Pith.

Information Retrieval · arXiv:2610.34124 v1 · submitted 25 Oct 2026 · 59 references

Decision-Native Recommendation with Large Language Models

Daoxuan Liu, Jiayi Liao, Jiancan Wu, An Zhang, Xiang Wang, Xiangnan He

The paper says

“Large language models (LLMs) bring world knowledge to user preference modeling in sequential recommendation, but generation-based recommenders express next-item predictions indirectly as text.”

more of the abstract

Generated outputs may fail to identify a unique candidate item, depend on candidate presentation order in the prompt, and lack explicit confidence in the recommendation. Recent studies explore Jev, a model for fast option selection with explicit confidence. Inspired by Jev, we introduce decision-native recommendation, a new paradigm that jointly models item choice and recommendation confidence using LLM representations. It requires valid candidate selection, consistency under candidate reordering, and confidence that reflects recommendation accuracy. We propose Direct Item Choice Estimation (DICE), which learns a personalized next-item distribution over supplied candidates from user interaction history and recommends the highest-scoring item with calibrated confidence. Tree-structured attention shares history computation across isolated candidate branches for order-consistent candidate scoring in a single forward pass. On three recommendation datasets, DICE achieves higher top-one accuracy than all evaluated baselines. Separate evaluations show unchanged item choices under tested candidate reorderings and end-to-end speedups of up to approximately 6 × over the evaluated generation-based methods. Its confidence estimates support selective recommendation by identifying more reliable predictions and guide additional computation for low-confidence examples in offline routing.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 27 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. Solid empirical paper on direct candidate scoring with LLMs, but the central accuracy comparison is partly confounded by unequal fine-tuning and missing a controlled ablation. This line is wrong

2 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 59 references say.

The paper says, at §4.1 (Baselines) and Table 1

The review did not quote this passage.

Pith says: Objection 1 of 2 · major

The comparison with direct-ranking baselines LlamaRec [50] and InvariRank [5] is potentially unfair because these methods are used in their original frozen form, while DICE adds LoRA adapters (rank 8) and a trained decision head. The paper states that 'All LLM methods use Qwen3-0.6B' and 'baselines follow their original training protocols' (Appendix A.3), but does not confirm whether LlamaRec/InvariRank were adapted with LoRA or otherwise fine-tuned.

more of objection 1

If they are frozen, the observed HR@1 gains (e.g., +6.14% on Goodreads) could be attributed to fine-tuning rather than to the decision-native approach itself. The paper should either (a) fine-tune these baselines under the same LoRA configuration, (b) report DICE with a frozen backbone (head-only) as a control, or (c) explicitly acknowledge this confound and adjust the central claim. Without this, the statement that 'DICE outperforms all evaluated baselines' in top-one accuracy is not conclusively supported.

The paper says, at §3.2, Eq. (10)

The review did not quote this passage.

Pith says: Objection 2 of 2 · major

Equation (10) claims exact equality (r^{tree}_i = r^{ind}_i) between tree-structured and independent encoding. The paper later acknowledges residual probability differences (Appendix A.8) and tests only 100 examples per dataset. While the empirical check is reasonable, the strong theoretical claim should be softened to reflect that the equality holds only in exact arithmetic due to shared prefix and position indices, with numerical deviations in practice.

more of objection 2

This is particularly important because the order-consistency guarantee (Eqs. (12)–(13)) and the speedup rely on this equivalence. A brief note in §3.2 would improve clarity.

5 minor objections

The paper says, at Abstract and §1

Pith says: minor objection 1 of 5

The abstract introduces Jev as a model for fast option selection, but the only citation is a blog post [1]. The paper would benefit from clarifying the status of Jev (e.g., as a system-one model) and perhaps adding a citation to a peer-reviewed publication if one exists; otherwise the reference is informal.

The paper says, at §3.2, Fig. 2

Pith says: minor objection 2 of 5

The figure caption and panel are somewhat dense. Consider adding a label for the 'tree mask: branches isolated' to make the attention pattern more immediately interpretable.

show 3 more

The paper says, at Table 2

Pith says: minor objection 3 of 5

The table shows HR@1, HR@3, NDCG@5 for backbone scaling. For completeness, include HR@5 to align with Table 1, or explain its omission.

