Pith. sign in

REVIEW 3 major objections 4 minor 45 references

STARec: An Efficient Agent Framework for Recommender Systems via Autonomous Deliberate Reasoning

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 The paper claims that recommender systems can beat full-data baselines using only 0.4% of training data when each user is an agent that ranks fast, reflects slowly, and updates a written preference profile.

desk verdict A coherent LLM agent pipeline with strong results, but the 0.4% data-efficiency claim is arithmetically wrong and the evaluation leaves open a target-label leakage in self-reflection. read the letter →

arxiv 2508.18812 v1 pith:EQUXKG4J submitted 2025-08-26 cs.AI

Chenghao Wu , Ruiyang Ren , Junjie Zhang , Ruirui Wang , Zhongrui Ma , Qi Ye , Wayne Xin Zhao This is my paper
classification cs.AI
keywords RecommendationAgentsDeliberateReasoningReinforcementLearningLLMchain-of-thoughtknowledgedistillationdual-processcognitiondataefficiency
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

STARec claims that recommendation accuracy improves dramatically when each user is modeled as an LLM agent that does two things: rank candidate items quickly, then deliberately reflect on mismatches between that ranking and real user feedback and rewrite its stored picture of the user. The training recipe—anchored reinforcement training—first distills structured reasoning from a stronger teacher model, then uses reinforcement learning with a ranking-based reward to sharpen both the ranking and the memory updates. The headline result is that a 7B-parameter agent trained on about 40,000 interactions reaches NDCG@10 of 77.16 on MovieLens 1M and 82.63 on Amazon CDs, exceeding strong full-data baselines. If correct, this means recommender systems can become both more accurate and far more data-efficient by giving agents a slow-thinking self-correction loop.

What carries the argument

The central mechanism is the dual-process agent cognition cycle combined with anchored reinforcement training. Each user is an agent with a natural-language memory storing demographic data, interaction history, and an evolving preference summary. Fast thinking generates an item ranking with chain-of-thought rationales; slow thinking identifies discrepancies between predicted and actual feedback and issues an updated preference summary. This cycle is trained in two anchored stages: SFT distillation from a stronger reasoning teacher instills foundational skills, and GRPO reinforcement learning with a rank-interval reward (positive for top ranks, negative beyond rank 20) simultaneously optimize

What would settle it

Run the released STARec model with the self-reflection loop disabled at test time (i.e., freeze the memory after training) and measure NDCG@10 on the same leave-one-out sets; if the gap to the full STARec number largely disappears, the reported gains come from test-time feedback. A complementary check: after the memory update, remove the ground-truth item from the candidate set and verify that the ranking of the remaining 19 candidates is unchanged.

Watch

Extended reading notes

Core claim

The paper's contribution is an agent framework, STARec, whose agents combine fast and slow thinking. A fast ranking pass produces an ordered list with chain-of-thought justifications from the agent's natural-language memory. A slow self-reflection pass then compares the agent's predicted feedback with the user's actual feedback and rewrites the preference profile before subsequent rankings. The training procedure, anchored reinforcement training, works in two stages: supervised distillation from a stronger reasoning teacher builds foundational skills (preference summarization, rationale generation, ranking logic), and GRPO-style reinforcement learning with a rank-based reward further adapts

Load-bearing premise

The paper does not state whether the self-reflection and memory-update steps run during evaluation and, if so, whether the held-out ground-truth item is included in the feedback that updates the profile before the final ranking; if it is, the reported gains partly reflect access to the answer.

Editorial extensions

If this is right

  • A 0.5B-parameter agent retains roughly 97% of the 7B model's SFT-stage performance and about 88% after RL, so lightweight agents can be deployed where latency and memory are constrained.
  • The self-reflection loop lets agents improve recommendation quality for low-activity users, directly addressing cold-start scenarios.
  • RL in this framework mainly amplifies solutions the SFT model can already sample, so a strong SFT initialization is a prerequisite for effective RL enhancement.
  • Generated chain-of-thought rationales provide interpretable explanations alongside ranked outputs, offering transparency for recommendation decisions.
  • Because only about 40,000 interactions are needed for training, the framework reduces training cost far below full-dataset supervised approaches while matching or exceeding their accuracy.

