Mobile GUI Agent Privacy Personalization with Trajectory Induced Preference Optimization Zhixin Lin1 , Jungang Li2,3 , Dongliang Xu1† , Shidong Pan4 Yibo Shi5 , Yuchi Liu6 , Yuecong Min7 , Yue Yao1† 1 Shandong University
2 The Hong Kong University of Science and Technology (Guangzhou)
3 The Hong Kong University of Science and Technology 6 Australian National University
4 New York University
5 Xi’an Jiaotong University
7 Institute of Computing Technology, Chinese Academy of Sciences
arXiv:2604.11259v1 [cs.AI] 13 Apr 2026
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 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, minimizing exposure), leading to logically different execution trajectories from utility-first users. This variable-length and structurally different execution trajectory makes standard preference optimization unstable and less informative. To address this, 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.
Keywords Mobile GUI agent, privacy personalization, preference optimization
1
Introduction
Unlike traditional voice assistants, which only answer questions in natural language, mobile GUI agents can directly operate apps to complete real user instructions, such as searching for information across apps, sending messages, booking tickets, navigating maps, adjusting system settings, managing emails, and completing shopping or service workflows [16, 27, 39]. In recent years, Multimodal Large Language Models (MLLMs) have enabled mobile GUI agents to perform complex tasks on mobile devices [12, 30]. The advancement of mobile GUI agents further improved task completion rate on these realistic smartphone tasks [9, 26, 36]. They are moving beyond proof-of-concept demonstrations toward practical assistants that can act on behalf of users in daily routines. However, from an end-user’s perspective, task completion alone does not necessarily imply user satisfaction [11, 29]. Users care not only about whether a task is completed, but also about how it is completed and what risks are incurred along the way, e.g., privacy † Corresponding authors.
Goal Observation History Privacy Preference
Privacy-first Minimal permission granting No personal data autofill Single-platform purchase
Utility-first Personalized Agent Preference-conditioned branching
Privacy preference
Expand permission granting
Personal data autofill Cross-platform comparison
Figure 1: Illustration of personalized trajectory selection for smartphone GUI agents. Given the same task goal, observation history, and user portrait, different privacy personas can induce different preference-conditioned branches, leading to distinct execution trajectories.
exposure. For the same task, there are often multiple feasible execution trajectories, and different users may prefer different trade-offs between utility and privacy risk. Shown in Fig. 1, privacy-sensitive users tend to behave more conservatively during task execution, for example, by reading privacy policies carefully, granting only necessary permissions, disabling personalized tracking, and logging out of accounts to reduce unnecessary exposure. [8, 19, 23] In contrast, utility-oriented users are more likely to accept default settings and follow more direct, higher-utility paths, even at the cost of greater privacy exposure, in order to reduce interaction friction and accomplish tasks more efficiently. Most existing research and systems optimize Success Rate or interaction efficiency, by assuming that each task has a single optimal trajectory, and ignoring trajectory induced preferences by user preference (e.g., the privacy preferences) [22, 40]. This gap significantly limits the real-world user experience of Mobile GUI agents. In this paper, we thereby focus on a practical yet often overlooked direction: personalized operation trajectory selection based on user privacy preferences in the Mobile GUI agent context. We define the personalized trajectory selection as follows: for multiple trajectories that can complete the same task, the agent should select one that best aligns with the preferences implied by the user preference. Specifically, we focus on privacy preference as a representative and
high-stakes user preference that critically affects the trustworthiness of Mobile GUI agents. In this work, privacy is operationalized as a user’s preference regarding how much personal information is disclosed, retained, tracked, or exposed during task execution. This preference is reflected in concrete behavioral choices, such as whether to grant optional permissions, accept personalized tracking, remain logged in, or clear traces after task completion. Inspired by Westin’s categorization of privacy attitudes [4] and further extending it, privacy preference allows us to clearly identify which steps are necessary and which steps should be avoided, making it an ideal entry point for evaluating “process personalization.” To support this setting, we build a new Privacy Preference dataset with paired trajectories under different privacy personas. An intuitive solution for personalized trajectory selection is preference-aligned training on an MLLM-based