short tuple of discrete codes, so that recommending well reduces to emitting the right SID. In this setting the model verbalizes its knowledge poorly, and text and SID tokens live in misaligned embedding spaces. Deep reasoning therefore rarely turns into a correct SID, and enabling explicit "thinking" often gives no gain or even hurts. The deeper cause is that a compact SID cannot hold content and collaborative signal at once: the two compete, and collaboration loses. Because a mis-predicted SID is a wrong recommendation, this caps accuracy directly. Costly multi-round training barely helps, and few methods try to supply the missing signal at inference time. What is missing is a reliable channel that carries collaborative signal into SID generation. We therefore propose a framework, guided by personalized natural language, that adds hierarchical collaborative cues as the model generates, without altering the backbone or retraining the SIDs. Rather than mapping language onto SIDs directly, it uses language to attach analyzable links between collaborative patterns and their audiences, restoring the collaborative signal that SIDs miss. The result is consistent gains in recommendation accuracy, grounding generation in collaborative structure at inference time rather than relying on explicit reasoning or retraining.

Figures

Figures reproduced from arXiv: 2607.27682 by Bowei He, Changjiang Han, Jikun Kang, Pinghua Gong, Qingyang Li, Xue Liu, Yaqiang Zang.

Figure 1
Figure 1. Figure 1: Full-SID hit@K on OneRec-1.7B/8B for direct decoding (no_think), native reasoning (think), and our reranking (Ours); all share the frozen backbone and de￾coding budget. A hit requires generating the target’s exact three-level ID (sa, sb, sc) within top K. Enabling reasoning lowers accuracy while Ours raises it. sonalized, much as chain-of-thought has advanced language tasks (Wei et al. 2022). That expectat… view at source ↗
Figure 2
Figure 2. Figure 2: PCA of text-token vs. three-level SID-token em [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview. Offline (top): co-view counts C are factorized with HPF into item factors ϕi ; an LLM turns each history into audience/behaviour tags; a bridge P(f | t) maps tags into the factor space and sums them into a user vector qu. Inference (bottom): over a frozen backbone, qu enters each level differently without widening the beam, expanding sa candidates, reranking sb, and inventory-masking sc. Coral ma… view at source ↗
Figure 4
Figure 4. Figure 4: Natural-language increment in the HPF factor space [PITH_FULL_IMAGE:figures/full_fig_p007_4.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

40 extracted references · 3 linked inside Pith

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  9. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  10. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  11. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  12. [12]

    and Vu, Trung and Heldt, Lukasz and Hong, Lichan and Tay, Yi and Tran, Vinh Q

    Rajput, Shashank and Mehta, Nikhil and Singh, Anima and Keshavan, Raghunandan H. and Vu, Trung and Heldt, Lukasz and Hong, Lichan and Tay, Yi and Tran, Vinh Q. and Samost, Jonah and Kula, Maciej and Chi, Ed H. and Sathiamoorthy, Maheswaran. Recommender Systems with Generative Retrieval. Advances in Neural Information Processing Systems (NeurIPS)

  13. [13]

    OneRec : Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment

    Deng, Jiaxin and Wang, Shiyao and Cai, Kuo and Ren, Lejian and Hu, Qigen and Ding, Weifeng and Luo, Qiang and Zhou, Guorui. OneRec : Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment. arXiv:2502.18965

    Pith/arXiv arXiv

  14. [14]

    Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation

    Zheng, Bowen and Hou, Yupeng and Lu, Hongyu and Chen, Yu and Zhao, Wayne Xin and Chen, Ming and Wen, Ji-Rong. Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation. IEEE 40th International Conference on Data Engineering (ICDE)

  15. [15]

    Learnable Item Tokenization for Generative Recommendation

    Wang, Wenjie and Bao, Honghui and Lin, Xinyu and Zhang, Jizhi and Li, Yongqi and Feng, Fuli and Ng, See-Kiong and Chua, Tat-Seng. Learnable Item Tokenization for Generative Recommendation. Proceedings of the 33rd ACM International Conference on Information and Knowledge Management (CIKM)

  16. [16]

    DiscRec : Disentangled Semantic-Collaborative Modeling for Generative Recommendation