The paper says, at Appendix A.8

Pith says: minor objection 4 of 5

The flat–tree agreement is reported as 100% on 100 examples. Since the paper acknowledges residual probability differences, clarify whether the 100% agreement refers to item choice or to exact match of logits. If it is only choice agreement, that is already sufficient; but the phrasing should be precise.

The paper says, at §4.3, Table 5

Pith says: minor objection 5 of 5

The FlipRate for SFT and S-DPO is computed over 20 reorderings per example. The paper states that 'nonunique generated answers are denoted ⊥' and transitions involving ⊥ count as flips. This is appropriate, but consider reporting ValidRatio separately for the original order to give a complete picture.

What would settle it

Fine-tune LlamaRec's verbalizer-based scoring (candidate-letter logits) with LoRA adapters on the same training data and cross-entropy objective used for DICE (keeping the backbone frozen, training only LoRA and the verbalizer head). Report HR@1 on Goodreads and LastFM. If fine-tuned LlamaRec achieves HR@1 within 0.5% of DICE-I, then DICE's claimed accuracy advantage reduces to a fine-tuning benefit rather than a paradigm improvement.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 27 Oct 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic partly agrees 58 s · 22,960 tokens read, 5,704 written
  3. Editor wrote the verdict above 19 s · 23,816 tokens read, 1,639 written
  4. Referee major revision: 2 major objections, 5 minor 23 s · 23,234 tokens read, 2,559 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The paper introduces decision-native recommendation and DICE, a clean method that scores candidate items directly from LLM representations rather than generating text. The idea is sensible and well-motivated: generation-based recommenders suffer from invalid outputs, order sensitivity, and no confidence. DICE addresses all three with a simple architecture—shared history encoding, LoRA adapters, a decision head, and temperature scaling. The empirical work is thorough: three datasets, multiple baselines, proper splits, standard deviations reported, and separate analyses for order consistency, latency, calibration, and routing. The code and data release is a plus.

The paper's main claim is that DICE outperforms all evaluated baselines in top-one accuracy. That claim is numerically true in Table 1, but there is a load-bearing confound that the authors do not adequately address. The direct-ranking baselines (LlamaRec, InvariRank) are used as originally published, which means they are frozen LLMs. DICE, by contrast, adds LoRA adapters (rank 8) and a trained decision head, updating about 1–2% of backbone parameters. The paper provides no controlled ablation where those baselines are also fine-tuned with LoRA under the same training budget. Without that, the advantage could stem from additional trainable capacity rather than the decision-native paradigm itself. This is the stress-tester's point and it is a fair one.

The cold-start evaluation also has a known limitation: items may have appeared in LLM pretraining. The authors acknowledge this but do not attempt to isolate it. The confidence routing uses fixed thresholds and a single temperature, and the paper does not show sensitivity to these choices. These are minor compared to the confound above.

Who is this for? Researchers working on LLM-based recommendation will want to read it, and the method is practical enough that practitioners could adopt it. The paper deserves serious peer review, but the authors should be asked to provide a controlled comparison where direct-ranking baselines receive the same LoRA fine-tuning. That would settle whether the paradigm itself drives the improvement.

Recommendation: accept for peer review. The core idea is timely, the experiments are otherwise rigorous, and the confound is addressable with additional experiments.

The strongest objection

The central accuracy claim (Table 1) rests on a comparison that may be unfair for the most relevant baselines. The paper states that 'All LLM methods use Qwen3-0.6B' and that baselines follow their original training protocols (Appendix A.3). LlamaRec [50] originally scores candidates via letter logits from a frozen LLM without fine-tuning; InvariRank [5] may also use a frozen model. DICE, by contrast, adds LoRA adapters (rank 8) and a learned decision head, training three epochs on recommendation data. If LlamaRec and InvariRank are used without any parameter update, then DICE's advantage (e.g., +6.14% HR@1 over LlamaRec on Goodreads) could be due to fine-tuning rather than the decision-native paradigm itself. The paper does not provide a controlled ablation where these baselines are also fine-tuned under the same LoRA setup. Without such a comparison, the claim 'DICE outperforms all evaluated baselines' (including direct-ranking methods) is not conclusively supported.

