arxiv: 2604.14972 · v2 · submitted 2026-04-16 · 💻 cs.IR

Recognition: unknown

SAGER: Self-Evolving User Policy Skills for Recommendation Agent

Zhen Tao , Riwei Lai , Chenyun Yu , Weixin Chen , Li Chen , Beibei Kong , Lei Cheng , Chengxiang Zhuo

show 2 more authors

Zang Li Qingqiang Sun

Authors on Pith no claims yet

Pith reviewed 2026-05-10 09:51 UTC · model grok-4.3

classification 💻 cs.IR

keywords LLM recommendation agentsself-evolving policiespersonalized reasoningchain-of-thoughtuser policy skillscontrastive reasoningrecommendation systems

0 comments

The pith

Personalizing reasoning rules in recommendation agents yields improvements distinct from memory updates.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

LLM-based recommendation agents typically personalize only through per-user memory while using a shared static reasoning prompt for all. This leaves the decision logic unchanged even after repeated failures. SAGER addresses this by giving each user an evolving policy skill, a natural language document of personalized principles that updates via interaction. It employs a two-representation architecture, contrastive chain-of-thought for flaw diagnosis, and skill-enhanced listwise reasoning. On four benchmarks, it reaches state-of-the-art results with gains that do not overlap with memory improvements, establishing that tailoring the reasoning process is a separate path to better performance.

Core claim

By equipping each user with a dedicated evolving policy skill that encodes personalized decision principles, SAGER allows the agent to interrogate and refine its reasoning logic upon failures through incremental contrastive chain-of-thought, rather than only accumulating memory of preferences, resulting in performance gains that are orthogonal to those from memory alone.

What carries the argument

The policy skill, a structured natural-language document encoding personalized decision principles that evolves continuously, supported by a two-representation architecture decoupling rich evolution from minimal inference injection and an incremental contrastive chain-of-thought engine.

Load-bearing premise

The incremental contrastive chain-of-thought engine can reliably diagnose reasoning flaws by contrasting accepted and unchosen items while the two-representation architecture keeps evolution separate from inference without introducing new biases or instability.

What would settle it

A controlled experiment on the same benchmarks where the policy skill evolution or contrastive diagnosis is disabled, showing no significant accuracy drop compared to the memory-only baseline.

Figures

Figures reproduced from arXiv: 2604.14972 by Beibei Kong, Chengxiang Zhuo, Chenyun Yu, Lei Cheng, Li Chen, Qingqiang Sun, Riwei Lai, Weixin Chen, Zang Li, Zhen Tao.

**Figure 1.** Figure 1: (a) Existing recommendation agents personalize what they know (memory) but reason about every user with a fixed, shared policy. (b) SAGER equips each user with a self-evolving policy skill that personalizes how the agent reasons, not just what it remembers. We share this vision, and observe that a key dimension of improvement remains largely unexplored. Current agents, including the self-evolving ones, foc… view at source ↗

**Figure 2.** Figure 2: The SAGER framework operates in four stages per interaction: ⃝1 Retrieve: the Memory Manager (LLMMem) curates top-k neighbors from the semantic graph and synthesizes collaborative facets Mcollab; ⃝2 Extract: the full policy skill repository S (t−1) u is distilled by LLMExt into a slim working skill S (t−1),slim u (∼30 tokens), respecting the Cognitive Injection Boundary; ⃝3 Reason: the Reasoning Agent (LLM… view at source ↗

**Figure 3.** Figure 3: iAgent with and without SAGER skill evolution across four datasets. Consistent improvements across all datasets demonstrate that skill evolution is effective even for static-profile agents. 0.40 0.50 0.60 0.70 0.80 0.90 Score Hit@1 Hit@3 NDCG@5 Hit@1 Hit@3 NDCG@5 Hit@1 Hit@3 NDCG@5 Hit@1 Hit@3 NDCG@5 Yelp GoodReads MovieTV Books Baseline +Skill Evolution +10.9% +5.7% +4.7% +11.0% +6.1% +5.0% +9.3% +5.2% +… view at source ↗