    Liu, Chang and Bai, Yimeng and Zhao, Xiaoyan and Zhang, Yang and Feng, Fuli and Rong, Wenge. DiscRec : Disentangled Semantic-Collaborative Modeling for Generative Recommendation. arXiv:2506.15576

    Pith/arXiv arXiv

  17. [17]

    Implicit Reasoning for Large Language Model-based Generative Recommendation

    He, Yinhan and Collins, Liam and Kumar, Bhuvesh and Li, Jundong and Shah, Neil and Loveland, Donald. Implicit Reasoning for Large Language Model-based Generative Recommendation. arXiv:2606.14142

    arXiv

  18. [18]

    OneRec-Think : In-Text Reasoning for Generative Recommendation

    Liu, Zhanyu and Wang, Shiyao and Wang, Xingmei and Zhang, Rongzhou and Deng, Jiaxin and Bao, Honghui and Zhang, Jinghao and Li, Wuchao and Zheng, PengFei and Wu, Xiangyu and Hu, Yifei and Hu, Qigen and Luo, Xinchen and Ren, Lejian and Zhang, Zixing and Wang, Qianqian and Cai, Kuo and Wu, Yunfan and Cheng, Hongtao and Cheng, Zexuan and Ren, Lu and Wang, Hu...

  19. [19]

    OneReason Technical Report

    OneRec Team. OneReason Technical Report. arXiv:2606.06260

    Pith/arXiv arXiv

  20. [20]

    and Blei, David M

    Gopalan, Prem and Hofman, Jake M. and Blei, David M. Scalable Recommendation with Hierarchical Poisson Factorization. Proceedings of the 31st Conference on Uncertainty in Artificial Intelligence (UAI)

  21. [21]

    Rethinking LLM -Based Recommendations: A Personalized Query-Driven Parallel Integration

    Han, Donghee and Song, Hwanjun and Yi, Mun Yong. Rethinking LLM -Based Recommendations: A Personalized Query-Driven Parallel Integration. Findings of the Association for Computational Linguistics: EMNLP 2025

    2025

  22. [22]

    Language Representations Can be What Recommenders Need: Findings and Potentials

    Sheng, Leheng and Zhang, An and Zhang, Yi and Chen, Yuxin and Wang, Xiang and Chua, Tat-Seng. Language Representations Can be What Recommenders Need: Findings and Potentials. International Conference on Learning Representations (ICLR)

  23. [23]

    Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting

    Turpin, Miles and Michael, Julian and Perez, Ethan and Bowman, Samuel R. Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting. Advances in Neural Information Processing Systems (NeurIPS)

  24. [24]

    OpenOneRec Technical Report

    OneRec Team. OpenOneRec Technical Report. arXiv:2512.24762

    arXiv

  25. [25]

    Self-Attentive Sequential Recommendation

    Kang, Wang-Cheng and McAuley, Julian J. Self-Attentive Sequential Recommendation. IEEE International Conference on Data Mining (ICDM)

  26. [26]

    BERT4Rec : Sequential Recommendation with Bidirectional Encoder Representations from Transformer

    Sun, Fei and Liu, Jun and Wu, Jian and Pei, Changhua and Lin, Xiao and Ou, Wenwu and Jiang, Peng. BERT4Rec : Sequential Recommendation with Bidirectional Encoder Representations from Transformer. ACM International Conference on Information and Knowledge Management (CIKM)

  27. [27]

    Session-based Recommendations with Recurrent Neural Networks

    Hidasi, Bal\'azs and Karatzoglou, Alexandros and Baltrunas, Linas and Tikk, Domonkos. Session-based Recommendations with Recurrent Neural Networks. International Conference on Learning Representations (ICLR)

  28. [28]

    BPR : Bayesian Personalized Ranking from Implicit Feedback

    Rendle, Steffen and Freudenthaler, Christoph and Gantner, Zeno and Schmidt-Thieme, Lars. BPR : Bayesian Personalized Ranking from Implicit Feedback. Conference on Uncertainty in Artificial Intelligence (UAI)

  29. [29]

    and Volinsky, Chris

    Koren, Yehuda and Bell, Robert M. and Volinsky, Chris. Matrix Factorization Techniques for Recommender Systems. Computer

