TrackRec: Iterative Alternating Feedback with Chain-of-Thought via Preference Alignment for Recommendation

Chi Lu; Junchen Wan; Peng Jiang; Qingpeng Cai; Rui Zhong; Wei Yang; Yu Xia; Zeyu Song

REVIEW 3 major objections 4 minor 60 references

TrackRec: Iterative Alternating Feedback with Chain-of-Thought via Preference Alignment for Recommendation

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Iterative generator–validator feedback lets a 7B LLM write preference summaries that improve recommenders.

desk verdict TrackRec is a plausible industrial co-training loop for LLM-generated preference summaries, but the self-referential reward and a sloppy table should give referees pause. read the letter →

arxiv 2508.15388 v1 pith:J27FOXAV submitted 2025-08-21 cs.IR

Yu Xia , Rui Zhong , Zeyu Song , Wei Yang , Junchen Wan , Qingpeng Cai , Chi Lu , Peng Jiang This is my paper

classification cs.IR

keywords recommendationsystemschain-of-thoughtreasoninglargelanguagemodelspreferencealignmentdirectoptimizationusersummarizationCTRpredictionadvertisingplatform

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

TrackRec claims that a small language model can learn to write useful chain-of-thought summaries of a user's preferences—without any hand-written reference summaries—by alternating two roles. One model, G, writes candidate 'RecCoT' texts; the other, V, scores each text by how confidently it predicts the user's actual click or rating. The best-scoring summaries are used to reinforce G, and G's improved summaries are then used to retrain V, round after round. The paper reports that this loop makes a 7B model outperform larger and fine-tuned LLM recommenders on MovieLens and Amazon-Book, and that the generated summaries, embedded as extra features, lift prediction accuracy across several standard CTR models. Deployed on an advertising platform, the same approach raised revenue by 2.3% and conversion rate by 1.6% in a 14-day test.

What carries the argument

RecCoT (recommendation chain-of-thought) is the central object: a generated text summary of a user's inferred preferences, used as explanation and as an auxiliary feature. The load-bearing mechanism is the alternating G/V feedback loop: G proposes N sampled RecCoTs, V scores each by the normalized probability of the user's actual label (Equation 1), the winner is the positive sample and the rest are negatives for Softmax-DPO (Equation 4), then V is retrained on G's best output ('Rec-tuning', Equation 6), and the pair is iterated for T rounds. This loop converts V's prediction confidence into a training signal for G, and G's improving summaries into training data for V.

What would settle it

On a held-out user set, sample N RecCoTs per user and score them with both V's reward and an external oracle (e.g., the downstream CTR model's loss or a validator trained on a disjoint period). If the RecCoT V ranks first does not beat a random RecCoT under the external oracle, or if swapping in an independently trained validator removes the reported AUC gains, the loop is optimizing V's biases rather than user preference.

Watch

Extended reading notes

Core claim

TrackRec's central claim is that recommendation chain-of-thought (RecCoT)—a short natural-language summary of what a user's history implies about their tastes—can be made reliable enough to improve real recommenders without collecting gold summaries. The method runs an alternating feedback loop: G samples several RecCoTs per user; V, asked whether the user will like a target item, gives each RecCoT a reward equal to its normalized probability of the user's actual behavior; the top-scoring RecCoT becomes the positive example and the rest negative examples for Softmax-DPO training of G. After G is updated, V is fine-tuned on G's greedily decoded RecCoT, and the two are retrained in alternating

Load-bearing premise

The load-bearing premise is that V's confidence in the label it predicts—computed on training data—is a reliable reward for the quality of the generated summary; if V is miscalibrated or shares G's blind spots, the alternating loop optimizes V's biases rather than true user preference.

Editorial extensions

If this is right

RecCoT from TrackRec acts as a model-agnostic auxiliary feature: the paper shows AUC and ACC gains when it is added to AutoInt, DeepFM, FiGNN, FiBiNet, DCN, DIN, DIEN, and Qwen-7B on both public datasets.
Iterations matter: performance rises from iteration 0 to iteration 3 in all four tracked configurations (Qwen-7B and DIN on both datasets), with AUC gains up to 1.56 percentage points.
A small model can be initialized cheaply: fine-tuning on roughly 5% of teacher-generated RecCoT data from GPT-4o is enough to lift the 7B generator's downstream recommendation performance.
The approach is deployable at scale: a 14-day A/B test on an advertising platform with hundreds of millions of users showed revenue +2.3% and CVR +1.6% overall, and +4.2%/+1.7% on long-tail traffic.