**Figure 4.** Figure 4: i2Agent with and without SAGER skill evolution. Skill evolution provides additive gains on top of dynamic profile updates, confirming orthogonality between profile-level memory and policy-level skill. 3.4 Ablation Study (RQ3) Component Ablation. SAGER integrates four interdependent designs: the policy skill itself, statistical initialization, listwise reasoning, and incremental CoT evolution. A natural qu… view at source ↗

**Figure 5.** Figure 5: Vanilla LLM with and without SAGER skill evolution. Skill evolution yields consistent improvements across all datasets, demonstrating the module’s practical utility. pronounced degradation and Books the smallest (−1.9%). The drop is concentrated on H@1, with H@3/H@5 degrading systematically less, reinforcing the asymmetric gain pattern: policy skills primarily improve top-rank decisional sharpness. • Listw… view at source ↗

**Figure 6.** Figure 6: Effect of warmup rounds τ on H@1, H@3, and H@5 (one panel per metric; bar colors distinguish datasets). τ = 2 achieves the best performance across most metrics and datasets (yellow highlight). 3.5 Design Sensitivity Analysis (RQ4) In this section, we examine how sensitive SAGER is to two key hyperparameters that govern the policy skill lifecycle: the number of warmup rounds τ (controlling evolution depth) … view at source ↗

**Figure 7.** Figure 7: Effect of slim skill injection length on H@1, H@3, and H@5 (one panel per metric; line [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

read the original abstract

Large language model (LLM) based recommendation agents personalize what they know through evolving per-user semantic memory, yet how they reason remains a universal, static system prompt shared identically across all users. This asymmetry is a fundamental bottleneck: when a recommendation fails, the agent updates its memory of user preferences but never interrogates the decision logic that produced the failure, leaving its reasoning process structurally unchanged regardless of how many mistakes it accumulates. To address this bottleneck, we propose SAGER (Self-Evolving Agent for Personalized Recommendation), the first recommendation agent framework in which each user is equipped with a dedicated policy skill, a structured natural-language document encoding personalized decision principles that evolves continuously through interaction. SAGER introduces a two-representation skill architecture that decouples a rich evolution substrate from a minimal inference-time injection, an incremental contrastive chain-of-thought engine that diagnoses reasoning flaws by contrasting accepted against unchosen items while preserving accumulated priors, and skill-augmented listwise reasoning that creates fine-grained decision boundaries where the evolved skill provides genuine discriminative value. Experiments on four public benchmarks demonstrate that SAGER achieves state-of-the-art performance, with gains orthogonal to memory accumulation, confirming that personalizing the reasoning process itself is a qualitatively distinct source of recommendation improvement.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SAGER adds per-user evolving policy skills to LLM rec agents to personalize reasoning, but the orthogonality claim and SOTA results need the full experiments to hold up.

read the letter

The punchline is that SAGER equips each user with an evolving policy skill to personalize the agent's reasoning process itself, on top of memory. This is new in the combination of per-user structured skills, the decoupled two-representation design, contrastive CoT diagnosis, and skill-augmented listwise reasoning. The paper does well at spelling out the limitation of static reasoning prompts in current LLM recommendation agents and at building a framework that lets the decision logic adapt through interaction. The experiments claim state-of-the-art results with gains orthogonal to memory on four benchmarks. If the full paper backs this with solid ablations and controls, that would be a real step. But the abstract alone gives no numbers or details, so the strength of the evidence is still unclear. A potential issue is whether the contrastive engine truly diagnoses reasoning flaws or just picks up on preference patterns already covered by memory. Since unchosen items in rec data are implicit, the updates could overlap rather than add a distinct source of improvement. The two-representation architecture aims to prevent this, but it needs careful testing to confirm no new biases creep in. This paper is aimed at researchers in LLM-based recommendation and agent personalization. Readers working on evolving agent behaviors would find the architectural ideas worth considering. It deserves a serious referee because the problem it targets is real and the proposed solution is coherent. I would recommend sending it to peer review so the authors can provide the missing experimental details and address questions about the orthogonality.

