Pith. sign in

REVIEW 4 major objections 4 minor 4 cited by

Thought-Augmented Planning for LLM-Powered Interactive Recommender Agent

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read TAIRA distills agent and expert experience into planning patterns and beats prior LLM agents on hit rate, ranking quality, and success rate.

desk verdict TAIRA is a solid BoT-style extension for interactive recommendation with consistent simulated wins, but every effectiveness number comes from an author-built GPT-4o user simulator, so the real-user advantage is unproven. read the letter →

arxiv 2506.23485 v1 pith:ZKJHHUZ6 submitted 2025-06-30 cs.CL cs.AIcs.IR

Haocheng Yu , Yaxiong Wu , Hao Wang , Wei Guo , Yong Liu , Yawen Li , Yuyang Ye , Junping Du
show 1 more author
Enhong Chen
This is my paper
classification cs.CLcs.AIcs.IR
keywords interactiverecommendationlargelanguagemodelsmulti-agentsystemsplanningthoughtpatterndistillationusersimulationgeneralization
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

TAIRA is an LLM-powered multi-agent recommender system built around a Manager Agent that decomposes each user request into phased subtasks and executes them with specialist agents. Its planning is supplied by Thought Pattern Distillation (TPD), which converts successful agent runs, expert-corrected failures, and direct human expert guidance into reusable three-part thought patterns. The paper's claim is that this thought augmentation lets the system handle complex intents—usage occasions, matching requests, bundles, ambiguity, and multi-occasion demands—far better than prior LLM agent planning methods. Across three Amazon datasets, TAIRA reports the highest hit rate, NDCG, and success rate on every dataset, with the largest margins on medium and difficult queries. A user-simulator study and a novel-task experiment back the claim that the gains come from reusable high-level planning guidance rather than memorized templates.

What carries the argument

The load-bearing mechanism is Thought Pattern Distillation (TPD), which produces a thought pattern with three parts: a task description used to retrieve the right pattern, a solution description giving conceptual-level guidance, and a thought template giving execution-level steps. Distillation sources are agent successes, agent failures corrected by human experts, and direct human expert experience; a prompting step merges these with any old pattern. The Manager Agent retrieves the top-K similar patterns by matching the query against task descriptions, selects the best one, and folds it into a hierarchical planning loop that revises the plan phase by phase as executor results arrive. This combination lets TAIRA split an ambiguous request into several concrete recommendation lists instead of issuing one premature list, which is the mechanism behind the reported gains on medium and difficult queries.

What would settle it

Run a live-user study with roughly 100 participants per difficulty level on the same three datasets, asking real shoppers to rate TAIRA and the strongest baseline's outputs without the simulator's exact-target bonus or whole-list zeroing; if the success-rate gap on medium and hard queries collapses or reverses, the simulator-driven conclusion fails. A cheaper check is to score the same logged outputs with a second, independently prompted simulator that does not know the target item or the scenario description.

Watch

Extended reading notes

Core claim

The paper's central discovery is that recommending for genuinely complex user intents is a planning problem, not a retrieval problem, and that planning can be improved by distilling reusable thought patterns from experience. TAIRA's Manager Agent matches each incoming query to a stored thought pattern, then generates a phased plan that iterates through hierarchical replanning ($P_{i+1}=H(P_i,I_i)$) until the task is complete. Executor agents search for attributes, retrieve items, and compile one or more ranked lists; the final response can contain multiple lists when the request is ambiguous or spans several scenarios. Measured by an LLM-based user simulator, TAIRA reports success rates up to 0.8824 on Amazon Clothing versus 0.7966 for the best baseline, and it retains a success-rate advantage over the strongest baseline on novel tasks whose matching thought pattern was withheld. The paper attributes this to the multi-scale structure of thought patterns: a solution description gives conceptual guidance that transfers to new tasks, while a thought template gives execution-level steps for familiar ones.

Load-bearing premise

The load-bearing premise is that the authors' LLM user simulator judges recommendations the way real users would; if that simulator's scoring—2 points only for the exact target item and zero for any list whose overall target misses—does not match real satisfaction, the reported advantages of TAIRA over the baselines may not transfer to actual users.