Reading between the lines

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

The authors frame TrackRec as improving reasoning, but the reward is entirely internal: V's confidence in the observed label. This makes the method a form of self-training with a learned reward model, so its ceiling is set by V's calibration; a V that shares G's blind spots would make the loop confidently reinforce the wrong summaries.
The same alternating generator–validator recipe could transfer to any task with abundant implicit labels and a cheap proxy scorer—for example, search snippets scored by click-through, or candidate explanations scored by downstream model accuracy—without needing human-written gold chains of thought.
Because the RecCoT is injected as BERT+MLP feature fields, a controlled test that removes the LLM text but keeps the extra feature path would separate gains from the reasoning content vs. gains from added model capacity; the paper does not run that isolation.
One testable extension: vary the number of samples N and iterations T. If V's reward is a faithful proxy, performance should plateau then degrade with too many iterations (overfitting to V); observing that pattern would confirm the loop is optimizing the proxy rather than true preference.

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.

Desk Editor's Note

read the letter

Short version: this is a plausible industrial co-training recipe for small-LLM-generated preference summaries, and the online A/B result is the most interesting thing in it. But the offline gains are tiny, the reward loop is self-referential without a fully convincing external anchor, and Table 7 has a blatant internal inconsistency that needs fixing before I'd trust any of the numbers.

What's genuinely new: the alternating G-V loop — Generator writes a user-preference CoT, Validator scores it, and they take turns upgrading each other — combined with S-DPO for preference alignment and a distillation warm-start from a larger LLM. That combination isn't in the cited RLPF or KAR work. The paper is also honest about the engineering constraints: small model, latency in production, long-tail gains. The ablations are reasonably thorough, running across seven backbones plus direct LLM recommendation, and the gains, while small, are consistent in sign.

The soft spots, in order. First, the reward for G is V's confidence in the observed label (Eq. 1), and V is fine-tuned on G's own outputs (Eq. 5). That is a co-adaptation loop with no human judgment or external coherence check on the CoTs. The fact that the CoTs also improve a standard backbone like DIN is some evidence they carry real signal, but it's indirect: the backbone is trained end-to-end with the CoT features, so any superficial correlation that helps the training loss could be exploited. A fixed-validator or data-free sanity check would help. Second, the offline improvements are tiny — AUC deltas of 0.001 to 0.009 — and there are no error bars, no seeds, no variance anywhere. The claim of statistical significance with a t-test at p<0.001 without reporting distributions is not convincing. Third, Table 7 reports MovieLens Qwen-7B ACC jumping from 0.7011 to 0.8047 with distillation, while the text says the improvement is 0.56%. That's a 10-point jump, not 0.56%; either the table cell or the text is wrong, and that kind of error makes all the other numbers suspect. Also minor: S-DPO is used without citation to the original S-DPO work.

The central idea is coherent and the online result, if real, is a meaningful business signal. I'm skeptical of the offline numbers as reported, but I wouldn't desk-reject it. It deserves a serious referee who should ask for error bars, a corrected Table 7, and either an external grounding of CoT quality or at least a fixed-validator ablation.

Referee Report

3 major / 4 minor

Summary. TrackRec is an LLM-based recommendation framework with a RecCoT generator G and a validator V. G samples multiple chain-of-thought summaries of user preference; V scores each summary by the probability it assigns to the observed user behavior label; softmax-DPO trains G toward high-scoring summaries; V is then fine-tuned on G's greedy outputs. Distillation from GPT-4o initializes G, and the RecCoT text is embedded via BERT and injected as auxiliary features into CTR/CVR backbones. Experiments on Amazon-Book, MovieLens, and an industrial advertising platform claim consistent AUC/ACC gains over strong baselines, plus online revenue +2.3% and CVR +1.6%.