Referee Report

3 major / 2 minor

Summary. The paper proposes SAGER, an LLM-based recommendation agent framework that equips each user with an evolving per-user 'policy skill' (a structured natural-language document of personalized decision principles). It introduces a two-representation skill architecture (decoupling evolution from inference), an incremental contrastive chain-of-thought engine (diagnosing flaws via accepted vs. unchosen items), and skill-augmented listwise reasoning. The central claim is that SAGER achieves SOTA on four public benchmarks with gains orthogonal to memory accumulation, establishing personalizing the reasoning process itself as a qualitatively distinct improvement source.

Significance. If the experimental results and orthogonality hold under rigorous validation, this would be a meaningful contribution to LLM recommendation agents by moving beyond memory-only personalization to reasoning personalization. The two-representation architecture and contrastive diagnosis mechanism are conceptually clean ideas that could generalize; credit is due for framing the asymmetry between memory and reasoning as a bottleneck and for attempting to isolate a new improvement axis.

major comments (3)

[Abstract and §4] Abstract and §4 (Experiments): The abstract asserts SOTA performance and 'orthogonal gains' on four benchmarks but supplies no numerical metrics, baseline comparisons, ablation tables, or statistical tests. Without these, the central claim that reasoning personalization is 'qualitatively distinct' cannot be evaluated and the orthogonality conclusion does not follow from the reported evidence.
[§3.2] §3.2 (Incremental contrastive CoT engine): The engine diagnoses reasoning flaws by contrasting accepted against unchosen items while 'preserving accumulated priors.' In implicit-feedback recommendation data, unchosen items are typically not explicit negatives but ranking artifacts or unobserved preferences; this risks the contrastive updates merely re-encoding user-item affinity signals already captured by memory, violating the orthogonality assumption. A targeted ablation or diagnostic experiment (e.g., comparing skill evolution with vs. without memory) is required to confirm the diagnosed flaws are reasoning-specific rather than redundant.
[§3.1] §3.1 (Two-representation skill architecture): The claim that the architecture 'keeps evolution separate from inference without introducing new biases or instability' is load-bearing for the orthogonality result. No analysis is provided on whether the minimal inference-time injection leaks evolved principles back into the base model or creates distribution shift over long interaction sequences.

minor comments (2)

[§2 and §3] The term 'policy skill' is introduced as a novel construct; provide a concise formal definition or pseudocode showing its structure, update rule, and exact injection point into the LLM prompt to avoid ambiguity with standard system prompts or memory entries.
[Figures and Tables] Figure 1 (framework overview) and Table 1 (benchmark results) would benefit from clearer labeling of the two representations and explicit indication of which rows isolate the skill-evolution component.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, providing clarifications and indicating revisions made to the manuscript.

read point-by-point responses

Referee: [Abstract and §4] Abstract and §4 (Experiments): The abstract asserts SOTA performance and 'orthogonal gains' on four benchmarks but supplies no numerical metrics, baseline comparisons, ablation tables, or statistical tests. Without these, the central claim that reasoning personalization is 'qualitatively distinct' cannot be evaluated and the orthogonality conclusion does not follow from the reported evidence.