Editorial extensions

If this is right

  • The reported improvement over the strongest baseline is concentrated in medium and difficult queries, so the method's value is in ambiguous, multi-item, and occasion-based requests rather than simple product lookup.
  • Ablations show that removing thought pattern matching causes the largest drop, so retrieving the right past experience is the most critical part of the pipeline.
  • When the matching thought pattern is withheld, TAIRA still out-scores the strongest baseline on novel tasks, indicating that conceptual solution descriptions transfer beyond memorized templates.
  • After a novel task is executed once, it becomes a new thought pattern, which makes the system self-improving in repeated deployment.
  • The multi-agent architecture and replanning loop are designed to extend to multi-turn dialogue, a direction the paper identifies as future work.

Reading between the lines

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

  • Editorial extension: because the user simulator is prompted with the same model family as the recommender and gives 2 points only to the exact target item, part of the measured advantage may reflect the simulator rewarding the system's multi-list, keyword-rich output style; a small live-user study is the direct test.
  • Editorial extension: the thought patterns are distilled on clothing, beauty, and music domains, but the TPD prompt itself is domain-neutral, so the same procedure could be lifted to other recommendation domains with a modest set of expert corrections.
  • Editorial extension: a sharper ablation would compare TAIRA against a variant that retrieves a random thought pattern instead of a matched one; that would separate the contribution of retrieval from the contribution of the pattern content itself.
  • Editorial extension: the single-turn evaluation likely understates the benefit of hierarchical replanning, since the loop's ability to revise plans from intermediate results becomes more valuable when users can ask follow-up questions.
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

4 major / 4 minor

Summary. The paper proposes TAIRA, a multi-agent LLM system for interactive recommendation that combines hierarchical planning with Thought Pattern Distillation (TPD), a method for extracting reusable, multi-scale thought patterns from agent successes, agent failures corrected by human experts, and direct expert experiences. The system is evaluated on three Amazon datasets (Clothing, Beauty, Music) using an author-built LLM user simulator that generates queries at three difficulty levels and scores recommendations. The central claim is that TAIRA significantly outperforms existing baselines (including Reflexion, ReAct, Plan&Solve, and prior interactive recommender agents such as InteRecAgent and MACRec) across HR@10, NDCG@10, and Success Rate, with the largest improvements on medium and difficult queries, while retaining reasonable generalization to novel tasks.

Significance. If the evaluation is trustworthy, TAIRA demonstrates a useful design pattern: combining hierarchical task decomposition with distilled, human-corrected thought patterns can improve LLM agents' handling of ambiguous and multi-faceted recommendation requests. The paper explicitly releases code, which is a strength, and the thought pattern library itself could be reused by other systems. However, the significance is currently bounded by the evaluation methodology: every effectiveness number in Tables 3-5 is produced by a GPT-4o-based user simulator that shares its backbone with the agent being evaluated and that was designed by the same team. The claimed advantage on complex, ambiguous queries is exactly where the simulator's scoring rules most strongly favor the multi-list, scenario-covering behavior that TAIRA's thought templates explicitly encourage. The lack of any human evaluation, inter-annotator agreement, or independent simulator check means the external validity of the central comparison is not yet established.

major comments (4)
  1. [Section 4.2, Figure A7] This is a load-bearing issue because the paper's abstract and conclusions generalize to 'interactive recommendation systems' and 'real-world scenarios,' not merely to the simulated environment.
  2. [Section 3.5, Appendix C, Section 4.1] This issue is load-bearing for the generalization claims in RQ4 and for a fair comparison with baselines.
  3. [Section 5.2, Tables 3-4] This is a presentation and reproducibility issue that directly affects the strength of the central empirical claim.
  4. [Section 3.3, Equation (1)] This is a minor-to-moderate concern about the precision of the claimed contribution, not a fatal flaw.
minor comments (4)
  1. [Section 3.5, line 1] This appears in the sentence describing TPD; please correct.
  2. [Table 6] Please reformat this table so the dataset statistics (number of items, attributes, meta-info fields) are clearly separated and readable.
  3. [Section 5.3] A sentence or two on how each baseline handles multi-step planning would improve readability.
  4. [Figure 4] Please add error bars and a clearer caption, or move the per-difficulty results to a table.

Circularity Check

1 steps flagged · score 4.0 of 10

Moderate self-referentiality: TAIRA's hard-task gains are scored by an author-built GPT-4o simulator whose 'targets together' criterion matches the multi-scenario thought template distilled into TAIRA.