Significance. If the claims hold, TrackRec is practically valuable: it is model-agnostic, uses only small LLMs at inference, and is deployed at scale. The paper's strengths are a clear two-component architecture, broad backbone coverage, and ablations showing monotone improvement over alternating-feedback iterations and distillation. The central weakness is that the reward signal for G is V's confidence on the observed label, with no external grounding: V is itself trained on G's outputs, so the loop can co-adapt without improving genuine preference analysis. The numerical evidence also contains internal inconsistencies, most notably Table 7's ACC jump. These issues must be resolved before the central claim is established.

major comments (3)

[Sec. 3.2.2, 3.3; Eq. (1), Algorithm 2] The reward for training G is V's probability of the observed user label for each sampled RecCoT (Eq. 1). V is then fine-tuned on G's own greedy decoded CoTs (Eq. 5). This closed loop has no external anchor: no human evaluation of RecCoT quality, no calibration analysis of V's probabilities, and no frozen-validator control. G can therefore exploit V's idiosyncratic patterns while V learns to read those patterns, without the CoT capturing genuine user preference. The claim that TrackRec 'continuously improves user preference analysis' is load-bearing and needs external validation, e.g., human annotations, a calibration plot, or a fixed-validator baseline.
[Table 7, Sec. 4.4.2] Table 7 reports MovieLens Qwen-7B ACC jumping from 0.7011 (w/o distillation) to 0.8047 (w/ distillation), an absolute +10.36 percentage point change, while Sec. 4.4.2 describes the improvement as 0.56%. The corresponding AUC change is only +0.0045, so the ACC value is implausible and internally inconsistent with both the text and the AUC. This appears to be a reporting error, but it undermines confidence in the distillation ablation results and must be corrected and re-verified.
[Tables 2 and 4; Sec. 4.1.5] All offline AUC gains are small (roughly +0.001 to +0.009), yet the paper reports no error bars, number of seeds, or confidence intervals. A t-test with p<0.001 is not interpretable without variance information. Similarly, Table 4 reports online revenue +2.3% and CVR +1.6%, but gives no A/B test duration, bucket size, significance level, or whether the metrics are adjusted for confounders. Please report multiple seeds with standard deviations and, for the online test, a standard significance statement.

minor comments (4)

[Sec. 3.4, Eq. (6)] Eq. (6) is called a loss, but as written it is a positive log-likelihood (no minus sign). Please clarify whether this is an objective to maximize or a loss to minimize.
[Sec. 3.2.1, Eq. (3) vs Eq. (4)] The S-DPO objective in Eq. (3) does not obviously match the loss in Eq. (4); the N-1 exponent and the role of β should be explained step by step.
[Sec. 4.1.4] Please report exact hyperparameters for reproducibility: distillation data fraction ('about 5%'), S-DPO β, LoRA rank (if used), grid-search ranges, and the number of random seeds.
[Throughout] Typos and unclear labels: 'lagre' in the Abstract, 'Amonzen Book' in Sec. 4.4.1, 'Alterating' in Algorithm 2, and 'ite ratio n' in the Figure 5 axis. Figure 1's axes and 'w/o w' labels should be self-explanatory.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the alternating G/V loop is an optimization heuristic, and the headline claims rest on external benchmarks.

full rationale

TrackRec's derivation chain is an empirical training loop rather than a logical derivation that reduces to its own inputs. G is trained to maximize V's confidence in the observed label (Eq. 1, Eq. 3), and V is trained to predict labels from G's outputs (Eq. 5, Eq. 6). This is a co-adaptation procedure, and the concern that the reward is self-referential is a validity/correctness threat: V may be miscalibrated, and G may exploit V's biases rather than learn true user preference. However, this is not circularity in the sense required here — no prediction is equivalent to a fitted input by construction. The paper's central performance claims (surpassing SOTA baselines in Table 2 and Table 3, and online revenue +2.3%, CVR +1.6% in Table 4) are evaluated on held-out test sets and online A/B tests, which are external to the training objective. No load-bearing self-citation was found; the same-author references [12, 43, 51] are not used to justify the core mechanism. The only notable anomaly is a numeric inconsistency in Table 7 (MovieLens Qwen-7B ACC jumps from 0.7011 to 0.8047 while the text reports a 0.56% improvement), which is a reporting error, not a circular step. Accordingly, the paper is self-contained with respect to its external benchmarks, and the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The paper's central mechanism rests on the assumption that a learned validator's confidence is a reliable reward for generated text quality, plus the standard assumption that user behavior labels capture true preference. The free parameters are mostly hyperparameters that are tuned by hand or grid search; the number of iterations is chosen based on the observed curve.

free parameters (6)

number of CoT samples N = 10
Set in Section 4.1.4; the preference alignment depends on this choice and no sensitivity analysis is provided.
sampling temperature t = 1.0
Used for multiple sampling in Algorithm 1; chosen by hand.
top_p = 0.9
Nucleus sampling parameter for the generator.
iterations T of alternating feedback = 3
Figure 5 shows gains up to iteration 3, so this is effectively chosen to maximize reported results.
learning rate = 1e-5
Used for optimizing both G and V; no ablation.
distillation data fraction = about 5%
Authors state 5% of distilled data is sufficient; no ablation on this fraction.