Mobile GUI agent, where a policy learns from paired feedback between chosen (more preference-consistent) and rejected (less consistent) trajectories. However, we find that existing preference optimization methods, such as Direct Preference Optimization (DPO) [25], are poorly matched to this problem. Trajectories from different privacy preferences are often structurally heterogeneous and length-mismatched. For example, for the same task goal, one trajectory may directly complete the target action, while another may additionally adjust privacy-related settings, deny unnecessary permissions, or clear traces after completion. As a result, the training signal is diluted, and gradients can be dominated by padding. The model may learn how to match the padded format rather than learning the true privacy-related differences. Moreover, privacy-relevant actions are often sparse but critical within a long trajectory; standard DPO is not sufficiently sensitive to these key steps, leading to unstable optimization and limited gains. To address the intrinsic heterogeneity, we propose Trajectory Induced Preference Optimization (TIPO), a preference optimization method for structurally heterogeneous trajectories. TIPO improves learning in two ways: it uses preference-intensity weighting to emphasize persona-relevant steps, and padding gating to suppress noise from alignment placeholders. Together, these designs make preference optimization more suitable for variable-length trajectory pairs. Results show that TIPO achieves the best overall performance among compared methods, reaching 65.60% in SR, 42.85 in PAS-S, 46.22 in Compliance, and 66.6% in PD, while preserving strong task executability. These results indicate that TIPO not only maintains task success but also more effectively aligns the generated trajectories with the target privacy persona and strengthens persona distinction across diverse mobile tasks. Further ablations show that preference-intensity weighting improves learning on privacy-critical steps, raising Compliance from 31.94 (DPO) to 38.93, while the full model further increases it to 46.22; similarly, PD improves from 59.26% to 62.96% and finally to 66.67%, demonstrating that padding gating and preference-intensity weighting are complementary and jointly yield the most stable and consistent gains. Further ablations indicate that preference-intensity weighting strengthens learning on critical privacy steps, while padding gating effectively suppresses unstable updates caused by alignment noise; their combination yields the most stable and consistent improvements. These results demonstrate that user preference can induce
structural trajectory differences in Mobile GUI agents, and that preference optimization mechanisms designed for such differences can effectively address the problem. In summary, our main contributions are: (1) We define mobile GUI agent personalization as a trajectory selection task, where the goal is not only to complete the task but also to choose a persona-consistent trajectory under the same task objective. (2) We build the privacy preference dataset, providing multitrajectory annotations for the same task goal under different privacy preferences, characterized by variable length and structural heterogeneity. (3) We propose the Trajectory Induced Preference Optimization (TIPO), which stabilizes preference optimization for variablelength, structurally heterogeneous trajectory feedback through preference-intensity weighting and padding gating, leading to significant improvements in experiments.
2 Related Work 2.1 Preference Optimization for Alignment Recent alignment research has increasingly shifted from Reinforcement Learning from Human Feedback (RLHF) toward simpler offline preference optimization objectives [18, 32, 33]. Direct Preference Optimization (DPO) [25] is a representative example, showing that preference alignment can be achieved with a direct classification-style objective without explicit reward modeling or reinforcement learning. Subsequent work has explored several variants of this paradigm, including reference-free or simplified objectives such as ORPO [6] and SimPO [21], contrastive formulations such as CPO [35], and broader theoretical perspectives such as IPO [1]. More recent studies further revisit reference mismatch, token-level weighting [38], and explicit preference objectives [7], suggesting that preference optimization remains an active and evolving direction. However, most of these methods are developed for settings in which the compared outputs are relatively homogeneous, such as two responses to the same prompt. Our setting is different. In Mobile GUI agents, user preferences can induce variable-length and structurally heterogeneous trajectories. A preferred trajectory may differ not only in local action choice, but also in whether certain steps are inserted, skipped, or reorganized. This makes standard response-level preference optimization less suitable, since the compared units are no longer naturally aligned. Our TIPO is therefore designed for trajectory-level preference learning under structural heterogeneity.