Reading between the lines

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

  • The paper does not state explicitly whether the self-reflection and memory-update loop runs at evaluation time; if it does and the held-out ground-truth item is included as 'actual feedback' before the final ranking, the reported NDCG numbers would partly reflect test-time leakage rather than learned generalization.
  • The same anchored-reinforcement recipe—teacher distillation followed by GRPO with a rule-based reward—could transfer to other structured-output tasks such as summarization-reranking, tool selection, or query rewriting, where a strong reasoning teacher and a metric-based reward exist.
  • The 'Best of N' analysis implies that an SFT-only model can approach the RL model's single-sample accuracy by sampling many times at inference, which offers a compute-expensive but RL-free alternative for deployment.
  • The reward intervals (only six buckets over rank position) are coarse approximations of NDCG; finer-grained or metric-continuous rewards might close the remaining gap to exact ranking objectives.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. STARec is an LLM-agent framework for sequential recommendation. Each user is represented by an agent with a natural-language memory: 'fast thinking' ranks candidate items, and 'slow thinking' reflects on discrepancies between predicted and observed feedback and updates the memory. Training uses a two-stage 'anchored reinforcement training' pipeline: SFT on CoT data distilled from DeepSeek-R1-Distill-Qwen-32B, followed by GRPO with a hand-designed, NDCG-inspired reward. Evaluation is on MovieLens-1M and Amazon CDs with a leave-one-out protocol (last interaction as ground truth, 19 random distractors), reporting NDCG@1/5/10/20. The paper's central claim is that STARec beats state-of-the-art baselines while using only 0.4% of the full training data.

Significance. If the results hold, the framework is a useful contribution: it demonstrates that a 7B base model can outperform much larger teacher-based prompting and full-data sequential baselines through distilled CoT plus RL, and it provides interpretable rationales. The work has genuine strengths: the ablations in §5.3.1 isolate SFT and self-reflection, the scaling analysis in §5.3.2 is informative, and the 'Best of N' analysis in §5.3.4 is a thoughtful way to separate knowledge acquisition from RL's 'success amplification.' However, significance is currently limited by an underspecified test-time evaluation boundary that could permit label leakage, and by a data-efficiency claim that is inconsistent with the paper's own dataset statistics. These issues must be resolved before the headline claims can be accepted.