assumptions (5)

domain assumption User behavior labels (clicks/ratings) are the correct ground truth for preference quality.
The entire reward and validation signal in Sections 3.2 and 3.3 is derived from y_i; if labels are noisy or proxy, the loop optimizes the proxy.
ad hoc to paper Validator V's output probability is a valid reward signal for RecCoT quality.
Eq. (1) uses P(V=Yes/No) as reward r_ij; no evidence that V's confidence correlates with true CoT quality. This is the paper's core mechanism and is not benchmarked externally.
domain assumption BERT-encoded natural-language RecCoT summaries provide information complementary to ID features in the backbone model.
Section 3.6 assumes this; ablations give indirect support, but no analysis of information overlap is provided.
ad hoc to paper Iterative alternating training of G and V converges to better recommendation performance.
No convergence analysis or theoretical guarantee; Figure 5 shows empirical gains for T=3 only.
standard math Softmax-DPO objective is a valid preference optimization loss.
Eq. (4) uses S-DPO; it is a known method, though the paper does not cite it.

invented entities (1)

RecCoT
purpose: Natural-language summary of inferred user preference, used as an explanation and as auxiliary features for downstream models.
RecCoT's quality is only measured through downstream AUC in this paper; no human evaluation or external benchmark validates the content of the summaries. It is an invented intermediate representation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TrackRec: Iterative Alternating Feedback with Chain-of-Thought via Preference Alignment for Recommendation." pith.science (2026). https://pith.science/paper/J27FOXAV

@misc{pith2026250815388,
  author       = {Pith},
  title        = {Pith review of: TrackRec: Iterative Alternating Feedback with Chain-of-Thought via Preference Alignment for Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J27FOXAV}},
  note         = {Machine review of arXiv:2508.15388}
}

abstract

The extensive world knowledge and powerful reasoning capabilities of large language models (LLMs) have attracted significant attention in recommendation systems (RS). Specifically, The chain of thought (CoT) has been shown to improve the performance of LLMs on complex reasoning tasks for RS. However, due to the fact that LLMs often suffer from hallucination issues, there is no guarantee that their reasoning CoT is effective. A key challenge is to further enhance the recommendation capabilities of LLMs through effective CoT reasonings. Therefore, we propose \textbf{TrackRec}, a framework designed to enhance reasoning capabilities of LLMs for RS. TrackRec specifically focuses on accurately inferring recommendation CoT \textbf{(RecCoT)} for user preference using the knowledge from LLMs. This RecCoT can serve both as an explanation for the LLM's completion of recommendation tasks and as auxiliary features to assist recommendation models in accomplishing recommendation tasks. TrackRec consists of a RecCoT generator $(G)$ and a RecCoT validator $(V)$. Furthermore, we design alternating feedback learning mechanism that $G$ undergoes direct preference optimization via feedback from $V$ to produce increasingly accurate RecCoT aligned with $V$'s standards. Meanwhile, $V$ is fine-tuned using the inference feedback from $G$ to enhance its validation capabilities in alignment with recommendation tasks. Through iterative alternating feedback learning between $G$ and $V$, TrackRec continuously improves the user preference analysis capability of $G$ and the validation capacity of $V$. Extensive experiments demonstrate the effectiveness of our approach, showing that it surpasses state-of-the-art methods. Moreover, TrackRec has been deployed on a lagre advertising platform with hundreds of millions of users, achieving substantial gains.

Figures

Figures reproduced from arXiv: 2508.15388 by the authors.