The reader identified the tree-equivalence assumption as the weakest methodological link, but that concern is mitigated by 100% item-choice agreement across 100 examples per run and dataset. A more load-bearing threat to the central accuracy claim is the inequitable comparison with direct-ranking baselines that may not have been fine-tuned. The paper includes fine-tuned generation-based baselines (SFT, S-DPO, NAPO) and outperforms them, which supports DICE's strength. However, the most tightly related methods (LlamaRec, InvariRank) are the ones that share the same 'score candidates directly' approach, and they are likely used without adaptation. If the advantage over those methods disappears after equal fine-tuning, the paper's headline claim is misleading. The test proposed would settle this cleanly. Nonetheless, the paper's overall evidence is strong enough that the verdict need not change; the CONDITIONAL verdict already flags that some conditions are unmet, and this concern reinforces the need for clearer baseline descriptions rather than invalidating the core method. I therefore recommend UNCHANGED but note that the paper should add this control in revision.

The referee's report

If substantiated, the decision-native paradigm could offer a principled way to combine world knowledge from LLMs with the categorical decision form required in recommendation, addressing known issues of output validity, order sensitivity, and lack of confidence in generation-based recommenders. The architecture and empirical results are of potential interest to the RecSys and IR communities. However, the central accuracy claim is partially weakened by an unequal comparison with direct-ranking baselines, as detailed below.

Pith's reading of the paper (a machine's summary, not the authors')

This paper argues that using large language models (LLMs) for recommendation does not require text generation. Instead, the authors propose a new paradigm called decision-native recommendation, where the LLM directly scores each candidate item and selects the highest-scoring one, simultaneously providing a calibrated confidence estimate. They instantiate this paradigm with Direct Item Choice Estimation (DICE), which uses tree-structured attention to share computation across candidates in a single forward pass while guaranteeing that scores do not depend on the order candidates are presented. On three public datasets, DICE achieves higher top-one accuracy than all compared baselines, including both generation-based and direct-ranking methods, while also offering end-to-end speedups of up to 6× and producing confidence values that reliably indicate whether the recommendation is likely to be correct.

What carries the argument. Tree-structured attention with shared prefix and isolated candidate branches, combined with a shared scoring head. The prefix encodes the user history; each candidate forms a separate branch that attends to the prefix and itself but not to other branches. Position indices per branch match those of independent encoding, so each candidate's representation is theoretically identical to what would be obtained from independent encoding, guaranteeing permutation equivariance. This design allows all candidate scores to be computed in a single LLM forward pass while preserving order consistency.

Core claim. The central claim is that recommendation can be reformulated as a categorical choice problem over supplied candidates using LLM representations, eliminating the need for token-by-token generation. The paper introduces DICE, which encodes the user's interaction history together with each candidate's description through a shared LLM, applies a shared scoring head to produce candidate logits, and normalizes these into a choice distribution. Critically, DICE employs tree-structured attention—where the history is a shared prefix and each candidate is an isolated branch—so that all candidate scores are computed in one forward pass and are permutation-equivariant: reordering the candidates merely permutes their scores without changing the values. The selected item's softmax probability is then calibrated via temperature scaling to serve as a decision confidence. Experiments show that DICE achieves higher Hit Rate@1 than all 13 evaluated baselines across Goodreads, LastFM, and Amazon Industrial datasets, maintains 100% item-choice agreement under candidate reordering, and provides well-calibrated confidence that enables selective recommendation and efficient model routing.

The premise it rests on. Tree-structured attention produces representations for each candidate that are exactly identical to those from independent encoding, which is necessary for the claimed permutation equivariance and for sharing computation without altering scores; the paper acknowledges small residual probability differences (App. A.8), implying numerical approximation.

If the paper is right (Pith's inferences, not the authors')

LLM-based recommenders can abandon text generation entirely, avoiding mismatched or ambiguous outputs and reducing inference latency.

Recommendation systems can use DICE's calibrated confidence to selectively present recommendations only when confidence is high, improving user trust.

Confidence-guided routing allows a small model to handle most queries and defer only low-confidence ones to a larger model, cutting computational cost while retaining high accuracy.

DICE's ability to score unseen items from their text descriptions supports cold-start recommendation without retraining.

The tree-structured attention pattern can be applied to other ranking and selection tasks that require order-consistent scoring over a variable set of candidates.

The paradigm could be extended to joint retrieval and selection, where the candidate pool itself is generated or expanded dynamically, and confidence informs whether the true item might be missing.