major comments (3)
  1. [§5.1.2, §4.1.3, Table 2] The test-time evaluation protocol is underspecified in a way that affects the central claim. §5.1.2 describes leave-one-out evaluation but does not state whether the self-reflection/memory-update loop of §4.1.3 runs during testing and whether it ever receives the held-out item's true label before the final ranking. Table 2's reflection prompt explicitly includes 'User's Actual Feedback,' and §4.1.3 says the agent's interaction record is updated with 'the current item and its associated feedback.' If the held-out last interaction's label is supplied to the agent before it produces the scored ranking, then the ground-truth answer enters the context and the Table 4 NDCG numbers are inflated. The Table 5 ablation shows this stage matters substantially: removing self-reflection drops NDCG@10 from 75.19 to 71.31 on ML-1M and from 80.66 to 76.88 on CDs. The authors must specify exactly what hap
  2. [Abstract, §1 bullet 3, §5.2(3), Table 3] The claim that STARec uses 'only 0.4% of the full training data' is contradicted by Table 3. The sampled training set is 40,000 interactions for both datasets, while the full interaction counts are 1,000,209 for ML-1M and 4,543,369 for CDs. These ratios are 40,000/1,000,209 = 4.0% and 40,000/4,543,369 = 0.88%, not 0.4%. The abstract, the contribution list, and §5.2 should be corrected, and the efficiency claim should be restated with the actual fractions. The relative performance comparison may survive this correction, but the stated efficiency factor is wrong by a factor of 2 to 10.
  3. [Table 4, §5.1.2] No variance or significance information is reported. The paper says each test instance is executed three times and the average reported, but no standard deviations, confidence intervals, or statistical tests are given. On ML-1M, STARecsample NDCG@10 (77.16) exceeds SASRecfull (76.51) by only 0.65 points; with 19 random negatives and a 1,000-user test set, this margin could easily be within sampling noise. The CDs margin is larger (82.63 vs. 79.47), but still needs error bars. Please report mean±std (or per-user bootstrap intervals) and significance tests, and state whether identical random distractor sets were used across all methods.
minor comments (4)
  1. [§5.1.1 vs. §5.3.3] Table 3 reports exactly 40,000 interactions for 1,000 sampled users, i.e., 40 interactions per user after truncation. This seems inconsistent with the low/medium activity user groups (10–24 and 25–39 interactions) analyzed in §5.3.3 and Figure 2a. Clarify whether the activity groups are defined on the truncated sample and how user-level variation arises.
  2. [§4.2.2] The reward bins (positions 1, 2–5, 6–10, 11–20, >20 with rewards +1, +0.5, 0, −0.5, −1) are described as 'inspired by NDCG' but no sensitivity analysis is provided. Since GRPO normalizes advantages, the absolute scale is less critical, but the bin boundaries are still free parameters. A short robustness check with different boundaries would strengthen the paper.
  3. [Table 2 and §4.1.3] The terms 'System's Prediction' and 'User's Actual Feedback' in the reflection prompt are not formally tied to the training/evaluation protocol. The authors should state explicitly whether these are observed labels, simulator outputs, or held-out labels, since this directly affects reproducibility.
  4. [Conclusion] Typo: 'estabilish' should be 'establish.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: STARec's empirical gains are not entailed by its training inputs; self-citations are contextual and the reward/evaluation alignment is direct objective alignment, not circular reasoning.

full rationale

The paper's central claim is that STARec's RL-enhanced agents outperform full-data baselines on held-out leave-one-out NDCG ranking (abstract, Table 4, §5.2). This result is not equivalent to its own inputs: the SFT data are teacher-generated rankings filtered by NDCG on the training split, the GRPO reward is a hand-set NDCG-inspired binning, and the evaluation uses a separate test split with random distractors. The RL objective and the evaluation metric both target NDCG, which is objective alignment rather than a derivation from the metric alone. The few self-citations (AgentCF, LLMRank, surveys, RecBole) are used as baselines or context and do not carry the argument. Two non-circular concerns are noted: the '0.4%' data-efficiency figure is inconsistent with Table 3's own 40,000/1,000,209 (≈4.0%) and 40,000/4,543,369 (≈0.88%) ratios, which is a reporting error rather than circularity; and §5.1.2 does not explicitly state whether the §4.1.3 memory-update/self-reflection loop runs at test time with access to the held-out item's actual feedback, which would be evaluation leakage. Neither concern is a circular step established by the manuscript, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central empirical claim rests on the assumptions listed above; the framework is an architectural and training pipeline, so it introduces no new physical or mathematical entities.