2.2
Personalization and User Modeling
Personalized alignment aims to move beyond population-level behavior and adapt models to users’ preferences, histories, and decision styles [5, 17, 34]. Existing work studies this problem through personalized preference learning, progressive adaptation, and benchmark construction. Representative examples include P-RLHF [13], which introduces personalized preference learning, PROPER [41], which formulates personalization as progressive refinement, and recent benchmarks such as PersonaLens [43] and Persona2Web [10],
Mobile GUI Agent Privacy Personalization with Trajectory Induced Preference Optimization
Utility-first Object Task: Open a video link in Chrome
…
Log in Privacy-first
Incognito
…
Shared steps Padding Step
Figure 2: A showcase of persona-induced trajectory divergence. Under the same task goal, Utility-first and Privacy-first share the same initial steps but diverge at a privacy-sensitive decision point, resulting different executable trajectories. which make personalized behavior increasingly measurable in conversational and agent settings. However, most existing personalization frameworks are developed for dialogue systems or web agents, where personalization is mainly reflected in response content or high-level decisions. In Mobile GUI agents, user preferences directly affect execution behavior, including permission handling, account states, privacy exposure, and risk-related action choices. As a result, personalization reshapes not only what the agent does, but also how the entire action trajectory is organized. This makes Mobile GUI personalization a problem of trajectory-level structural variation, highlighting the need for a trajectory-centric framework such as ours.
2.3
Mobile GUI Agents and Mobile Privacy
Recent progress in multimodal large language models has led to the rapid development of Mobile GUI agents and mobile interaction benchmarks. [22, 28, 31] Early systems such as AppAgent [40] and AppAgent-v2 [14] demonstrate the feasibility of autonomous mobile app operation, while benchmarks such as AndroidWorld [26], GUIOdyssey [20], and SPA-Bench [3] make evaluation more realistic and systematic. More recent agent systems, including UITARS [24] and Mobile-Agent-v3 [37], further push this direction toward stronger grounding, longer-horizon execution, and more practical deployment. Recent studies have also started to examine privacy in mobile agent settings, shifting attention from general task execution to privacy-related risks and protections. However, existing work still mainly focuses on task success, privacy awareness [15], or information protection [42]. We study how user-specific privacy preferences reshape the execution trajectory itself under the same task goal. This makes privacy preference a trajectory-selection problem rather than only a detection or protection problem, motivating our trajectory-centric method.
3
Problem Definition
We define Mobile GUI agents personalization as a trajectory selection task. Specifically, given the same task and initial UI state,
different user preferences may induce different preferred execution strategies, leading to trajectories with systematically different structures and lengths. In this work, we focus on a high-stakes preference privacy preference, and aim to generate trajectories that are more consistent with the target privacy preference while preserving task feasibility. Figure 2 showcase a visualized example. We define the agent input as 𝑥 = (𝑔, 𝑜, ℎ, 𝑝), where 𝑔 is the task goal, 𝑜 is the current UI observation, ℎ denotes a bounded interaction history, and 𝑝 denotes the privacy persona. To enable a controlled study, we instantiate 𝑝 in a simplified binary form, with two representative profiles: Privacy-first and Utility-first. Given 𝑥, the agent generates a step sequence 𝑦 = (𝑦1, . . . , 𝑦 |𝑦 | ), where each 𝑦𝑡 denotes a specific action (e.g., “tap search” or “open the JD app”). Since privacy preference may introduce additional defensive actions or suppress utility-oriented ones, trajectory lengths are often different across persona branches. We construct persona-conditioned preference training samples as triplets (𝑥, 𝑦 +, 𝑦 − ), where 𝑦 + denotes the trajectory that is more aligned with the privacy persona specified in 𝑥, and 𝑦 − denotes a less aligned alternative under the same task context. This follows the standard preference-pair formulation used in DPO. Let 𝜋𝜃 (𝑦 | 𝑥) denote the agent policy parameterized by 𝜃 . The objective is to encourage the policy to assign higher probability to the personaaligned trajectory than to the less aligned one under the same context, i.e., 𝜋𝜃 (𝑦 + | 𝑥) > 𝜋𝜃 (𝑦 − | 𝑥). Importantly, this preference relation is defined over privacy alignment rather than task success, as both trajectories are assumed to be feasible solutions to the task.
4 Method 4.1 Preliminaries Direct Preference Optimization (DPO). We adopt DPO as the base preference optimization framework. Given a persona-conditioned preference pair (𝑥, 𝑦 +, 𝑦 − ), where 𝑦 + and 𝑦 − denote the preferred and less preferred trajectories under the same context 𝑥, DPO encourages the policy model 𝜋𝜃 to assign a higher relative likelihood to 𝑦 + than to 𝑦 − with respect to a fixed reference policy 𝜋ref . In standard DPO, this preference is defined at the trajectory level. Specifically, the sequence-level preference score can be written as h 𝑧 (𝜃 ) = 𝛽 log 𝜋𝜃 (𝑦 + | 𝑥) − log 𝜋𝜃 (𝑦 − | 𝑥) i − log 𝜋ref (𝑦 + | 𝑥) − log 𝜋 ref (𝑦 − | 𝑥) . where 𝛽 > 0 controls the sharpness of the preference signal. The DPO objective is then written as 𝐿DPO = E (𝑥,𝑦 + ,𝑦 − ) softplus(−𝑧(𝜃 )) . Limitations of DPO and motivation of our methods. DPO implicitly treats aligned positions in a preference pair as equally valid comparison units. This assumption becomes problematic in our setting, where persona-induced trajectories may differ substantially in both structure and length. After alignment, some positions correspond to genuine persona-related decisions, while others are introduced only for padding, leading to semantic noise and making uniform supervision less suitable for trajectory preference learning.
Preferred trajectory s1
s2
s3
s4
Less preferred trajectory
s5
s1
s2
s3
s4
s5
s6
s1
s2
s3
s4
s5
s6
Padding s1
s2
s3
s4
s5
S6*
𝑳step-DPO = 𝔼 𝕩,𝕪+ ,𝕪− softplus −𝑧𝑡 𝜃 𝑳TIPO = 𝔼 𝕩,𝕪+,𝕪− softplus −𝒎𝒕 𝜶𝒕 𝑧𝑡 𝜃 Padding Gating
,
𝜶𝒕 = 𝐜𝐥𝐢𝐩
− 𝐒𝐜𝐨𝐫𝐞 𝒔+ 𝒕 − 𝐒𝐜𝐨𝐫𝐞 𝒔𝒕 , 𝟎, 𝟏 𝚫𝒎𝒂𝒙
𝜸
Preference Intensity Weighting
Figure 3: Comparison between step-DPO and TIPO on aligned trajectory pairs. While step-DPO treats positions uniformly, including alignment-induced placeholders, TIPO highlights persona-critical steps via preference-intensity weighting and reduces placeholder noise through padding gating.
As a result, standard DPO faces two limitations in our setting: • Padding alignment introduces semantic placeholder noise. To compare variable-length trajectories, the chosen and rejected branches must be aligned to a common length. However, the resulting no_action placeholders do not carry genuine preference information, yet they still enter the loss calculation. • Uniform treatment of aligned positions ignores step importance. In standard DPO, all aligned positions are treated as equally informative training units. Consequently, personacritical steps may be diluted by a large number of neutral or placeholder positions. As a result, standard DPO becomes less effective on variablelength, structurally heterogeneous trajectory pairs. This limitation motivates TIPO, which improves preference learning by emphasizing persona-critical steps and suppressing alignment-induced noise.
4.2
TIPO
Since our method operates on aligned trajectory steps, we further decompose this sequence-level comparison into step-wise preference signals. Let 𝑥𝑡 denote the step-level planning context at aligned step 𝑡, including the current task state and the relevant interaction history up to that step. We define h 𝑧𝑡 (𝜃 ) = 𝛽 log 𝜋𝜃 (𝑦𝑡+ | 𝑥𝑡 ) − log 𝜋𝜃 (𝑦𝑡− | 𝑥𝑡 ) i − log 𝜋ref (𝑦𝑡+ | 𝑥𝑡 ) − log 𝜋 ref (𝑦𝑡− | 𝑥𝑡 ) . This step-wise form is a decomposition of the original trajectorylevel DPO objective tailored to our aligned-action setting, rather than the standard DPO formulation itself. 𝐿step-DPO = − log 𝜎 (𝑧𝑡 (𝜃 )) . As shown in Fig. 3, TIPO improves step-level preference optimization from two complementary aspects: it emphasizes persona-critical aligned positions through preference intensity weighting, and suppresses noise through a padding gating mechanism.