  1. fitted input called prediction [Section 4.2 / Figure A7 (User Simulator); Section 3.5 / Figure B4 (Thought Pattern 4); Appendix C / Figure C2 (Expert Correction)]
    "First, you need to determine whether these targets together can fully meet your requirements. If not, then the recommendation will be considered a failure. ... Based on the knowledge, divide the user’s need into about three specific scenarios. Only one recommendation isn’t enough. ... Note that, there is still a gap between user simulators and real users of interactive recommender systems."

    The claimed advantage on ambiguous/hard queries is not anchored to any independent user signal. The evaluator (Figure A7) defines success for multi-intent queries by asking whether the union of recommended targets 'together can fully meet your requirements,' and a list whose target fails is scored 0 for all items. The thought pattern distilled for exactly this kind of query (Figure B4) instructs the agent to 'divide the user’s need into about three specific scenarios' because 'only one recommendation isn’t enough' — the same success criterion. The expert correction in Figure C2 that produced this template uses the same logic ('necessary to give recommendations in various possible directions').

full rationale

TAIRA's central contribution is a system and an empirical evaluation, not a mathematical derivation, so there is no equation-level circularity. The reported numbers are generated by a single author-built GPT-4o user simulator (Section 4.2, Figure A7) applied uniformly to all baselines; within that protocol the comparison is self-consistent and no fitted parameters are hidden. I find no load-bearing self-citation: the citations to [51]/[53] for query construction and [49] for thought-pattern inspiration are methodological, and the method explicitly extends [49]. The main concern is partial self-referentiality of the evaluation: the simulator's success condition for ambiguous requests ('targets together can fully meet your requirements') is mirrored by the TPD thought template for exactly those requests ('divide the user’s need into about three specific scenarios. Only one recommendation isn’t enough'), and the expert corrections that shape the templates use the same multi-direction rule. Because the thought patterns and the evaluator were authored by the same team, the large gains on hard tasks are at least partly a measure of the agent following a scoring rule already embedded in its prompts, rather than of real-user satisfaction. The paper candidly admits the simulator–real-user gap, but no human study or independent simulator is provided to break the self-reference. Weighing all factors, this is a moderate partial-circularity flag, not a full reduction.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The central claim rests on a hand-built thought-pattern library, a hand-tuned LLM user simulator, and a set of design constants such as top-5 retrieval and about-three scenario decomposition. None of these are fitted to an external benchmark, and they are not independently validated against real users, so they carry the main evidential burden beyond the code.

free parameters (3)
  • Top-K retrieved thought patterns = K=5 (top-5)
    The Manager Agent retrieves the top-5 thought patterns by similarity before selecting the most relevant one; K is a hand-chosen constant with no ablation over K.
  • Number of target scenarios or clothing categories in thought templates = about three
    Thought templates instruct the agent to divide user needs into about three specific scenarios or clothing types; this affects recommendation coverage and output length and is not varied in experiments.
  • User simulator scoring weights = 0.5, 1, 2 points
    The simulator gives 1 point for a relevant item, 2 points for the exact sample product, and 0.5 for items that meet requirements but not preferences; these hand-set weights shape all reported metrics.
assumptions (3)
  • domain assumption An LLM user simulator can stand in for real user judgments of recommendation quality.
    Section 4.2 uses GPT-4o to act as a shopper and score recommendation lists; the paper itself acknowledges a gap between simulators and real users, yet all headline results depend on this assumption.
  • domain assumption The last interaction in a user's history is the correct target item, and atomic queries built from it represent authentic user needs.
    Section 4.1 follows common sequential recommendation practice [51], but this modeling choice defines the evaluation data and may not reflect how real users phrase requests.
  • domain assumption High-level thought patterns distilled from successful and corrected experiences improve planning on new queries without retraining.
    This is the core hypothesis of TPD in Section 3.5; the ablations support it only within the author-designed simulated environment, not against independent human judgment.
invented entities (1)
  • Thought Pattern library
    purpose: A persistent set of distilled, human-readable planning templates (task description, solution description, thought template) that the Manager Agent retrieves and follows when decomposing user requests.
    Appendix B shows several examples, but the library has no external falsifiable handle beyond the agent's own prompts; its value is only measured through the author-designed user simulator.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Thought-Augmented Planning for LLM-Powered Interactive Recommender Agent." pith.science (2026). https://pith.science/paper/ZKJHHUZ6