Because DICE is order-consistent by construction, it eliminates position bias that plagues generation-based and listwise recommenders, potentially improving fairness across items.

The tree-sharing idea generalizes to any scenario where a fixed context is scored against many options—for example, multiple-choice question answering or document reranking—and may inspire similar efficiency gains.

The paper's confidence calibration uses a single learned temperature; future work might adaptively adjust calibration per user or per item category as distributions shift.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Large language models (LLMs) bring world knowledge to user preference modeling in sequential recommendation, but generation-based recommenders express next-item predictions indirectly as text. Generated outputs may fail to identify a unique candidate item, depend on candidate presentation order in the prompt, and lack explicit confidence in the recommendation. Recent studies explore Jev, a model for fast option selection with explicit confidence. Inspired by Jev, we introduce decision-native recommendation, a new paradigm that jointly models item choice and recommendation confidence using LLM representations. It requires valid candidate selection, consistency under candidate reordering, and confidence that reflects recommendation accuracy. We propose Direct Item Choice Estimation (DICE), which learns a personalized next-item distribution over supplied candidates from user interaction history and recommends the highest-scoring item with calibrated confidence. Tree-structured attention shares history computation across isolated candidate branches for order-consistent candidate scoring in a single forward pass. On three recommendation datasets, DICE achieves higher top-one accuracy than all evaluated baselines. Separate evaluations show unchanged item choices under tested candidate reorderings and end-to-end speedups of up to approximately 6 × over the evaluated generation-based methods. Its confidence estimates support selective recommendation by identifying more reliable predictions and guide additional computation for low-confidence examples in offline routing.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-27

The authors' main claim, as the classifier picked it out 0.97

“On three recommendation datasets, DICE achieves higher top-one accuracy than all evaluated baselines.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.10 · states limitation 0.13 · human participants 0.06 · controlled comparison 0.89 · randomized assignment 0.02 · states quantitative result 0.98 · compares with prior results 0.97 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Large language models (LLMs) bring world knowledge to user preference modeling in sequential recommendation, but generation-based recommenders express next-item predictions indirectly as text. background 1.00 0.06 0.03 asserted 1.00
Generated outputs may fail to identify a unique candidate item, depend on candidate presentation order in the prompt, and lack explicit confidence in the recommendation. background 0.53 0.10 0.03 hedged possibility 0.99
Recent studies explore Jev, a model for fast option selection with explicit confidence. background 1.00 0.04 0.03 asserted 0.98
Inspired by Jev, we introduce decision-native recommendation, a new paradigm that jointly models item choice and recommendation confidence using LLM representations. contribution 1.00 0.87 0.17 asserted 0.99
It requires valid candidate selection, consistency under candidate reordering, and confidence that reflects recommendation accuracy. contribution 0.60 0.13 0.08 asserted 0.97
We propose Direct Item Choice Estimation (DICE), which learns a personalized next-item distribution over supplied candidates from user interaction history and recommends the highest-scoring item with calibrated confidence. method 0.83 0.87 0.17 asserted 0.99
Tree-structured attention shares history computation across isolated candidate branches for order-consistent candidate scoring in a single forward pass. method 0.99 0.78 0.04 asserted 1.00
On three recommendation datasets, DICE achieves higher top-one accuracy than all evaluated baselines. result 1.00 0.95 0.53 asserted 1.00
Separate evaluations show unchanged item choices under tested candidate reorderings and end-to-end speedups of up to approximately $6\times$ over the evaluated generation-based methods. result 1.00 0.94 0.18 asserted 1.00
Its confidence estimates support selective recommendation by identifying more reliable predictions and guide additional computation for low-confidence examples in offline routing. interpretation 0.53 0.90 0.16 asserted 0.98

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.96

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-27: 1 request, 7,457 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (6)
Figure 1 of the paper
Figure 1. Generation-based recommendation and DICE for
Figure 2 of the paper
Figure 2. DICE-I scoring and confidence-guided routing. Icons represent item texts. Mask rows attend to columns, with causal
Figure 3 of the paper
Figure 3. Mean end-to-end latency (𝐾 = 20, 200 examples per run and dataset). All paths include tokenization. Flat/Tree de￾note DICE-I layouts; arrows give flat-to-tree speedups. Lower is better. retains an advantage across candidate counts. Both DICE vari￾ants exceed SFT, S-DPO, and NAPO at every tested candidate count (
Figure 4 of the paper
Figure 4. Mean HR@1 across candidate counts (𝐾 includes the target). The horizontal axis is logarithmic. 4.4 RQ3: Decision Confidence and Applications We compare confidence calibration across methods (
Figure 5 of the paper
Figure 5. DICE-I confidence. (a) LastFM calibration before and
Figure 6 of the paper
Figure 6. DICE-I routing on LastFM: mean accuracy versus

Figures are the authors', reproduced from the paper.