  30. [30]

    Neural Collaborative Filtering

    He, Xiangnan and Liao, Lizi and Zhang, Hanwang and Nie, Liqiang and Hu, Xia and Chua, Tat-Seng. Neural Collaborative Filtering. International Conference on World Wide Web (WWW)

  31. [31]

    Recommendation as Language Processing ( RLP ): A Unified Pretrain, Personalized Prompt & Predict Paradigm ( P5 )

    Geng, Shijie and Liu, Shuchang and Fu, Zuohui and Ge, Yingqiang and Zhang, Yongfeng. Recommendation as Language Processing ( RLP ): A Unified Pretrain, Personalized Prompt & Predict Paradigm ( P5 ). ACM Conference on Recommender Systems (RecSys)

  32. [32]

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

    Bao, Keqin and Zhang, Jizhi and Zhang, Yang and Wang, Wenjie and Feng, Fuli and He, Xiangnan. TALLRec : An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation. ACM Conference on Recommender Systems (RecSys)

  33. [33]

    LLaRA : Large Language-Recommendation Assistant

    Liao, Jiayi and Li, Sihang and Yang, Zhengyi and Wu, Jiancan and Yuan, Yancheng and Wang, Xiang and He, Xiangnan. LLaRA : Large Language-Recommendation Assistant. International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR)

  34. [34]

    CoLLM : Integrating Collaborative Embeddings Into Large Language Models for Recommendation

    Zhang, Yang and Feng, Fuli and Zhang, Jizhi and Bao, Keqin and Wang, Qifan and He, Xiangnan. CoLLM : Integrating Collaborative Embeddings Into Large Language Models for Recommendation. IEEE Transactions on Knowledge and Data Engineering

  35. [35]

    Text Is All You Need: Learning Language Representations for Sequential Recommendation

    Li, Jiacheng and Wang, Ming and Li, Jin and Fu, Jinmiao and Shen, Xin and Shang, Jingbo and McAuley, Julian J. Text Is All You Need: Learning Language Representations for Sequential Recommendation. ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)

  36. [36]

    A Survey on Large Language Models for Recommendation

    Wu, Likang and Zheng, Zhi and Qiu, Zhaopeng and Wang, Hao and Gu, Hongchao and Shen, Tingjia and Qin, Chuan and Zhu, Chen and Zhu, Hengshu and Liu, Qi and Xiong, Hui and Chen, Enhong. A Survey on Large Language Models for Recommendation. World Wide Web

  37. [37]

    Autoregressive Image Generation using Residual Quantization

    Lee, Doyup and Kim, Chiheon and Kim, Saehoon and Cho, Minsu and Han, Wook-Shin. Autoregressive Image Generation using Residual Quantization. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  38. [38]

    and Sathiamoorthy, Maheswaran and Zheng, Yilin and Hong, Lichan and Heldt, Lukasz and Wei, Li and Tandon, Devansh and Chi, Ed H

    Singh, Anima and Vu, Trung and Mehta, Nikhil and Keshavan, Raghunandan H. and Sathiamoorthy, Maheswaran and Zheng, Yilin and Hong, Lichan and Heldt, Lukasz and Wei, Li and Tandon, Devansh and Chi, Ed H. and Yi, Xinyang. Better Generalization with Semantic IDs : A Case Study in Ranking for Recommendations. ACM Conference on Recommender Systems (RecSys)

  39. [39]

    and Le, Quoc V

    Wei, Jason and Wang, Xuezhi and Schuurmans, Dale and Bosma, Maarten and Ichter, Brian and Xia, Fei and Chi, Ed H. and Le, Quoc V. and Zhou, Denny. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Advances in Neural Information Processing Systems (NeurIPS)

  40. [40]

    and Chi, Ed H

    Wang, Xuezhi and Wei, Jason and Schuurmans, Dale and Le, Quoc V. and Chi, Ed H. and Narang, Sharan and Chowdhery, Aakanksha and Zhou, Denny. Self-Consistency Improves Chain of Thought Reasoning in Language Models. International Conference on Learning Representations (ICLR)