free parameters (2)
  • Ranking reward bins for GRPO = rank 1: +1.0; ranks 2-5: +0.5; ranks 6-10: 0.0; ranks 11-20: -0.5; outside top-20: -1.0
    Hand-set in §4.2.2, described as inspired by the NDCG metric; no ablation or sensitivity analysis is reported. These bins directly shape the RL policy and therefore influence all headline results.
  • RL hyperparameters = batch size 64, KL coefficient 1e-3, rollouts 8, learning rate 1e-6, 1 epoch
    Chosen by hand in §5.1.4; no grid search or sensitivity study is presented. They are auxiliary rather than scientific quantities, but they affect the reported results.
assumptions (4)
  • standard math GRPO with KL regularization is a sound RL optimizer for LLM policies (Eq. 3).
    The paper relies on DeepSeek-R1 and DeepSeekMath findings for GRPO stability; it does not re-derive or verify this in the recommendation setting.
  • domain assumption Rationales and rankings produced by DeepSeek-R1-Distill-Qwen-32B are high-quality supervision for teaching slow thinking.
    SFT anchoring (§4.2.1) distills from this teacher; the paper only screens samples by format and NDCG, so this assumption carries much of the method's strength.
  • ad hoc to paper Simulated feedback derived from binary ratings (rating > 3 positive) is a faithful proxy for real user preference evolution.
    The RL and self-reflection loops (§4.1.3, §4.2.2) are trained and evaluated against this simulated signal; there is no online user study or human evaluation to validate the proxy.
  • domain assumption Leave-one-out test items are ranked without access to their ground-truth feedback.
    Required for the evaluation to be valid; the paper does not explicitly state whether the self-reflection loop consumes the held-out feedback before ranking.

how reviews work

0 comments
Cite this review

Pith. "Pith review of STARec: An Efficient Agent Framework for Recommender Systems via Autonomous Deliberate Reasoning." pith.science (2026). https://pith.science/paper/EQUXKG4J

@misc{pith2026250818812,
  author       = {Pith},
  title        = {Pith review of: STARec: An Efficient Agent Framework for Recommender Systems via Autonomous Deliberate Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EQUXKG4J}},
  note         = {Machine review of arXiv:2508.18812}
}
read the original abstract

While modern recommender systems are instrumental in navigating information abundance, they remain fundamentally limited by static user modeling and reactive decision-making paradigms. Current large language model (LLM)-based agents inherit these shortcomings through their overreliance on heuristic pattern matching, yielding recommendations prone to shallow correlation bias, limited causal inference, and brittleness in sparse-data scenarios. We introduce STARec, a slow-thinking augmented agent framework that endows recommender systems with autonomous deliberative reasoning capabilities. Each user is modeled as an agent with parallel cognitions: fast response for immediate interactions and slow reasoning that performs chain-of-thought rationales. To cultivate intrinsic slow thinking, we develop anchored reinforcement training - a two-stage paradigm combining structured knowledge distillation from advanced reasoning models with preference-aligned reward shaping. This hybrid approach scaffolds agents in acquiring foundational capabilities (preference summarization, rationale generation) while enabling dynamic policy adaptation through simulated feedback loops. Experiments on MovieLens 1M and Amazon CDs benchmarks demonstrate that STARec achieves substantial performance gains compared with state-of-the-art baselines, despite using only 0.4% of the full training data.

Figures

Figures reproduced from arXiv: 2508.18812 by the authors.

Figure 1
Figure 1. Overview of the proposed STARec framework. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Performance comparisons of STARec-1.5B w.r.t. dif [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 25 canonical work pages

  1. [1]

    Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michał Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, and Torsten Hoefler. 2024. Graph of thoughts: solving elaborate problems with large language models. In Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Si...

    work page 2024

  2. [2]

    Guangyao Chen, Siwei Dong, Yu Shu, Ge Zhang, Jaward Sesay, Börje Karlsson, Jie Fu, and Yemin Shi. 2024. AutoAgents: A Framework for Automatic Agent Generation. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence. 22–30

    work page 2024

  3. [3]

    Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, Yaxi Lu, Yi-Hsin Hung, Chen Qian, Yujia Qin, Xin Cong, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. 2024. AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors. In The Twelfth International Conference on Learning Representations

    work page 2024

  4. [4]

    Yuheng Cheng, Ceyao Zhang, Zhengwen Zhang, Xiangrui Meng, Sirui Hong, Wenhao Li, Zihao Wang, Zekai Wang, Feng Yin, Junhua Zhao, and Xiuqiang He

    work page

  5. [5]

    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

  6. [6]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai D...

    arXiv 2025

  7. [7]

    Binzong Geng, Zhaoxin Huan, Xiaolu Zhang, Yong He, Liang Zhang, Fajie Yuan, Jun Zhou, and Linjian Mo. 2024. Breaking the Length Barrier: LLM-Enhanced CTR Prediction in Long Textual User Behaviors. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 2311–2315

    work page 2024

  8. [8]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5). In Proceedings of the 16th ACM Conference on Recommender Systems. 299–315

    work page 2022

Show all 45 references
  1. [9]

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. 2024. MetaGPT: Meta Programming for A Multi-Agent Collaborat...