References (59)
  1. [50] Zhenrui Yue, Sara Rabhi, Gabriel de Souza Pereira Moreira, Dong Wang, and Even Oldridge. 2023. LlamaRec: Two-Stage Recommendation using Large Language Models for Ranking. https://arxiv.org/abs/2311.02089 arXiv:2311.02089 on PithPith's reviewarXiv
  2. [5] Ethan Bito, Yongli Ren, and Estrid He. 2026. One Pass, Any Order: Position- Invariant Listwise Reranking for LLM-Based Recommendation. InSIGIR. ACM, New York, NY, USA, 3625–3629. doi:10.1145/3805712.3809952 on Pithdoi:10.1145/3805712.3809952
  3. [1] Diogo Almeida. 2026. Introducing System One Models & Jev. https://typesafe. ai/blog/introducing-system-one-models-and-jev TypeSafe AI, September 15, 2026 on Pith
  4. [2] Keqin Bao, Jizhi Zhang, Wenjie Wang, Yang Zhang, Zhengyi Yang, Yanchen Luo, Chong Chen, Fuli Feng, and Qi Tian. 2025. A Bi-Step Grounding Paradigm for Large Language Models in Recommendation Systems.ACM Transactions on Recommender Systems3, 4 (2025), 53:1–53:27. doi:10.1145/3716393 on Pithdoi:10.1145/3716393
  5. [3] Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He on Pith
  6. [4] Ethan Bito, Yongli Ren, and Estrid He. 2025. Evaluating Position Bias in Large Language Model Recommendations. https://arxiv.org/abs/2508.02020 arXiv:2508.02020 on PithPith's reviewarXiv
  7. [6] Ethan Bito, Yongli Ren, and Estrid He. 2026. Position Bias Undermines Preference Consistency in Listwise LLM-Based Reranking. InRecSys. ACM, New York, NY, USA, 728–732. doi:10.1145/3773078.3831801 on Pithdoi:10.1145/3773078.3831801
  8. [7] Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D. Lee, Deming Chen, and Tri Dao. 2024. Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads. InICML, Vol. 235. PMLR, Vienna, Austria, 5209– on Pith
  9. [8] Lingjiao Chen, Matei Zaharia, and James Zou. 2024. FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance.Trans. Mach. Learn. Res.2024 (2024). https://openreview.net/forum?id=cSimKw5p6R on Pith
  10. [9] Yuxin Chen, Junfei Tan, An Zhang, Zhengyi Yang, Leheng Sheng, Enzhi Zhang, Xiang Wang, and Tat-Seng Chua. 2024. On Softmax Direct Preference Optimiza- tion for Recommendation. InNeurIPS, Vol. 37. Curran Associates, Inc., Vancouver, BC, Canada, 27463–27489. doi:10.52202/079017-0863 on Pithdoi:10.52202/079017-0863
  11. [10] Zhi-Yuan Chen, Siyu Lu, Qiang Liu, Xingxing Wang, and Yankai Lin. 2026. From Prediction to Understanding: Leveraging Reasoning in Large Language Model- based Recommendations. InWWW. Association for Computing Machinery, New York, NY, USA, 6171–6182. doi:10.1145/3774904.3792283 on Pithdoi:10.1145/3774904.3792283
  12. [11] Shrey Desai and Greg Durrett. 2020. Calibration of Pre-trained Transformers. In EMNLP (1). Association for Computational Linguistics, Stroudsburg, PA, USA, 295–302. doi:10.18653/v1/2020.emnlp-main.21 on Pithdoi:10.18653/v1/2020.emnlp-main.21
  13. [12] Chenlu Ding, Daoxuan Liu, Jiancan Wu, Xingyu Hu, Junkang Wu, Haitao Wang, Yongkang Wang, Xingxing Wang, and Xiang Wang. 2025. On Negative-aware Preference Optimization for Recommendation. https://arxiv.org/abs/2508.09653 arXiv:2508.09653 on PithPith's reviewarXiv
  14. [13] Yonatan Geifman and Ran El-Yaniv. 2017. Selective Classification for Deep Neural Networks. InNIPS. Curran Associates, Inc., Long Beach, CA, USA, 4878–4887. https://proceedings.neurips.cc/paper/2017/hash/ 4a8423d5e91fda00bb7e46540e2b0cf1-Abstract.html on Pith