Authors: We agree that the abstract would be strengthened by including key numerical results to support the claims upfront. In the revised manuscript, we have updated the abstract to report specific performance metrics (e.g., relative improvements over baselines), mention of ablation studies, and reference to statistical significance tests. The full set of tables, baseline comparisons, ablations, and statistical analyses remain detailed in Section 4, which we believe provides sufficient evidence for the SOTA results and orthogonality of gains from reasoning personalization. revision: yes
Referee: [§3.2] §3.2 (Incremental contrastive CoT engine): The engine diagnoses reasoning flaws by contrasting accepted against unchosen items while 'preserving accumulated priors.' In implicit-feedback recommendation data, unchosen items are typically not explicit negatives but ranking artifacts or unobserved preferences; this risks the contrastive updates merely re-encoding user-item affinity signals already captured by memory, violating the orthogonality assumption. A targeted ablation or diagnostic experiment (e.g., comparing skill evolution with vs. without memory) is required to confirm the diagnosed flaws are reasoning-specific rather than redundant.

Authors: This concern about implicit feedback data is well-taken, as unchosen items may reflect ranking artifacts rather than explicit negatives. Our contrastive CoT engine is structured to isolate reasoning discrepancies in the decision process (e.g., why an item was not selected given the current policy) while explicitly preserving priors to avoid conflating with affinity signals. To directly validate orthogonality, we have added a targeted ablation in the revised manuscript comparing skill evolution with and without memory accumulation. The results demonstrate that the diagnosed flaws and performance gains are reasoning-specific and do not reduce to re-encoding memory signals. revision: yes
Referee: [§3.1] §3.1 (Two-representation skill architecture): The claim that the architecture 'keeps evolution separate from inference without introducing new biases or instability' is load-bearing for the orthogonality result. No analysis is provided on whether the minimal inference-time injection leaks evolved principles back into the base model or creates distribution shift over long interaction sequences.

Authors: The two-representation design intentionally uses a minimal inference-time injection to decouple evolution from inference and reduce leakage risk. We acknowledge that explicit long-sequence analysis was not included in the original submission. In the revised manuscript, we have added experiments and discussion quantifying potential leakage of evolved principles and distribution shifts over extended interaction sequences, confirming that the architecture maintains separation without introducing measurable biases or instability. revision: yes

Circularity Check

0 steps flagged

No significant circularity in SAGER framework derivation

full rationale

The paper introduces SAGER as a novel framework with independent components including a two-representation skill architecture that decouples evolution from inference, an incremental contrastive chain-of-thought engine for diagnosing flaws via accepted vs. unchosen items, and skill-augmented listwise reasoning. These are defined as new constructs without any equations, fitted parameters, or self-referential reductions that would make the claimed orthogonality to memory accumulation or SOTA gains tautological by construction. Performance claims rest on experiments across four public benchmarks rather than definitional equivalence or self-citation chains. No load-bearing self-citations, uniqueness theorems, or ansatzes smuggled via prior work appear in the provided text, confirming the derivation is self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review limits visibility into parameters and assumptions; the framework implicitly relies on LLM capabilities for natural-language skill evolution and contrastive reasoning.

axioms (1)

domain assumption LLMs can reliably update and apply structured natural-language policy documents through interaction without catastrophic forgetting or hallucination of prior rules.
Invoked by the claim that skills evolve continuously while preserving accumulated priors.

invented entities (1)

policy skill no independent evidence
purpose: Structured natural-language document encoding personalized decision principles that evolves per user.
New postulated artifact introduced to decouple reasoning personalization from memory; no independent evidence provided in abstract.

pith-pipeline@v0.9.0 · 5542 in / 1187 out tokens · 25149 ms · 2026-05-10T09:51:42.053890+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 18 canonical work pages · 10 internal anchors

[1]

MemRec: Collaborative Memory-Augmented Agentic Recommender System

URLhttps://github.com/langchain-ai/langchain. Weixin Chen, Yuhan Zhao, Jingyuan Huang, Zihe Ye, Clark Mingxuan Ju, Tong Zhao, Neil Shah, Li Chen, and Yongfeng Zhang. Memrec: Collaborative memory-augmented agentic recommender system.arXiv preprint arXiv:2601.08816,

work page internal anchor Pith review Pith/arXiv arXiv
[2]