    2024

  2. [10]

    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. In Advances in Information Retrieval: 46th European Conference on Information Retrieval, ECIR 2024, Glasg...

    2024

  3. [11]

    Jian Jia, Yipei Wang, Yan Li, Honggang Chen, Xuehan Bai, Zhaocheng Liu, Jian Liang, Quan Chen, Han Li, Peng Jiang, and Kun Gai. 2024. LEARN: Knowl- edge Adaptation from Large Language Model to Recommendation for Practical Industrial Application. arXiv preprint arXiv:2405.03988 (2024)

    2024 arXiv

  4. [12]

    Daniel Kahneman. 2011. Thinking, fast and slow . macmillan

    2011

  5. [13]

    Shyam Sundar Kannan, Vishnunandan L. N. Venkatesh, and Byung-Cheol Min

  6. [14]

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

    2020 arXiv

  7. [15]

    In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    SMART-LLM: Smart Multi-Agent Robot Task Planning using Large Lan- guage Models. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). 12140–12147

    2024

  8. [16]

    Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wang, Xiuyi Chen, Yingying Zhang, Fei Yin, Jiahua Dong, Zhijiang Guo, Le Song, and Cheng-Lin Liu. 2025. From System 1 to System 2: A Survey of Reasoning Large...

    2025 arXiv

  9. [17]

    Yuanchun Li, Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Yuan, Guohong Liu, Jiacheng Liu, Wenxing Xu, Xiang Wang, Yi Sun, Rui Kong, Yile Wang, Hanfei Geng, Jian Luan, Xuefeng Jin, Zilong Ye, Guanjing Xiong, Fan Zhang, Xiang Li, Mengwei Xu, Zhijun Li, Peng Li, Yang Liu, Ya-Qin Zha...

    2024 arXiv

  10. [18]

    Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun. 2024. ChatDev: Communicative Agents for Software Development. In Proceedings of the 62nd Annual Meeting of the A...

    2024

  11. [19]

    Bernstein

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. 1–22

    2023

  12. [20]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv preprint arXiv:2402.03300 (2024)

    2024 arXiv

  13. [21]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  14. [22]

    Zhongxiang Sun, Zihua Si, Xiaoxue Zang, Kai Zheng, Yang Song, Xiao Zhang, and Jun Xu. 2024. Large Language Models Enhanced Collaborative Filtering. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. 2178–2188

    2024

  15. [23]

    Yashar Talebirad and Amirhossein Nadiri. 2023. Multi-Agent Collaboration: Harnessing the Power of Intelligent LLM Agents. arXiv preprint arXiv:2306.03314 (2023)

    2023 arXiv

  16. [24]

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2024. HybridFlow: A Flexible and Efficient RLHF Framework. arXiv preprint arXiv: 2409.19256 (2024)

    2024 arXiv

  17. [25]

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, Wayne Xin Zhao, Zhewei Wei, and Jirong Wen. 2024. A survey on large language model based autonomous agents. Frontiers of Computer Science 18, 6 (2024), 186345

    2024

  18. [26]

    Yancheng Wang, Ziyan Jiang, Zheng Chen, Fan Yang, Yingxue Zhou, Eunah Cho, Xing Fan, Yanbin Lu, Xiaojiang Huang, and Yingzhen Yang. 2024. RecMind: Large Language Model Powered Agent For Recommendation. In Findings of the Association for Computational Linguistics: NAACL 2024 . ...

    2024

  19. [27]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Con- gcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, Haiqing Guo, Han Zhu, Hao Ding...