  15. [14] Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Person- alized Prompt & Predict Paradigm (P5). InRecSys. Association for Computing Machinery, New York, NY, USA, 299–315. doi:10.1145/... on Pithdoi:10.1145/3523227.3546767
  16. [15] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On Calibration of Modern Neural Networks. InICML. PMLR, Sydney, Australia, 1321–1330. https://proceedings.mlr.press/v70/guo17a.html on Pith
  17. [16] Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. In IJCAI. ijcai.org, California, 1725–1731. doi:10.24963/ijcai.2017/239 on Pithdoi:10.24963/ijcai.2017/239
  18. [17] Huizhong Guo, Tianjun Wei, Dongxia Wang, Yingpeng Du, Ziyan Wang, Jie Zhang, and Zhu Sun. 2026. Think When Needed: Model-Aware Reasoning Routing for LLM-based Ranking. InSIGIR. ACM, New York, NY, USA, 471–481. doi:10.1145/3805712.3809709 on Pithdoi:10.1145/3805712.3809709
  19. [18] Neha Gupta, Harikrishna Narasimhan, Wittawat Jitkrittum, Ankit Singh Rawat, Aditya Krishna Menon, and Sanjiv Kumar. 2024. Language Model Cascades: Token-Level Uncertainty And Beyond. InICLR. OpenReview.net, Vienna, Austria, 4147–4180. https://openreview.net/forum?id=KgaBScZ4VI on Pith
  20. [19] Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk on Pith
show 39 more
  1. [20] Yupeng Hou, Shanlei Mu, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. 2022. Towards Universal Sequence Representation Learning for Recom- mender Systems. InKDD. ACM, New York, NY, USA, 585–593. doi:10.1145/ 3534678.3539381 on Pith
  2. [21] Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian J. McAuley, and Wayne Xin Zhao. 2024. Large Language Models are Zero-Shot Rankers for Recommender Systems. InECIR (2). Springer, Cham, 364–381. doi:10.1007/978- 3-031-56060-6_24 on Pithdoi:10.1007/978-
  3. [22] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InICLR. OpenReview.net, Virtual. https://openreview.net/ forum?id=nZeVKeeFYf9 on Pith
  4. [23] Jianchao Ji, Zelong Li, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Juntao Tan, and Yongfeng Zhang. 2024. GenRec: Large Language Model for Generative Recommendation. InECIR (3). Springer, Cham, 494–502. doi:10.1007/978-3-031- 56063-7_42 on Pithdoi:10.1007/978-3-031-
  5. [24] Chumeng Jiang, Jiayin Wang, Xinjie Lin, Zhiqiang Guo, Hengliang Luo, and Min Zhang. 2026. SelfDR: Self-Distillation from Reasoning for LLM-Based Recom- mendation. https://arxiv.org/abs/2609.03313 arXiv:2609.03313 on PithPith's reviewarXiv
  6. [25] Dongming Jiang, Yi Li, and Bingzhe Li. 2026. Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents. https://arxiv.org/abs/2609.23986 arXiv:2609.23986 on PithPith's reviewarXiv
  7. [26] Amita Kamath, Robin Jia, and Percy Liang. 2020. Selective Question Answer- ing under Domain Shift. InACL. Association for Computational Linguistics, Stroudsburg, PA, USA, 5684–5696. doi:10.18653/v1/2020.acl-main.503 on Pithdoi:10.18653/v1/2020.acl-main.503
  8. [27] Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recom- mendation. InICDM. IEEE, Singapore, 197–206. doi:10.1109/ICDM.2018.00035 on Pithdoi:10.1109/icdm.2018.00035
  9. [28] Sein Kim, Hongseok Kang, Seungyoon Choi, Donghyun Kim, Minchul Yang, and Chanyoung Park. 2024. Large Language Models meet Collaborative Filtering: An Efficient All-round LLM-based Recommender System. InKDD. ACM, New York, NY, USA, 1395–1406. doi:10.1145/3637528.3671931 on Pithdoi:10.1145/3637528.3671931