**Figure 2.** The issues related to introducing chain-of-thought [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The architecture of TrackRec, consists of a recommendation CoT (RecCoT) generator [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Example prompts for TrackRec, including recommendation CoT reasoning prompt denoted as [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Performance of the iterative alternating learning mechanism on different datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 15 canonical work pages

[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.arXiv preprint arXiv:2303.08774 (2023)

arXiv 2023
[2]

Sanjeev Arora. 2024. From Word-prediction to Complex Skills: Compositional Thinking and Metacognition in LLMs. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1–1

work page 2024
[3]

Amos Azaria, Avinatan Hassidim, Sarit Kraus, Adi Eshkol, Ofer Weintraub, and Irit Netanely. 2013. Movie recommender system for profit maximization. In Proceedings of the 7th ACM conference on Recommender systems. 121–128

work page 2013
[4]

Zhuoxi Bai, Ning Wu, Fengyu Cai, Xinyi Zhu, and Yun Xiong. 2024. Finetuning Large Language Model for Personalized Ranking.arXiv preprint arXiv:2405.16127 (2024)

arXiv 2024
[5]

Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InProceedings of the 17th ACM Conference on Recommender Systems. 1007–1014

2023
[6]

Jin Chen, Zheng Liu, Xu Huang, Chenwang Wu, Qi Liu, Gangwei Jiang, Yuanhao Pu, Yuxuan Lei, Xiaolong Chen, Xingmei Wang, et al. 2023. When large language models meet personalization: Perspectives of challenges and opportunities.arXiv preprint arXiv:2307.16376(2023)

arXiv 2023
[7]

Zeyu Cui, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022. M6-rec: Generative pretrained language models are open-ended recommender systems.arXiv preprint arXiv:2205.08084(2022)

arXiv 2022
[8]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805(2018)

arXiv 2018

Show all 60 references

[9]

Wenqi Fan, Zihuai Zhao, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Jiliang Tang, and Qing Li. 2023. Recommender systems in the era of large language models (llms).arXiv preprint arXiv:2307.02046(2023)

2023 arXiv
[10]

Tianyu Gao, Adam Fisch, and Danqi Chen. 2020. Making pre-trained language models better few-shot learners.arXiv preprint arXiv:2012.15723(2020)

2020 arXiv
[11]

Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. 2023. Chat-rec: Towards interactive and explainable llms-augmented recommender system.arXiv preprint arXiv:2303.14524(2023)

2023 arXiv
[12]

Hao Gu, Rui Zhong, Yu Xia, Wei Yang, Chi Lu, Peng Jiang, and Kun Gai. 2025. R4ec: A Reasoning, Reflection, and Refinement Framework for Recommendation Systems.arXiv preprint arXiv:2507.17249(2025)

2025 arXiv
[13]

Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. CoRRabs/1703.04247 (2017). arXiv:1703.04247 http://arxiv.org/abs/1703.04247

2017 arXiv
[14]

Joseph F Hair Jr, Barry J Babin, and Nina Krey. 2017. Covariance-based structural equation modeling in the Journal of Advertising: Review and recommendations. Journal of Advertising46, 1 (2017), 163–177

2017
[15]

Ruining He and Julian McAuley. 2016. VBPR: visual bayesian personalized ranking from implicit feedback. InProceedings of the AAAI conference on artificial intelligence, Vol. 30

2016
[16]

Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. 2024. Large language models are zero-shot rankers for recommender systems. InEuropean Conference on Information Retrieval. Springer, 364–381

2024
[17]

Tongwen Huang, Zhiqi Zhang, and Junlin Zhang. 2019. FiBiNET: combining fea- ture importance and bilinear feature interaction for click-through rate prediction. InProceedings of the 13th ACM conference on recommender systems. 169–177

2019
[18]

Xuanwen Huang, Kaiqiao Han, Yang Yang, Dezheng Bao, Quanjin Tao, Ziwei Chai, and Qi Zhu. 2024. Can GNN be Good Adapter for LLMs?. InProceedings of the ACM on Web Conference 2024. 893–904

2024
[19]

Xu Huang, Jianxun Lian, Yuxuan Lei, Jing Yao, Defu Lian, and Xing Xie. 2023. Recommender ai agent: Integrating large language models for interactive recom- mendations.arXiv preprint arXiv:2308.16505(2023)

2023 arXiv
[20]

Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners.Advances in neural information processing systems35 (2022), 22199–22213

2022
[21]

Anis Koubaa. 2023. GPT-4 vs. GPT-3.5: A concise showdown. (2023)

2023
[22]

Chunyuan Li, Haotian Liu, Liunian Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, et al . 2022. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. Advances in Neural Information Processing Systems...