From Local to Global: A Graph RAG Approach to Query-Focused Summarization

Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization.arXiv preprint arXiv:2404.16130,

work page internal anchor Pith review arXiv
[3]

arXiv preprint arXiv:2303.14524 , year=

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

work page arXiv
[4]

Rethinking Recommendation Paradigms: From Pipelines to Agentic Recommender Systems

Jinxin Hu, Hao Deng, Lingyu Mu, Hao Zhang, Shizhun Wang, Yu Zhang, and Xiaoyi Zeng. Rethink- ing recommendation paradigms: From pipelines to agentic recommender systems.arXiv preprint arXiv:2603.26100,

work page internal anchor Pith review Pith/arXiv arXiv
[5]

ReRec: Reasoning-Augmented LLM-based Recommendation Assistant via Reinforcement Fine-tuning

Jiani Huang, Shijie Wang, Liangbo Ning, Wenqi Fan, and Qing Li. Rerec: Reasoning-augmented llm- based recommendation assistant via reinforcement fine-tuning.arXiv preprint arXiv:2604.07851,

work page internal anchor Pith review Pith/arXiv arXiv
[6]

Self-evolverec: Self-evolving recommender systems with llm-based directional feedback.arXiv preprint arXiv:2602.12612,

Sein Kim, Sangwu Park, Hongseok Kang, Wonjoong Kim, Jimin Seo, Yeonjun In, Kanghoon Yoon, and Chanyoung Park. Self-evolverec: Self-evolving recommender systems with llm-based directional feedback.arXiv preprint arXiv:2602.12612,

work page arXiv
[7]

Recnet: Self-evolving preference propagation for agentic recommender systems.arXiv preprint arXiv:2601.21609,

15 Bingqian Li, Xiaolei Wang, Junyi Li, Weitao Li, Long Zhang, Sheng Chen, Wayne Xin Zhao, and Ji-Rong Wen. Recnet: Self-evolving preference propagation for agentic recommender systems. arXiv preprint arXiv:2601.21609,

work page arXiv
[8]

Is chatgpt a good recom- mender? a preliminary study

Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang. Is chatgpt a good recom- mender? a preliminary study. InProceedings of the CIKM 2023 Workshop on Recommendation with Generative Models,

2023
[9]

Llm-rec: Personalized recommendation via prompting large language models

Hanjia Lyu, Song Jiang, Hanqing Zeng, Yinglong Xia, Qifan Wang, Si Zhang, Ren Chen, Chris Leung, Jiajie Tang, and Jiebo Luo. Llm-rec: Personalized recommendation via prompting large language models. InFindings of the Association for Computational Linguistics: NAACL 2024, pages 583–612,

2024
[10]

Justifying recommendations using distantly-labeled reviews and fine-grained aspects

Jianmo Ni, Jiacheng Li, and Julian McAuley. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 188–197,

2019
[11]

MemGPT: Towards LLMs as Operating Systems

URL https://openai.com/index/hello-gpt-4o/. Charles Packer, Vivian Fang, Shishir_G Patil, Kevin Lin, Sarah Wooders, and Joseph_E Gonzalez. Memgpt: Towards llms as operating systems.arXiv preprint arXiv:2310.08560,

work page internal anchor Pith review arXiv
[12]

Zep: A Temporal Knowledge Graph Architecture for Agent Memory

Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. Zep: a temporal knowledge graph architecture for agent memory.arXiv preprint arXiv:2501.13956,

work page internal anchor Pith review arXiv
[13]

Personax: A recommen- dation agent-oriented user modeling framework for long behavior sequence

Yunxiao Shi, Wujiang Xu, Zhang Zeqi, Xing Zi, Qiang Wu, and Min Xu. Personax: A recommen- dation agent-oriented user modeling framework for long behavior sequence. InFindings of the Association for Computational Linguistics: ACL 2025, pages 5764–5787,

2025
[14]