  10. [29] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam R. Kosiorek, Seungjin Choi, and Yee Whye Teh. 2019. Set Transformer: A Framework for Attention-based Permutation-Invariant Neural Networks. InICML. PMLR, Long Beach, California, USA, 3744–3753. https://proceedings.mlr.press/v97/lee19d.html on Pith
  11. [30] Jiacheng Li, Ming Wang, Jin Li, Jinmiao Fu, Xin Shen, Jingbo Shang, and Julian J. McAuley. 2023. Text Is All You Need: Learning Language Representations for Sequential Recommendation. InKDD. ACM, New York, NY, USA, 1258–1267. doi:10.1145/3580305.3599519 on Pithdoi:10.1145/3580305.3599519
  12. [31] Lei Li, Yongfeng Zhang, Dugang Liu, and Li Chen. 2024. Large Language Models for Generative Recommendation: A Survey and Visionary Discus- sions. InLREC/COLING. ELRA and ICCL, Torino, Italia, 10146–10159. https: //aclanthology.org/2024.lrec-main.886 on Pith
  13. [32] Yubo Li, Yidi Miao, Ramayya Krishnan, and Rema Padman. 2026. JEV-as-a-Judge: Accept When Confident, Escalate When Unsure. https://arxiv.org/abs/2609.26550 arXiv:2609.26550 on PithPith's reviewarXiv
  14. [33] Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. LLaRA: Large Language-Recommendation Assistant. In SIGIR. ACM, New York, NY, USA, 1785–1795. doi:10.1145/3626772.3657690 on Pithdoi:10.1145/3626772.3657690
  15. [34] Xinyu Lin, Wenjie Wang, Yongqi Li, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua on Pith
  16. [35] Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2024. Fine- Tuning LLaMA for Multi-Stage Text Retrieval. InSIGIR. Association for Comput- ing Machinery, New York, NY, USA, 2421–2425. doi:10.1145/3626772.3657951 on Pithdoi:10.1145/3626772.3657951
  17. [36] Xupeng Miao, Gabriele Oliaro, Zhihao Zhang, Xinhao Cheng, Zeyu Wang, Zhengxin Zhang, Rae Ying Yee Wong, Alan Zhu, Lijie Yang, Xiaoxiang Shi, Chunan Shi, Zhuoming Chen, Daiyaan Arfeen, Reyna Abhyankar, and Zhihao Jia. 2024. SpecInfer: Accelerating Large Language Model Serving w... on Pithdoi:10.1145/3620666.3651335
  18. [37] Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M. Waleed Kadous, and Ion Stoica. 2025. RouteLLM: Learning to Route LLMs from Preference Data. InICLR. OpenReview.net, Sin- gapore, 34433–34448. https://proceedings.iclr.cc/paper_files/pape... on Pith
  19. [38] Qwen Team. 2025. Qwen3 Technical Report. https://arxiv.org/abs/2505.09388 arXiv:2505.09388 on PithPith's reviewarXiv
  20. [39] Swarnadeep Saha, Archiki Prasad, Justin Chih-Yao Chen, Peter Hase, Elias Stengel-Eskin, and Mohit Bansal. 2025. System 1.x: Learning to Balance Fast and Slow Planning with Language Models. InICLR. OpenReview.net, Singapore, 38652–38679. https://openreview.net/forum?id=zd0iX5xBhA on Pith
  21. [40] Jiaxi Tang and Ke Wang. 2018. Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding. InWSDM. Association for Computing Machinery, New York, NY, USA, 565–573. doi:10.1145/3159652.3159656 on Pithdoi:10.1145/3159652.3159656
  22. [41] Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D. Manning. 2023. Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback. InEMNLP. ... on Pithdoi:10.18653/v1/2023.emnlp-
  23. [42] Alicia Tsai, Adam Kraft, Long Jin, Chenwei Cai, Anahita Hosseini, Taibai Xu, Zemin Zhang, Lichan Hong, Ed H. Chi, and Xinyang Yi. 2024. Leveraging LLM Reasoning Enhances Personalized Recommender Systems. InACL (Findings). Association for Computational Linguistics, Stroudsburg,... on Pithdoi:10.18653/v1/2024.findings-acl.780