2022
[23]

Zekun Li, Zeyu Cui, Shu Wu, Xiaoyu Zhang, and Liang Wang. 2019. Fi-gnn: Modeling feature interactions via graph neural networks for ctr prediction. In Proceedings of the 28th ACM international conference on information and knowledge management. 539–548

2019
[24]

Wensheng Lu, Jianxun Lian, Wei Zhang, Guanghua Li, Mingyang Zhou, Hao Liao, and Xing Xie. 2024. Aligning Large Language Models for Controllable Recommendations.arXiv preprint arXiv:2403.05063(2024)

2024 arXiv
[25]

Hanjia Lyu, Song Jiang, Hanqing Zeng, Yinglong Xia, and Jiebo Luo. 2023. Llm- rec: Personalized recommendation via prompting large language models.arXiv preprint arXiv:2307.15780(2023)

2023 arXiv
[26]

Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback.Advances in neural information processing systems35 (...

2022
[27]

Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. 2019. Autoint: Automatic feature interaction learning via self- attentive neural networks. InProceedings of the 28th ACM international conference on information and knowledge management....

2019
[28]

Zayne Sprague, Fangcong Yin, Juan Diego Rodriguez, Dongwei Jiang, Manya Wadhwa, Prasann Singhal, Xinyu Zhao, Xi Ye, Kyle Mahowald, and Greg Durrett
[29]

Qi Sun, Kun Huang, Xiaocui Yang, Rong Tong, Kun Zhang, and Soujanya Poria
[30]

Weiwei Sun, Lingyong Yan, Xinyu Ma, Pengjie Ren, Dawei Yin, and Zhaochun Ren. 2023. Is chatgpt good at search? investigating large language models as re-ranking agent.arXiv preprint arXiv:2304.09542(2023)

2023 arXiv
[31]

InProceedings of the ACM on Web Conference 2024

Consistency guided knowledge retrieval and denoising in llms for zero- shot document-level relation triplet extraction. InProceedings of the ACM on Web Conference 2024. 4407–4416

2024
[32]

Lei Wang, Wanyu Xu, Yihuai Lan, Zhiqiang Hu, Yunshi Lan, Roy Ka-Wei Lee, and Ee-Peng Lim. 2023. Plan-and-solve prompting: Improving zero-shot chain- of-thought reasoning by large language models.arXiv preprint arXiv:2305.04091 (2023)

2023 arXiv
[33]

Aaron Van den Oord, Sander Dieleman, and Benjamin Schrauwen. 2013. Deep content-based music recommendation.Advances in neural information processing systems26 (2013). TrackRec: Iterative Alternating Feedback with Chain-of-Thought via Preference Alignment for Recommendation Con...

2013
[34]

Yancheng Wang, Ziyan Jiang, Zheng Chen, Fan Yang, Yingxue Zhou, Eunah Cho, Xing Fan, Xiaojiang Huang, Yanbin Lu, and Yingzhen Yang. 2023. Rec- mind: Large language model powered agent for recommendation.arXiv preprint arXiv:2308.14296(2023)

2023 arXiv
[35]

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

2017
[36]

Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al
[37]

Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners.arXiv preprint arXiv:2109.01652(2021)

2021 arXiv
[38]

Jiaxing Wu, Lin Ning, Luyang Liu, Harrison Lee, Neo Wu, Chao Wang, Sushant Prakash, Shawn O’Banion, Bradley Green, and Jun Xie. 2024. RLPF: Reinforcement Learning from Prediction Feedback for User Summarization with LLMs.arXiv preprint arXiv:2409.04421(2024)

2024 arXiv
[39]

Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al . 2023. A survey on large language models for recommendation.arXiv preprint arXiv:2305.19860(2023)

2023 arXiv
[40]

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems35 (2022), 24824–24837

2022
[41]

Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. Towards open-world recommendation with knowledge augmentation from large language models. In Proceedings of the 18th ACM Conference on Recommender ...

2024
[42]

Yunjia Xi, Weiwen Liu, Jianghao Lin, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, Rui Zhang, and Yong Yu. 2023. Towards open-world recommen- dation with knowledge augmentation from large language models.arXiv preprint arXiv:2306.10933(2023)

2023 arXiv
[43]

Songhao Wu, Quan Tu, Hong Liu, Jia Xu, Zhongyi Liu, Guannan Zhang, Ran Wang, Xiuying Chen, and Rui Yan. 2024. Unify Graph Learning with Text: Unleashing LLM Potentials for Session Search. InProceedings of the ACM on Web Conference 2024. 1509–1518

2024
[44]

An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Cheng- peng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzhe...