@misc{pith2026250623485,
  author       = {Pith},
  title        = {Pith review of: Thought-Augmented Planning for LLM-Powered Interactive Recommender Agent},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZKJHHUZ6}},
  note         = {Machine review of arXiv:2506.23485}
}
read the original abstract

Interactive recommendation is a typical information-seeking task that allows users to interactively express their needs through natural language and obtain personalized recommendations. Large language model-powered (LLM-powered) agents have become a new paradigm in interactive recommendations, effectively capturing users' real-time needs and enhancing personalized experiences. However, due to limited planning and generalization capabilities, existing formulations of LLM-powered interactive recommender agents struggle to effectively address diverse and complex user intents, such as intuitive, unrefined, or occasionally ambiguous requests. To tackle this challenge, we propose a novel thought-augmented interactive recommender agent system (TAIRA) that addresses complex user intents through distilled thought patterns. Specifically, TAIRA is designed as an LLM-powered multi-agent system featuring a manager agent that orchestrates recommendation tasks by decomposing user needs and planning subtasks, with its planning capacity strengthened through Thought Pattern Distillation (TPD), a thought-augmentation method that extracts high-level thoughts from the agent's and human experts' experiences. Moreover, we designed a set of user simulation schemes to generate personalized queries of different difficulties and evaluate the recommendations based on specific datasets. Through comprehensive experiments conducted across multiple datasets, TAIRA exhibits significantly enhanced performance compared to existing methods. Notably, TAIRA shows a greater advantage on more challenging tasks while generalizing effectively on novel tasks, further validating its superiority in managing complex user intents within interactive recommendation systems. The code is publicly available at:https://github.com/Alcein/TAIRA.

Figures

Figures reproduced from arXiv: 2506.23485 by the authors.

Figure 1
Figure 1. Examples of recommendation involving diverse [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. TAIRA’s overall architecture. The Manager Agent locates a matching Thought Pattern for each user query, then [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The prompt of Thought Pattern Distillation (TPD). [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: SR of Reflexion and TAIRA across three difficulty levels. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Executing Case of TAIRA. REFERENCES [1] Suha Khalil Assayed, Khaled Shaalan, and Manar Alkhatib. 2022. A Chatbot Intent Classifier for Supporting High School Students. EAI Endorsed Transactions on Scalable Information Systems 10, 3 (Dec. 2022), e1. https://doi.org/10.4…

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Attacking and Defending Multi-Agent Collaborative Filtering Systems Through Connectivity

    cs.IR 2026-08 conditional novelty 6.0 of 10

    In agent-based collaborative filtering, attack spread and privacy leakage grow with interaction connectivity, but the effect is asymmetric between user and item agents and differs between early and steady-state phases.

  2. Autonomous Information Seeking: A Roadmap for Agentic Recommender Systems

    cs.IR 2026-07 accept novelty 5.0 of 10

    Agentic recommender systems are organized by agent role (assisted, as-recommender, as-simulator) crossed with autonomy levels L2–L5, yielding a roadmap of architectures, evaluation limits, and open challenges.

  3. Why Thinking Hurts: Diagnosing and Rectifying Linguistic Inertia in Large Language Models for Recommendation

    cs.IR 2026-02 conditional novelty 5.0 of 10

    Chain-of-thought reasoning degrades semantic-ID recommendation accuracy through 'linguistic inertia,' and a training-free compression-plus-contrastive decoding fix restores and often improves accuracy.

  4. SelfAug: Mitigating Catastrophic Forgetting in Retrieval-Augmented Generation via Distribution Self-Alignment

    cs.CL 2025-09 conditional novelty 5.0 of 10

    Adding a KL penalty between fine-tuned and original model logits on input tokens during RAG fine-tuning reduces catastrophic forgetting while preserving downstream performance.

Reference graph

Works this paper leans on

63 extracted references · 19 canonical work pages · cited by 4 Pith papers

  1. [1]

    Suha Khalil Assayed, Khaled Shaalan, and Manar Alkhatib. 2022. A Chatbot Intent Classifier for Supporting High School Students. EAI Endorsed Transactions on Scalable Information Systems 10, 3 (Dec. 2022), e1. https://doi.org/10.4108/ eetsis.v10i2.2948

    work page 2022

  2. [2]

    Anna Bodonhelyi, Efe Bozkir, Shuo Yang, Enkelejda Kasneci, and Gjergji Kasneci

    work page

  3. [3]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin...

    arXiv 2020

  4. [4]

    CB Chandrakala, Rohit Bhardwaj, and Chetana Pujari. 2024. An intent recognition pipeline for conversational AI. International Journal of Information Technology 16, 2 (2024), 731–743

    work page 2024

  5. [5]

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu

  6. [6]

    Jiabao Fang, Shen Gao, Pengjie Ren, Xiuying Chen, Suzan Verberne, and Zhaochun Ren. 2024. A Multi-Agent Conversational Recommender System. arXiv:cs.IR/2402.01135 https://arxiv.org/abs/2402.01135

    arXiv 2024

  7. [7]

    arXiv:cs.CL/2402.03216

    BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. arXiv:cs.CL/2402.03216

    arXiv

  8. [8]

    Fernando Fernández-Martínez, Cristina Luna-Jiménez, Ricardo Kleinlein, David Griol, Zoraida Callejas, and Juan Manuel Montero. 2022. Fine-tuning BERT models for intent recognition using a frequency cut-off strategy for domain- specific vocabulary extension. Applied Sciences 12, 3 (2022), 1610

    work page 2022

Show all 63 references
  1. [9]

    Yue Feng, Shuchang Liu, Zhenghai Xue, Qingpeng Cai, Lantao Hu, Peng Jiang, Kun Gai, and Fei Sun. 2023. A Large Language Model Enhanced Conversational Recommender System. arXiv:cs.IR/2308.06212 https://arxiv.org/abs/2308.06212

    2023 arXiv

  2. [10]

    Chongming Gao, Wenqiang Lei, Xiangnan He, Maarten de Rijke, and Tat-Seng Chua. 2021. Advances and challenges in conversational recommender systems: A survey. AI open 2 (2021), 100–126

    2021

  3. [11]

    Mauajama Firdaus, Hitesh Golchha, Asif Ekbal, and Pushpak Bhattacharyya. 2021. A deep multi-task model for dialogue act classification, intent detection and slot filling. Cognitive Computation 13 (2021), 626–645

    2021

  4. [12]

    Wei Guo, Hao Wang, Luankang Zhang, Jin Yao Chin, Zhongzhou Liu, Kai Cheng, Qiushi Pan, Yi Quan Lee, Wanqi Xue, Tingjia Shen, et al . 2024. Scal- ing New Frontiers: Insights into Large Recommendation Models. arXiv preprint arXiv:2412.00714 (2024)

    2024 arXiv

  5. [13]

    Hongchao Gu, Dexun Li, Kuicai Dong, Hao Zhang, Hang Lv, Hao Wang, Defu Lian, Yong Liu, and Enhong Chen. 2025. RAPID: Efficient Retrieval-Augmented Long Text Generation with Writing Planning and Information Discovery. arXiv preprint arXiv:2503.00751 (2025)

    2025 arXiv

  6. [14]

    Wanwei He, Yinpei Dai, Min Yang, Jian Sun, Fei Huang, Luo Si, and Yongbin Li

  7. [15]

    Chen He, Denis Parra, and Katrien Verbert. 2016. Interactive recommender systems: A survey of the state of the art and future research challenges and opportunities. Expert Systems with Applications 56 (2016), 9–27. https://doi.org/ 10.1016/j.eswa.2016.02.013

    2016 doi

  8. [16]

    Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024. Understanding the planning of LLM agents: A survey. arXiv:cs.AI/2402.02716 https://arxiv.org/abs/ 2402.02716

    2024 arXiv

  9. [17]

    Dietmar Jannach, Ahtsham Manzoor, Wanling Cai, and Li Chen. 2021. A Survey on Conversational Recommender Systems. ACM Comput. Surv. 54, 5, Article 105 (May 2021), 36 pages. https://doi.org/10.1145/3453154

    2021 doi

  10. [18]

    Xu Huang, Jianxun Lian, Yuxuan Lei, Jing Yao, Defu Lian, and Xing Xie. 2024. Recommender AI Agent: Integrating Large Language Models for Interactive Recommendations. arXiv:cs.IR/2308.16505 https://arxiv.org/abs/2308.16505

    2024 arXiv

  11. [19]

    Khang Nhut Lam, Nam Nhat Le, and Jugal Kalita. 2021. Building a Chatbot on a Closed Domain using RASA. In Proceedings of the 4th International Conference on Natural Language Processing and Information Retrieval (NLPIR ’20) . Association for Computing Machinery, New York, NY, U...

    2021

  12. [20]

    Wenqiang Lei, Xiangnan He, Yisong Miao, Qingyun Wu, Richang Hong, Min-Yen Kan, and Tat-Seng Chua. 2020. Estimation-Action-Reflection: Towards Deep Interaction Between Conversational and Recommender Systems. In Proceedings of the 13th International Conference on Web Search and ...

    2020

  13. [21]

    Mourad Jbene, Smail Tigani, Rachid Saadane, and Abdellah Chehri. 2022. An LSTM-based Intent Detector for Conversational Recommender Systems. In 2022 IEEE 95th Vehicular Technology Conference: (VTC2022-Spring). IEEE, Helsinki, 1–5. https://doi.org/10.1109/VTC2022-Spring54318.20...

    2022

  14. [22]

    Chuang Li, Yang Deng, Hengchang Hu, Min-Yen Kan, and Haizhou Li. 2024. Incor- porating External Knowledge and Goal Guidance for LLM-based Conversational Recommender Systems. arXiv:cs.CL/2405.01868 https://arxiv.org/abs/2405.01868

    2024 arXiv

  15. [23]

    Chaofan Li, Zheng Liu, Shitao Xiao, and Yingxia Shao. 2023. Making Large Lan- guage Models A Better Foundation For Dense Retrieval. arXiv:cs.CL/2312.15503

    2023

  16. [24]

    Wenqiang Lei, Gangyi Zhang, Xiangnan He, Yisong Miao, Xiang Wang, Liang Chen, and Tat-Seng Chua. 2020. Interactive Path Reasoning on Graph for Conversational Recommendation. In Proceedings of the 26th ACM SIGKDD In- ternational Conference on Knowledge Discovery & Data Mining (...

    2020

  17. [25]

    Weiwen Liu, Wei Guo, Yong Liu, Ruiming Tang, and Hao Wang. 2023. User Behavior Modeling with Deep Learning for Recommendation: Recent Advances. In Proceedings of the 17th ACM Conference on Recommender Systems . 1286–1287

    2023

  18. [26]

    Kelong Mao, Zhicheng Dou, Fengran Mo, Jiewen Hou, Haonan Chen, and Hongjin Qian. 2023. Large Language Models Know Your Contextual Search Intent: A Prompting Framework for Conversational Search. arXiv:cs.IR/2303.06573 https: //arxiv.org/abs/2303.06573

    2023 arXiv

  19. [27]

    Raymond Li, Samira Kahou, Hannes Schulz, Vincent Michalski, Laurent Charlin, and Chris Pal. 2018. Towards deep conversational recommendations. In Proceed- ings of the 32nd International Conference on Neural Information Processing Systems (NIPS’18). Curran Associates Inc., Red ...

    2018

  20. [28]

    Rizou, A

    S. Rizou, A. Theofilatos, A. Paflioti, E. Pissari, I. Varlamis, G. Sarigiannidis, and K.Ch. Chatzisavvas. 2023. Efficient intent classification and entity recognition for university administrative services employing deep learning models. Intelligent Systems with Applications 1...

    2023 doi

  21. [29]

    Amrita Saha, Mitesh Khapra, and Karthik Sankaranarayanan. [n.d.]. Towards Building Large Scale Multimodal Domain-Aware Conversation Systems. 32, 1 ([n. d.]). https://doi.org/10.1609/aaai.v32i1.11331

    doi

  22. [30]

    Julian McAuley. [2014]. Amazon product data. https://jmcauley.ucsd.edu/data/ amazon/index_2014.html

    2014

  23. [31]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: language agents with verbal reinforcement learn- ing. In Advances in Neural Information Processing Systems , A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine ...

    2023

  24. [32]

    Yubo Shu, Haonan Zhang, Hansu Gu, Peng Zhang, Tun Lu, Dongsheng Li, and Ning Gu. 2024. RAH! RecSys–Assistant–Human: A Human-Centered Recommen- dation Framework With LLM Agents. IEEE Transactions on Computational Social Systems 11, 5 (2024), 6759–6770. https://doi.org/10.1109/T...

    2024

  25. [33]

    Tingjia Shen, Hao Wang, Chuan Qin, Ruijun Sun, Yang Song, Defu Lian, Heng- shu Zhu, and Enhong Chen. 2025. GenKI: Enhancing Open-Domain Question Answering with Knowledge Integration and Controllable Generation in Large Language Models. arXiv:cs.CL/2505.19660 https://arxiv.org/...

    2025

  26. [34]

    Pengfei Sun, Dingjie Song, Yawen Ouyang, Zhen Wu, and Xinyu Dai. 2023. Episode-Based Prompt Learning for Any-Shot Intent Detection. In Natural Lan- guage Processing and Chinese Computing , Fei Liu, Nan Duan, Qingting Xu, and Yu Hong (Eds.). Springer Nature Switzerland, Cham, 29–41

    2023

  27. [35]

    Yueming Sun and Yi Zhang. 2018. Conversational Recommender System. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval (SIGIR ’18) . Association for Computing Machinery, New York, NY, USA, 235–244. https://doi.org/10.1145/3209978.3210002

    2018

  28. [36]

    Xiaoshuai Song, Keqing He, Pei Wang, Guanting Dong, Yutao Mou, Jingang Wang, Yunsen Xian, Xunliang Cai, and Weiran Xu. 2023. Large Language Models Meet Open-World Intent Discovery and Recognition: An Evaluation of ChatGPT. arXiv:cs.CL/2310.10176 https://arxiv.org/abs/2310.10176

    2023 arXiv

  29. [37]

    Quan Tu, Shen Gao, Yanran Li, Jianwei Cui, Bin Wang, and Rui Yan. 2022. Conver- sational Recommendation via Hierarchical Information Modeling. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’22) . Assoc...

    2022

  30. [38]

    Dobrescu, Adrián Sánchez-Miguel, Guillermo Cubero, and Ramón Hervás

    Laura Villa, David Carneros-Prado, Cosmin C. Dobrescu, Adrián Sánchez-Miguel, Guillermo Cubero, and Ramón Hervás. 2024. Comparative Analysis of Generic and Fine-Tuned Large Language Models for Conversational Agent Systems. Ro- botics 13, 5 (2024), 1. https://doi.org/10.3390/ro...

    2024 doi

  31. [39]

    Andrew Trotman, Antti Puurula, and Blake Burgess. 2014. Improvements to BM25 and Language Models Examined. InProceedings of the 19th Australasian Document Computing Symposium (ADCS ’14) . Association for Computing Machinery, New York, NY, USA, 58–65. https://doi.org/10.1145/26...

    2014

  32. [40]

    Hao Wang, Wei Guo, Luankang Zhang, Jin Yao Chin, Yufei Ye, Huifeng Guo, Yong Liu, Defu Lian, Ruiming Tang, and Enhong Chen. 2025. Generative Large Recommendation Models: Emerging Trends in LLMs for Recommendation. arXiv preprint arXiv:2502.13783 (2025)

    2025 arXiv

  33. [41]

    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:cs.CL/2305.04091 https://arxiv.org/abs/2305.04091

    2023 arXiv

  34. [42]

    Ivan Vulić, Pei-Hao Su, Sam Coope, Daniela Gerz, Paweł Budzianowski, Iñigo Casanueva, Nikola Mrkšić, and Tsung-Hsien Wen. 2021. ConvFiT: Conversational Fine-Tuning of Pretrained Language Models. arXiv:cs.CL/2109.10126 https: //arxiv.org/abs/2109.10126

    2021 arXiv

  35. [43]

    Xiaolei Wang, Kun Zhou, Ji-Rong Wen, and Wayne Xin Zhao. 2022. Towards Unified Conversational Recommender Systems via Knowledge-Enhanced Prompt Learning. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’22) . Association for Computi...

    2022

  36. [44]

    Zhefan Wang, Yuanqing Yu, Wendi Zheng, Weizhi Ma, and Min Zhang. 2024. MACRec: A Multi-Agent Collaboration Framework for Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’24) . Association...

    2024 doi

  37. [45]

    Xiaolei Wang, Xinyu Tang, Xin Zhao, Jingyuan Wang, and Ji-Rong Wen. 2023. Rethinking the Evaluation for Conversational Recommendation in the Era of Large Language Models. In Proceedings of the 2023 Conference on Empirical Meth- ods in Natural Language Processing . Association ...

    2023 doi

  38. [46]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems , S. Koyejo, S. Mohamed, A. Agar-...

    2022

  39. [47]

    Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. 2021. Fashion IQ: A New Dataset Towards Retrieving Images by Natural Language Feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVP...