  24. [43] Qi Wang, Jindong Li, Shiqi Wang, Qianli Xing, Runliang Niu, He Kong, Rui Li, Guodong Long, Yi Chang, and Chengqi Zhang. 2024. Towards Next-Generation LLM-based Recommender Systems: A Survey and Beyond. https://arxiv.org/ abs/2410.19744 arXiv:2410.19744 on PithPith's reviewarXiv
  25. [44] Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & Cross Network for Ad Click Predictions. InADKDD. ACM, New York, NY, USA, 12:1–12:7. doi:10.1145/3124749.3124754 on Pithdoi:10.1145/3124749.3124754
  26. [45] Xinfeng Wang, Jin Cui, Yoshimi Suzuki, and Fumiyo Fukumoto. 2024. RDRec: Rationale Distillation for LLM-based Recommendation. InACL (Short Papers). Association for Computational Linguistics, Stroudsburg, PA, USA, 65–74. doi:10. 18653/V1/2024.ACL-SHORT.6 on Pith
  27. [46] Yan Wang, Zhixuan Chu, Xin Ouyang, Simeng Wang, Hongyan Hao, Yue Shen, Jinjie Gu, Siqiao Xue, James Y. Zhang, Qing Cui, Longfei Li, Jun Zhou, and Sheng Li. 2024. LLMRG: Improving Recommendations through Large Language Model Reasoning Graphs. InAAAI. AAAI Press, Vancouver, Cana... on Pithdoi:10.1609/aaai.v38i17.29887
  28. [47] Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen. 2024. A survey on large language models for recommendation.World Wide Web27, 5 (2024), 60. doi:10.1007/s11280-024-01291-2 on Pithdoi:10.1007/s11280-024-01291-2
  29. [48] Tiantong Wu and Wei Yang Bryan Lim. 2026. REFLEX with Jev for Efficient Selec- tive Control in LLM Agents. https://arxiv.org/abs/2609.26532 arXiv:2609.26532 on PithPith's reviewarXiv
  30. [49] Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. 2024. Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs. InICLR. OpenReview.net, Vienna, Austria, 23650–23678. https://openreview.net/forum?id=gjeQKFxFpZ on Pith
  31. [51] Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabás Póczos, Ruslan Salakhutdinov, and Alexander J. Smola. 2017. Deep Sets. InNIPS. Curran Asso- ciates, Inc., Long Beach, CA, USA, 3391–3401. https://proceedings.neurips.cc/ paper/2017/hash/f22e4747da1aa27e363d86d40ff442fe... on Pith
  32. [52] Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, and Xiangnan He on Pith
  33. [53] Guorui Zhou, Xiaoqiang Zhu, Chengru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click-Through Rate Prediction. InKDD. ACM, New York, NY, USA, 1059–1068. doi:10.1145/3219819.3219823 on Pithdoi:10.1145/3219819.3219823
  34. [54] Kunlun Zhu, Tao Feng, Xu Han, Heng Wang, Ling Yang, Xiangru Tang, Jiaxuan You, and Heng Ji. 2026. JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs. https://arxiv.org/abs/2610.21412 arXiv:2610.21412. A Evaluation Details Anonymo... on PithPith's reviewarXiv
  35. [2016] Session-based Recommendations with Recurrent Neural Networks. InICLR (Poster). ICLR, San Juan, Puerto Rico, 10 pages. https://arxiv.org/abs/1511.06939 on PithPith's reviewarXiv
  36. [2023] TALLRec: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation. InRecSys. ACM, New York, NY, USA, 1007–1014. doi:10.1145/3604915.3608857 on Pithdoi:10.1145/3604915.3608857
  37. [2024] Bridging Items and Language: A Transition Paradigm for Large Language Model-Based Recommendation. InKDD. ACM, New York, NY, USA, 1816–1826. doi:10.1145/3637528.3671884 on Pithdoi:10.1145/3637528.3671884
  38. [2025] CoLLM: Integrating Collaborative Embeddings into Large Language Models for Recommendation.IEEE Transactions on Knowledge and Data Engineering37, 5 (2025), 2329–2340. doi:10.1109/TKDE.2025.3540912 on Pithdoi:10.1109/tkde.2025.3540912
  39. [5235] https://proceedings.mlr.press/v235/cai24b.html on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment