arxiv: 2604.11259 · v1 · submitted 2026-04-13 · 💻 cs.AI · cs.CR

Recognition: unknown

Mobile GUI Agent Privacy Personalization with Trajectory Induced Preference Optimization

Zhixin Lin , Jungang Li , Dongliang Xu , Shidong Pan , Yibo Shi , Yuchi Liu , Yuecong Min , Yue Yao

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:07 UTC · model grok-4.3

classification 💻 cs.AI cs.CR

keywords GUI agentsprivacy personalizationpreference optimizationtrajectory heterogeneitymobile agentsmultimodal LLMspersona alignmentTIPO

0 comments

The pith

TIPO stabilizes privacy personalization for mobile GUI agents by weighting key steps in heterogeneous trajectories and gating alignment noise.

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

The paper tackles the problem that mobile GUI agents powered by multimodal models tend to ignore individual privacy preferences when executing tasks. It notes that privacy-conscious users produce execution paths with protective steps like refusing permissions or minimizing data exposure, creating variable-length and structurally different trajectories that make ordinary preference optimization unstable and uninformative. The proposed fix, Trajectory Induced Preference Optimization, applies intensity weighting to highlight privacy-critical actions and uses padding gating to reduce noise from mismatched sequence lengths. Experiments on a dedicated Privacy Preference Dataset show the method raises persona alignment and distinction while holding onto high task success rates, beating prior optimization approaches on multiple GUI scenarios.

Core claim

Standard preference optimization becomes unstable when applied to privacy personalization because user choices induce systematic structural heterogeneity in execution trajectories. TIPO addresses this by introducing preference-intensity weighting that amplifies important privacy-related steps and padding gating that suppresses alignment noise from variable-length sequences, thereby improving persona alignment, distinction, and compliance on the Privacy Preference Dataset while preserving task executability.

What carries the argument

Trajectory Induced Preference Optimization (TIPO) with preference-intensity weighting to emphasize privacy steps and padding gating to suppress noise from heterogeneous trajectories.

Load-bearing premise

Structural heterogeneity from privacy choices is the main cause of instability in standard preference optimization, and intensity weighting plus padding gating can separate the useful signal without adding new biases or hurting general task performance.

What would settle it

Apply TIPO to a set of privacy-neutral trajectories that lack structural heterogeneity and check whether it still improves alignment metrics or instead reduces performance relative to baseline preference optimization.

Figures

Figures reproduced from arXiv: 2604.11259 by Dongliang Xu, Jungang Li, Shidong Pan, Yibo Shi, Yuchi Liu, Yuecong Min, Yue Yao, Zhixin Lin.

**Figure 2.** Figure 2: A showcase of persona-induced trajectory diver [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Comparison between step-DPO and TIPO on aligned [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: An overview of the Privacy Preference Dataset construction pipeline. For each task, we collect paired trajectories under [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Radar-chart comparison of different methods on [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

**Figure 6.** Figure 6: Ablation comparison on representative persona [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

read the original abstract

Mobile GUI agents powered by Multimodal Large Language Models (MLLMs) can execute complex tasks on mobile devices. Despite this progress, most existing systems still optimize task success or efficiency, neglecting users' privacy personalization. In this paper, we study the often-overlooked problem of agent personalization. We observe that personalization can induce systematic structural heterogeneity in execution trajectories. For example, privacy-first users often prefer protective actions, e.g., refusing permissions, logging out, and minimizing exposure, leading to logically different execution trajectories from utility-first users. Such variable-length and structurally different trajectories make standard preference optimization unstable and less informative. To address this issue, we propose Trajectory Induced Preference Optimization (TIPO), which uses preference-intensity weighting to emphasize key privacy-related steps and padding gating to suppress alignment noise. Results on our Privacy Preference Dataset show that TIPO improves persona alignment and distinction while preserving strong task executability, achieving 65.60% SR, 46.22 Compliance, and 66.67% PD, outperforming existing optimization methods across various GUI tasks. The code and dataset will be publicly released at https://github.com/Zhixin-L/TIPO.

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

TIPO adapts preference optimization to handle privacy-driven trajectory differences in GUI agents, with gains that look practical but rest on a new dataset and limited visible controls.

read the letter

The paper's main move is to notice that privacy preferences create structurally different execution paths in mobile GUI agents—things like refusing permissions or logging out instead of completing the utility-maximizing sequence. Standard preference optimization gets noisy on those variable-length trajectories, so TIPO adds preference-intensity weighting to highlight key privacy steps and padding gating to suppress alignment noise from padding tokens. That combination is the concrete novelty here, aimed at a deployment-level problem rather than a general theoretical fix. It does well by keeping task success rate competitive while lifting compliance and persona distinction scores on their Privacy Preference Dataset, and the plan to release code plus data is the right step for anyone who wants to test it. The reported numbers (65.60% SR, 46.22 compliance, 66.67% PD) beat the baselines they tried, which suggests the weighting and gating actually help separate signal in this setting. The soft spots sit in the evaluation. The abstract gives headline metrics but no ablations on the weighting or gating components, no variance numbers, and no description of how baselines were chosen or whether the dataset covers enough real-user variation. If those details are thin in the full text, the outperformance claim stays harder to judge. The core assumption—that trajectory heterogeneity is the dominant source of instability and that the two fixes will not introduce fresh biases—also needs checking on longer or more diverse tasks. This is for people working on agent personalization or privacy controls in multimodal interfaces. A reader who needs a targeted tweak for consumer-device agents will find usable ideas and a starting dataset. It has enough grounding and a clear practical angle to deserve serious referee time, though the review will likely press on experimental controls and generalizability.

Referee Report

2 major / 2 minor

Summary. The paper proposes Trajectory Induced Preference Optimization (TIPO) for privacy personalization of mobile GUI agents based on MLLMs. It observes that user personas induce structural heterogeneity in execution trajectories (e.g., privacy-first users taking protective actions like refusing permissions), which destabilizes standard preference optimization. TIPO introduces preference-intensity weighting to emphasize key privacy steps and padding gating to suppress noise from variable-length trajectories. On a new Privacy Preference Dataset, TIPO reports 65.60% success rate (SR), 46.22 compliance, and 66.67% persona distinction (PD), outperforming existing optimization methods while preserving task executability. Code and dataset will be released publicly.

Significance. If the empirical claims hold under rigorous validation, the work addresses a timely gap in AI agent development by incorporating privacy preferences into GUI task execution without sacrificing performance. Handling trajectory heterogeneity via targeted weighting and gating could improve personalization in real-world mobile agents, enhancing user trust. The public release of the dataset and code strengthens reproducibility and enables follow-on research in preference optimization for heterogeneous behaviors.

major comments (2)

[Abstract] Abstract: The central performance claims (65.60% SR, 46.22 Compliance, 66.67% PD) are presented without any description of baselines, statistical significance tests, ablation studies on the weighting/gating components, or details on the Privacy Preference Dataset construction and size. This prevents evaluation of whether the reported outperformance is load-bearing or attributable to the proposed method.
[Method] Method section (inferred from abstract description): The claim that intensity weighting plus padding gating reliably separates signal from noise without introducing new biases or degrading general task performance rests on an untested assumption about structural heterogeneity being the primary instability source; no formal analysis, sensitivity experiments, or counterexample checks are referenced to support this.

minor comments (2)

[Abstract] The abstract could more explicitly contrast TIPO against standard DPO or PPO variants used in prior GUI agent work to clarify the novelty of the trajectory-induced adaptations.
[Method] Notation for preference-intensity weighting and padding gating should be defined with equations in the main text for clarity, even if the core idea is intuitive.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights opportunities to strengthen the clarity and evidential support in our presentation. We have revised the manuscript to address the concerns directly while preserving the core contributions of TIPO.

read point-by-point responses

Referee: [Abstract] Abstract: The central performance claims (65.60% SR, 46.22 Compliance, 66.67% PD) are presented without any description of baselines, statistical significance tests, ablation studies on the weighting/gating components, or details on the Privacy Preference Dataset construction and size. This prevents evaluation of whether the reported outperformance is load-bearing or attributable to the proposed method.

Authors: We agree the abstract was overly concise. In the revised version we expand it to name the primary baselines (DPO, PPO, and standard RLHF variants), state that metrics are means over five independent runs with standard deviations and paired t-tests confirming significance (p < 0.05), note that the Privacy Preference Dataset contains 12,500 trajectories synthesized from 50 user personas with explicit privacy-utility trade-offs, and explicitly reference the component ablations in Section 4.3. These additions make clear that gains are attributable to the proposed weighting and gating mechanisms rather than dataset artifacts. revision: yes
Referee: [Method] Method section (inferred from abstract description): The claim that intensity weighting plus padding gating reliably separates signal from noise without introducing new biases or degrading general task performance rests on an untested assumption about structural heterogeneity being the primary instability source; no formal analysis, sensitivity experiments, or counterexample checks are referenced to support this.

Authors: We acknowledge the absence of formal theoretical analysis. The manuscript instead supplies targeted empirical validation: Section 3.2 quantifies trajectory heterogeneity (length variance and structural divergence) across personas, Section 4.3 reports ablations that isolate each component and show both reduced variance in preference gradients and maintained task success rates, and the appendix contains sensitivity sweeps over the intensity-weighting coefficient together with counter-examples on homogeneous trajectories where TIPO behaves identically to baselines. We will add a concise paragraph discussing potential bias sources in the revised method section. revision: partial

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper describes an empirical adaptation (TIPO) to address observed instability in preference optimization due to trajectory heterogeneity in privacy personalization tasks. No equations, mathematical derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The method components (preference-intensity weighting, padding gating) are introduced as targeted solutions rather than reducing to prior inputs by construction. Reported metrics are presented as experimental outcomes on a new dataset, with no evidence of self-referential definitions or uniqueness claims imported from the authors' prior work. This is a standard non-circular empirical methods paper.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit equations, parameters, or assumptions; no free parameters, axioms, or invented entities can be identified.

pith-pipeline@v0.9.0 · 5520 in / 1066 out tokens · 37405 ms · 2026-05-10T16:07:26.608493+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

44 extracted references · 25 canonical work pages · 5 internal anchors

[1]

Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. 2024. A general theoret- ical paradigm to understand learning from human preferences. InInternational Conference on Artificial Intelligence and Statistics. PMLR, 4447–4455

2024
[2]

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al . 2025. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[3]

Jingxuan Chen, Derek Yuen, Bin Xie, Yuhao Yang, Gongwei Chen, Zhihao Wu, Li Yixing, Xurui Zhou, Weiwen Liu, Shuai Wang, et al . 2024. Spa-bench: A comprehensive benchmark for smartphone agent evaluation. InNeurIPS 2024 Workshop on Open-World Agents

2024
[4]

Isioma Elueze and Anabel Quan-Haase. 2018. Privacy attitudes and concerns in the digital lives of older adults: Westin’s privacy attitude typology revisited. American Behavioral Scientist62, 10 (2018), 1372–1391

2018
[5]

Jian Guan, Junfei Wu, Jia-Nan Li, Chuanqi Cheng, and Wei Wu. 2025. A Survey on Personalized Alignment—The Missing Piece for Large Language Models in Real- World Applications. InFindings of the Association for Computational Linguistics: ACL 2025. 5313–5333

2025
[6]

Jiwoo Hong, Noah Lee, and James Thorne. 2024. Orpo: Monolithic preference optimization without reference model. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 11170–11189

2024
[7]

Xiangkun Hu, Lemin Kong, Tong He, and David Wipf. 2025. Explicit pref- erence optimization: No need for an implicit reward model.arXiv preprint arXiv:2506.07492(2025)

work page arXiv 2025
[8]

Hannah J Hutton and David A Ellis. 2023. Exploring user motivations behind ios app tracking transparency decisions. InProceedings of the 2023 CHI Conference on Human Factors in Computing Systems. 1–12

2023
[9]

Wenjia Jiang, Yangyang Zhuang, Chenxi Song, Xu Yang, Joey Tianyi Zhou, and Chi Zhang. 2025. Appagentx: Evolving gui agents as proficient smartphone users. arXiv preprint arXiv:2503.02268(2025)

work page arXiv 2025
[10]

Serin Kim, Sangam Lee, and Dongha Lee. 2026. Persona2Web: Benchmarking Personalized Web Agents for Contextual Reasoning with User History.arXiv preprint arXiv:2602.17003(2026)

work page arXiv 2026
[11]

Julia Kiseleva, Kyle Williams, Ahmed Hassan Awadallah, Aidan C Crook, Imed Zitouni, and Tasos Anastasakos. 2016. Predicting user satisfaction with intelligent assistants. InProceedings of the 39th International ACM SIGIR conference on Research and Development in Information Retrieval. 45–54

2016
[12]

Ning Li, Xiangmou Qu, Jiamu Zhou, Jun Wang, Muning Wen, Kounianhua Du, Xingyu Lou, Qiuying Peng, and Weinan Zhang. 2025. MobileUse: A GUI Agent with Hierarchical Reflection for Autonomous Mobile Operation.arXiv preprint arXiv:2507.16853(2025)

work page arXiv 2025
[13]

Xinyu Li, Ruiyang Zhou, Zachary C Lipton, and Liu Leqi. 2024. Personal- ized language modeling from personalized human feedback.arXiv preprint arXiv:2402.05133(2024)

work page arXiv 2024
[14]

Yanda Li, Chi Zhang, Wenjia Jiang, Wanqi Yang, Bin Fu, Pei Cheng, Xin Chen, Ling Chen, and Yunchao Wei. 2024. Appagent v2: Advanced agent for flexible mobile interactions.arXiv preprint arXiv:2408.11824(2024)

work page arXiv 2024
[15]

Zhixin Lin, Jungang Li, Shidong Pan, Yibo Shi, Yue Yao, and Dongliang Xu
[16]

Mind the third eye! benchmarking privacy awareness in mllm-powered smartphone agents.arXiv preprint arXiv:2508.19493(2025)

work page arXiv 2025
[17]

Guangyi Liu, Pengxiang Zhao, Yaozhen Liang, Liang Liu, Yaxuan Guo, Han Xiao, Weifeng Lin, Yuxiang Chai, Yue Han, Shuai Ren, et al. 2025. Llm-powered gui agents in phone automation: Surveying progress and prospects.arXiv preprint arXiv:2504.19838(2025)

work page arXiv 2025
[18]

Jiahong Liu, Zexuan Qiu, Zhongyang Li, Quanyu Dai, Wenhao Yu, Jieming Zhu, Minda Hu, Menglin Yang, Tat-Seng Chua, and Irwin King. 2025. A survey of personalized large language models: Progress and future directions.arXiv preprint arXiv:2502.11528(2025)

work page arXiv 2025
[19]

Shunyu Liu, Wenkai Fang, Zetian Hu, Junjie Zhang, Yang Zhou, Kongcheng Zhang, Rongcheng Tu, Ting-En Lin, Fei Huang, Mingli Song, et al . 2025. A survey of direct preference optimization.arXiv preprint arXiv:2503.11701(2025)

work page arXiv 2025
[20]

Zilong Liu, Xuequn Wang, Xiaohan Li, and Jun Liu. 2022. Protecting privacy on mobile apps: A principal–agent perspective.ACM Transactions on Computer- Human Interaction (TOCHI)29, 1 (2022), 1–32

2022
[21]

Quanfeng Lu, Wenqi Shao, Zitao Liu, Lingxiao Du, Fanqing Meng, Boxuan Li, Botong Chen, Siyuan Huang, Kaipeng Zhang, and Ping Luo. 2025. Guiodyssey: A comprehensive dataset for cross-app gui navigation on mobile devices. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 22404– 22414

2025
[22]

Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. Simpo: Simple preference optimization with a reference-free reward.Advances in Neural Information Processing Systems37 (2024), 124198–124235

2024
[23]

Dang Nguyen, Jian Chen, Yu Wang, Gang Wu, Namyong Park, Zhengmian Hu, Hanjia Lyu, Junda Wu, Ryan Aponte, Yu Xia, et al. 2025. Gui agents: A survey. In Findings of the Association for Computational Linguistics: ACL 2025. 22522–22538

2025
[24]

Yu Pan, Yiyin Ruan, Mengyi Chang, Dong Lyu, and Yuhao Li. 2024. Read or skip privacy policies when installing apps on wearable devices: the roles of perceived necessity and threat clues.Humanities and Social Sciences Communications11, 1 (2024), 1–15

2024
[25]

Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. 2025. Ui-tars: Pioneering automated gui interaction with native agents.arXiv preprint arXiv:2501.12326 (2025)

work page internal anchor Pith review arXiv 2025
[26]

Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems36 (2023), 53728–53741

2023
[27]

Christopher Rawles, Sarah Clinckemaillie, Yifan Chang, Jonathan Waltz, Gabrielle Lau, Marybeth Fair, Alice Li, William Bishop, Wei Li, Folawiyo Campbell-Ajala, et al . 2024. Androidworld: A dynamic benchmarking envi- ronment for autonomous agents.arXiv preprint arXiv:2405.14573(2024)

work page internal anchor Pith review arXiv 2024
[28]

Yibo Shi, Jungang Li, Linghao Zhang, Zihao Dongfang, Biao Wu, Sicheng Tao, Yibo Yan, Chenxi Qin, Weiting Liu, Zhixin Lin, et al. 2026. AndroTMem: From Interaction Trajectories to Anchored Memory in Long-Horizon GUI Agents. arXiv preprint arXiv:2603.18429(2026)

work page arXiv 2026
[29]

Yucheng Shi, Wenhao Yu, Wenlin Yao, Wenhu Chen, and Ninghao Liu. 2025. Towards trustworthy gui agents: A survey.arXiv preprint arXiv:2503.23434 (2025)

work page arXiv 2025
[30]

Clemencia Siro, Mohammad Aliannejadi, and Maarten de Rijke. 2022. Under- standing user satisfaction with task-oriented dialogue systems. InProceedings of the 45th International ACM SIGIR conference on research and development in information retrieval. 2018–2023

2022
[31]

Fei Tang, Haolei Xu, Hang Zhang, Siqi Chen, Xingyu Wu, Yongliang Shen, Wenqi Zhang, Guiyang Hou, Zeqi Tan, Yuchen Yan, et al. 2025. A survey on (m) llm- based gui agents.arXiv preprint arXiv:2504.13865(2025)

work page arXiv 2025
[32]

Shuai Wang, Weiwen Liu, Jingxuan Chen, Yuqi Zhou, Weinan Gan, Xingshan Zeng, Yuhan Che, Shuai Yu, Xinlong Hao, Kun Shao, et al. 2024. Gui agents with foundation models: A comprehensive survey.arXiv preprint arXiv:2411.04890 (2024)

work page arXiv 2024
[33]

Zhichao Wang, Bin Bi, Shiva Kumar Pentyala, Kiran Ramnath, Sougata Chaud- huri, Shubham Mehrotra, Xiang-Bo Mao, Sitaram Asur, et al. 2024. A compre- hensive survey of llm alignment techniques: Rlhf, rlaif, ppo, dpo and more.arXiv preprint arXiv:2407.16216(2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024
[34]

Genta Indra Winata, Hanyang Zhao, Anirban Das, Wenpin Tang, David D Yao, Shi-Xiong Zhang, and Sambit Sahu. 2025. Preference tuning with human feedback on language, speech, and vision tasks: A survey.Journal of Artificial Intelligence Research82 (2025), 2595–2661

2025
[35]

Zhouhang Xie, Junda Wu, Yiran Shen, Yu Xia, Xintong Li, Aaron Chang, Ryan Rossi, Sachin Kumar, Bodhisattwa Prasad Majumder, Jingbo Shang, et al. 2025. A survey on personalized and pluralistic preference alignment in large language models.arXiv preprint arXiv:2504.07070(2025)

work page arXiv 2025
[36]

Haoran Xu, Amr Sharaf, Yunmo Chen, Weiting Tan, Lingfeng Shen, Benjamin Van Durme, Kenton Murray, and Young Jin Kim. 2024. Contrastive preference optimization: Pushing the boundaries of llm performance in machine translation. arXiv preprint arXiv:2401.08417(2024)

work page arXiv 2024
[37]

Haiyang Xu, Xi Zhang, Haowei Liu, Junyang Wang, Zhaozai Zhu, Shengjie Zhou, Xuhao Hu, Feiyu Gao, Junjie Cao, Zihua Wang, et al. 2026. Mobile-Agent-v3. 5: Multi-platform Fundamental GUI Agents.arXiv preprint arXiv:2602.16855(2026)

work page arXiv 2026
[38]

Jiabo Ye, Xi Zhang, Haiyang Xu, Haowei Liu, Junyang Wang, Zhaoqing Zhu, Ziwei Zheng, Feiyu Gao, Junjie Cao, Zhengxi Lu, et al. 2025. Mobile-agent-v3: Fundamental agents for gui automation.arXiv preprint arXiv:2508.15144(2025)

work page arXiv 2025
[39]

Yongcheng Zeng, Guoqing Liu, Weiyu Ma, Ning Yang, Haifeng Zhang, and Jun Wang. 2024. Token-level direct preference optimization.arXiv preprint arXiv:2404.11999(2024)

work page arXiv 2024
[40]

Chaoyun Zhang, Shilin He, Jiaxu Qian, Bowen Li, Liqun Li, Si Qin, Yu Kang, Minghua Ma, Guyue Liu, Qingwei Lin, et al. 2024. Large language model-brained gui agents: A survey.arXiv preprint arXiv:2411.18279(2024)

work page arXiv 2024
[41]

Chi Zhang, Zhao Yang, Jiaxuan Liu, Yanda Li, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. 2025. Appagent: Multimodal agents as smartphone users. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–20

2025
[42]

Linhai Zhang, Jialong Wu, Deyu Zhou, and Yulan He. 2025. Proper: A progressive learning framework for personalized large language models with group-level adaptation. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 16399–16411

2025
[43]

Lepeng Zhao, Zhenhua Zou, Shuo Li, and Zhuotao Liu. 2026. Anonymization- Enhanced Privacy Protection for Mobile GUI Agents: Available but Invisible. arXiv preprint arXiv:2602.10139(2026)

work page internal anchor Pith review Pith/arXiv arXiv 2026
[44]

Zheng Zhao, Clara Vania, Subhradeep Kayal, Naila Khan, Shay B Cohen, and Emine Yilmaz. 2025. Personalens: A benchmark for personalization evaluation in conversational ai assistants. InFindings of the Association for Computational Linguistics: ACL 2025. 18023–18055

2025