Recbot: Agent-based recommendation system.arXiv preprint arXiv:2509.21317,

Jiakai Tang, Yujie Luo, Xunke Xi, Fei Sun, Xueyang Feng, Sunhao Dai, Chao Yi, Dian Chen, Zhujin Gao, Yang Li, et al. Interactive recommendation agent with active user commands.arXiv preprint arXiv:2509.21317,

work page arXiv
[15]

Voyager: An Open-Ended Embodied Agent with Large Language Models

Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291,

work page internal anchor Pith review arXiv
[16]

Self-evolving recommen- dation system: End-to-end autonomous model optimization with llm agents.arXiv preprint arXiv:2602.10226,

Haochen Wang, Yi Wu, Daryl Chang, Li Wei, and Lukasz Heldt. Self-evolving recommen- dation system: End-to-end autonomous model optimization with llm agents.arXiv preprint arXiv:2602.10226,

work page arXiv
[17]

Self-Consistency Improves Chain of Thought Reasoning in Language Models

16 Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdh- ery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171,

work page Pith review arXiv
[18]

Recmind: Large language model powered agent for recommendation

Yancheng Wang, Ziyan Jiang, Zheng Chen, Fan Yang, Yingxue Zhou, Eunah Cho, Xing Fan, Yanbin Lu, Xiaojiang Huang, and Yingzhen Yang. Recmind: Large language model powered agent for recommendation. InFindings of the Association for Computational Linguistics: NAACL 2024, pages 4351–4364, 2024a. Zhefan Wang, Yuanqing Yu, Wendi Zheng, Weizhi Ma, and Min Zhang....

2024
[19]

Internalizing multi-agent reasoning for accurate and efficient llm-based recommendation.arXiv preprint arXiv:2602.09829,

Yang Wu, Haoze Wang, Qian Li, Jun Zhang, Huan Yu, and Jie Jiang. Internalizing multi-agent reasoning for accurate and efficient llm-based recommendation.arXiv preprint arXiv:2602.09829,

work page arXiv
[20]

SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning

Peng Xia, Jianwen Chen, Hanyang Wang, Jiaqi Liu, Kaide Zeng, Yu Wang, Siwei Han, Yiyang Zhou, Xujiang Zhao, Haifeng Chen, et al. Skillrl: Evolving agents via recursive skill-augmented reinforcement learning.arXiv preprint arXiv:2602.08234,

work page internal anchor Pith review arXiv
[21]

Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward

Renjun Xu and Yang Yan. Agent skills for large language models: Architecture, acquisition, security, and the path forward.arXiv preprint arXiv:2602.12430,

work page internal anchor Pith review arXiv
[22]

Memory-R1: Enhancing Large Language Model Agents to Manage and Utilize Memories via Reinforcement Learning

Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-mem: Agentic memory for llm agents. InAdvances in Neural Information Processing Systems, 2025a. Wujiang Xu, Yunxiao Shi, Zujie Liang, Xuying Ning, Kai Mei, Kun Wang, Xi Zhu, Min Xu, and Yongfeng Zhang. iAgent: LLM agent as a shield between user and recommender systems. In Findin...

work page internal anchor Pith review arXiv 2025
[23]

On generative agents in recommendation

An Zhang, Yuxin Chen, Leheng Sheng, Xiang Wang, and Tat-Seng Chua. On generative agents in recommendation. InProceedings of the 47th international ACM SIGIR conference on research and development in Information Retrieval, pages 1807–1817, 2024a. Junjie Zhang, Yupeng Hou, Ruobing Xie, Wenqi Sun, Julian McAuley, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. Ag...

work page arXiv 2024
[24]

reflection

17 A Experimental Setup and Implementation Details A.1 Dataset Details We utilize four datasets widely used in recommendation research, encompassing diverse domains such as e-commerce, social reading, entertainment, and local services. As mentioned in Section 3, we adopt the versions of these datasets augmented with natural language user instructions from...

2019