    2021

  40. [48]

    Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M

    Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. Finetuned Language Models Are Zero-Shot Learners. arXiv:cs.CL/2109.01652 https://arxiv.org/abs/2109.01652

    2022 arXiv

  41. [49]

    Gonzalez, and Bin Cui

    Ling Yang, Zhaochen Yu, Tianjun Zhang, Shiyi Cao, Minkai Xu, Wentao Zhang, Joseph E. Gonzalez, and Bin Cui. 2024. Buffer of Thoughts: Thought-Augmented Reasoning with Large Language Models. arXiv:cs.CL/2406.04271 https://arxiv. org/abs/2406.04271

    2024 arXiv

  42. [50]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. ReAct: Synergizing Reasoning and Acting in Language Models. arXiv preprint arXiv:2210.03629 (2022)

    2022 arXiv

  43. [51]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al . 2024. A survey on large language models for recommendation. World Wide Web 27, 5 (2024), 60

    2024

  44. [52]

    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 (WWW ’24) . Association for...

    2024

  45. [53]

    Jiaqing Zhang, Mingjia Yin, Hao Wang, Yawen Li, Yuyang Ye, Xingyu Lou, Junping Du, and Enhong Chen. 2025. TD3: Tucker Decomposition Based Dataset Distilla- tion Method for Sequential Recommendation. arXiv preprint arXiv:2502.02854 (2025)

    2025 arXiv

  46. [54]

    Mingjia Yin, Hao Wang, Wei Guo, Yong Liu, Suojuan Zhang, Sirui Zhao, Defu Lian, and Enhong Chen. 2024. Dataset regeneration for sequential recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3954–3965

    2024

  47. [55]

    Kun Zhou, Wayne Xin Zhao, Shuqing Bian, Yuanhang Zhou, Ji-Rong Wen, and Jingsong Yu. 2020. Improving Conversational Recommender Systems via Knowledge Graph based Semantic Fusion. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Minin...

    2020

  48. [56]

    Kun Zhou, Wayne Xin Zhao, Hui Wang, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. Leveraging Historical Interaction Data for Improving Conversational Recommender System. In Proceedings of the 29th ACM International Conference on Information & Knowledge Mana...

    2020

  49. [57]

    Bruce Croft

    Yongfeng Zhang, Xu Chen, Qingyao Ai, Liu Yang, and W. Bruce Croft. 2018. Towards Conversational Search and Recommendation: System Ask, User Respond. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management (CIKM ’18) . Association for Com...

    2018

  50. [58]

    Lixi Zhu, Xiaowen Huang, and Jitao Sang. 2024. How Reliable is Your Simulator? Analysis on the Limitations of Current LLM-based User Simulators for Conver- sational Recommendation. In Companion Proceedings of the ACM Web Conference 2024 (WWW ’24) . Association for Computing Ma...

    2024

  51. [59]

    Yuqi Zhu, Shuofei Qiao, Yixin Ou, Shumin Deng, Ningyu Zhang, Shiwei Lyu, Yue Shen, Lei Liang, Jinjie Gu, and Huajun Chen. 2024. Knowagent: Knowledge- augmented planning for llm-based agents. arXiv preprint arXiv:2403.03101 (2024)

    2024 arXiv

  52. [60]

    Yuanhang Zhou, Kun Zhou, Wayne Xin Zhao, Cheng Wang, Peng Jiang, and He Hu. 2022. C 2-CRS: Coarse-to-Fine Contrastive Learning for Conversational Rec- ommender System. InProceedings of the Fifteenth ACM International Conference on Web Search and Data Mining (WSDM ’22) . Associ...

    2022

  53. [63]

    content” The next agent to complete this task is: “next agent name

    Jie Zou, Yifan Chen, and Evangelos Kanoulas. 2020. Towards Question-based Recommender Systems. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’20) . Association for Computing Machinery, New York, NY, US...

    2020

  54. [2022]

    In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’22)

    Unified Dialog Model Pre-training for Task-Oriented Dialog Understanding and Generation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’22) . Association for Computing Machinery, New York, NY, USA, 187...

  55. [2024]

    arXiv:cs.HC/2402.02136 https://arxiv.org/abs/2402

    User Intent Recognition and Satisfaction with Large Language Models: A User Study with ChatGPT. arXiv:cs.HC/2402.02136 https://arxiv.org/abs/2402. 02136

    arXiv

Pith tools

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