    2025 arXiv

  20. [28]

    Dingjun Wu, Jing Zhang, and Xinmei Huang. 2023. Chain of Thought Prompting Elicits Knowledge Augmentation. InFindings of the Association for Computational Linguistics: ACL 2023. 6519–6534

    2023

  21. [29]

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. 2023. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. arXiv pre...

    2023 arXiv

  22. [30]

    Yuling Wang, Changxin Tian, Binbin Hu, Yanhua Yu, Ziqi Liu, Zhiqiang Zhang, Jun Zhou, Liang Pang, and Xiao Wang. 2024. Can Small Language Models be Good Reasoners for Sequential Recommendation?. In Proceedings of the ACM Web Conference 2024. 3876–3887

    2024

  23. [31]

    Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, Rui Zheng, Xiaoran Fan, Xiao Wang, Limao Xiong, Yuhao Zhou, Weiran Wang, Changhao Jiang, Yicheng Zou, Xiangyang Liu, Zhangyue Yin, Shihan Dou, Rongxiang Weng, W...

    2025

  24. [32]

    Violet Xiang, Charlie Snell, Kanishk Gandhi, Alon Albalak, Anikait Singh, Chase Blagden, Duy Phung, Rafael Rafailov, Nathan Lile, Dakota Mahan, Louis Castri- cato, Jan-Philipp Franken, Nick Haber, and Chelsea Finn. 2025. Towards System 2 Reasoning in LLMs: Learning How to Thin...

    2025 arXiv

  25. [33]

    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

  26. [34]

    Griffiths, Yuan Cao, and Karthik Narasimhan

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: deliberate problem solving with large language models. In Proceedings of the 37th International Conference on Neural Information Processing Systems

    2023

  27. [35]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. In The Eleventh International Conference on Learning Representations . STARec: An Efficient Agent Framework for Recom...

    2023

  28. [36]

    Li Yang, Anushya Subbiah, Hardik Patel, Judith Yue Li, Yanwei Song, Reza Mirghaderi, and Vikram Aggarwal. 2024. Item-Language Model for Conver- sational Recommendation. arXiv preprint arXiv:2406.02844 (2024)

    2024 arXiv

  29. [37]

    Junjie Zhang, Yupeng Hou, Ruobing Xie, Wenqi Sun, Julian McAuley, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2024. AgentCF: Collaborative Learning with Autonomous Language Agents for Recommender Systems. In Proceedings of the ACM Web Conference 2024. 3679–3689

    2024

  30. [38]

    Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. 2024. ExpeL: LLM agents are experiential learners. In Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applications of Artific...

    2024

  31. [39]

    An Zhang, Yuxin Chen, Leheng Sheng, Xiang Wang, and Tat-Seng Chua. 2024. On Generative Agents in Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1807–1817

    2024

  32. [40]

    Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, Yingqian Min, Zhichao Feng, Xinyan Fan, Xu Chen, Pengfei Wang, Wendi Ji, Yaliang Li, Xiaoling Wang, and Ji-Rong Wen. 2021. RecBole: Towards a Unified, Com...

    2021

  33. [41]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong W...

    2023 arXiv

  34. [42]

    Wayne Xin Zhao, Yupeng Hou, Xingyu Pan, Chen Yang, Zeyu Zhang, Zihan Lin, Jingsen Zhang, Shuqing Bian, Jiakai Tang, Wenqi Sun, Yushuo Chen, Lanling Xu, Gaowei Zhang, Zhen Tian, Changxin Tian, Shanlei Mu, Xinyan Fan, Xu Chen, and Ji-Rong Wen. 2022. RecBole 2.0: Towards a More U...

    2022

  35. [45]

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: S...

    2024 arXiv

  36. [2017]

    arXiv preprint arXiv:1707.06347 (2017)

    Proximal Policy Optimization Algorithms. arXiv preprint arXiv:1707.06347 (2017)

    2017 arXiv

  37. [2024]

    arXiv preprint arXiv:2401.03428 (2024)

    Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv preprint arXiv:2401.03428 (2024)

    2024 arXiv

Pith tools

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