2024 arXiv
[45]

Fan Yang, Zheng Chen, Ziyan Jiang, Eunah Cho, Xiaojiang Huang, and Yanbin Lu. 2023. Palr: Personalization aware llms for recommendation.arXiv preprint arXiv:2305.07622(2023)

2023 arXiv
[46]

Yu Xia, Rui Zhong, Hao Gu, Wei Yang, Chi Lu, Peng Jiang, and Kun Gai. 2025. Hierarchical Tree Search-based User Lifelong Behavior Modeling on Large Lan- guage Model. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retriev...

2025
[47]

Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2023. Recommendation as instruction following: A large language model empowered recommendation approach.arXiv preprint arXiv:2305.07001(2023)

2023 arXiv
[48]

Yang Zhang, Keqin Bao, Ming Yan, Wenjie Wang, Fuli Feng, and Xiangnan He
[49]

Jizhi Zhang, Keqin Bao, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Is chatgpt fair for recommendation? evaluating fairness in large lan- guage model recommendation. InProceedings of the 17th ACM Conference on Recommender Systems. 993–999

2023
[50]

Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, and Xiangnan He
[51]

Rui Zhao, Rui Zhong, Haoran Zheng, Wei Yang, Chi Lu, Beihong Jin, Peng Jiang, and Kun Gai. 2025. Hierarchical Sequence ID Representation of Large Language Models for Large-scale Recommendation Systems. InCompanion Proceedings of the ACM on Web Conference 2025. 641–650

2025
[52]

arXiv:2406.03210 [cs.IR] https://arxiv.org/abs/2406.03210

Text-like Encoding of Collaborative Information in Large Language Models for Recommendation. arXiv:2406.03210 [cs.IR] https://arxiv.org/abs/2406.03210

arXiv
[53]

Yuhui Zhang, Hao Ding, Zeren Shui, Yifei Ma, James Zou, Anoop Deoras, and Hao Wang. 2021. Language models as recommender systems: Evaluations and limitations. InI (Still) Can’t Believe It’s Not Better! NeurIPS 2021 Workshop

2021
[54]

Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep Interest Evolution Network for Click-Through Rate Prediction.Proceedings of the AAAI Conference on Artificial Intelligence33, 01 (Jul. 2019), 5941–5948. doi:10.1609/aaai.v33i0...

2019 doi
[55]

Guorui Zhou, Xiaoqiang Zhu, Chenru 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. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining(...

2018
[57]

Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models.arXiv preprint arXiv:2303.18223(2023)

2023 arXiv
[58]

Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. 2022. Least-to- most prompting enables complex reasoning in large language models.arXiv preprint arXiv:2205.10625(2022)

2022 arXiv
[2022]

Emergent abilities of large language models.arXiv preprint arXiv:2206.07682 (2022)

2022 arXiv
[2023]

Collm: Integrating collaborative embeddings into large language models for recommendation.arXiv preprint arXiv:2310.19488(2023)

2023 arXiv
[2024]

To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning.arXiv preprint arXiv:2409.12183(2024)

2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[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.arXiv preprint arXiv:2303.08774 (2023)

arXiv 2023

[2] [2]

Sanjeev Arora. 2024. From Word-prediction to Complex Skills: Compositional Thinking and Metacognition in LLMs. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1–1

work page 2024

[3] [3]

Amos Azaria, Avinatan Hassidim, Sarit Kraus, Adi Eshkol, Ofer Weintraub, and Irit Netanely. 2013. Movie recommender system for profit maximization. In Proceedings of the 7th ACM conference on Recommender systems. 121–128

work page 2013

[4] [4]

Zhuoxi Bai, Ning Wu, Fengyu Cai, Xinyi Zhu, and Yun Xiong. 2024. Finetuning Large Language Model for Personalized Ranking.arXiv preprint arXiv:2405.16127 (2024)

arXiv 2024

[5] [5]

Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InProceedings of the 17th ACM Conference on Recommender Systems. 1007–1014

2023

[6] [6]

Jin Chen, Zheng Liu, Xu Huang, Chenwang Wu, Qi Liu, Gangwei Jiang, Yuanhao Pu, Yuxuan Lei, Xiaolong Chen, Xingmei Wang, et al. 2023. When large language models meet personalization: Perspectives of challenges and opportunities.arXiv preprint arXiv:2307.16376(2023)

arXiv 2023

[7] [7]

Zeyu Cui, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022. M6-rec: Generative pretrained language models are open-ended recommender systems.arXiv preprint arXiv:2205.08084(2022)

arXiv 2022

[8] [8]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805(2018)

arXiv 2018