Preference intensity weighting. Standard DPO treats all aligned positions uniformly during optimization, which may dilute personacritical decisions in long trajectories. In our setting, however, different aligned positions contribute unequally to privacy preference expression: some steps (e.g., denying permissions or disabling tracking) are highly informative for persona discrimination, while many neutral steps mainly serve task execution and carry little preference information. To address this issue, we assign each aligned step pair a preference intensity weight 𝛼𝑡 , so that persona-relevant positions contribute more strongly to optimization. Specifically, for each aligned step pair (𝑠𝑡+, 𝑠𝑡− ), we first compute a persona-aware score difference Δ𝑠𝑡 = Score(𝑠𝑡+ ) − Score(𝑠𝑡− ), where Score(·) is derived from a rule-based action scoring scheme with LLM assistance. Concretely, our annotation protocol specifies preference-related action categories and their corresponding scores, while the LLM is used only to assist semantic normalization and resolve cases where surface forms differ but the underlying action intent is equivalent. The detailed scoring rules are provided in the appendix. We then map Δ𝑠𝑡 to a normalized step weight 𝛾 Δ𝑠𝑡 , 0, 1 , 𝛼𝑡 = clip Δmax where Δmax is the maximum score difference used for normalization, and 𝛾 ≥ 0 controls the sharpness of the mapping. A larger 𝛾 assigns relatively higher weights to steps with stronger persona relevance. Padding gating mechanism. Although preferenceintensity weighting highlights persona-relevant steps, aligned trajectory pairs still contain no_action placeholders introduced by variable-length alignment. These positions do not carry genuine preference information and may interfere with optimization if treated the same as valid semantic steps. To suppress such alignmentinduced noise, we introduce a padding gate 𝑚𝑡 : ( 0, if the chosen is no_action, 𝑚𝑡 = 1, otherwise. The gated weighted preference score is then defined as (gate)
𝑧ˆ𝑡
(𝜃 ) = 𝑚𝑡 𝛼𝑡 𝑧𝑡 (𝜃 ).
Finally, the TIPO objective is h i (gate) 𝐿TIPO = E (𝑥𝑡 ,𝑦𝑡+ ,𝑦𝑡− ) softplus −ˆ𝑧𝑡 (𝜃 ) .
5 Experiments and Results 5.1 Dataset Construction To study privacy preference driven personalized trajectory selection, we build the Privacy Preference dataset using the pipeline shown in Fig. 4, which combines real persona role-playing with on-device human trajectory collection. Inspired by Westin’s categorization of privacy attitudes [4], we instantiate privacy preference in a simplified binary setting with two personas: Privacy-first and Utility-first. This design captures the utility-privacy trade-off central to our problem, while remaining extensible to finer-grained privacy personas in future work.
Mobile GUI Agent Privacy Personalization with Trajectory Induced Preference Optimization
Sample 151 Tasks from 8 Categories
Personal Role-Playing Manually Task Review
Stage 0. Task Setup
Privacy-first Utility-first
- Collect trajectories by ADB-based logging. - Step are recorded.
40 hours
8 labelers
Stage 1. Trajectory Collection
Manually match length of 𝒚+ = 𝒚− Cross-check each trajectory for privacy, utility, and persona consistency. 10 hours
Stage 2. Human Verification
Original pair:
Aligned pair:
𝑦+ : s1→s2→s3→s4 𝑦+ : s1→s2→s3→s4 𝑦− : s1→s2→s4 𝑦− : s1→s2→ no_action→s4 Divergence
Insert no_action after key divergence between 𝒚+ and 𝒚−
Stage 3. Pref. Pair Construction and Alignment
Figure 4: An overview of the Privacy Preference Dataset construction pipeline. For each task, we collect paired trajectories under both Privacy-first and Utility-first personas, followed by human verification and pairwise alignment to construct trajectory-level preference pairs. The dataset covers eight high-frequency task categories (Shopping, Payment, Browsing, Food Delivery, Sharing, Account, Backup, and Reservations). In total, it contains 151 task instances. Each task is annotated with paired trajectories under Privacy-first and Utility-first personas, resulting in 302 trajectories and approximately 8.2k annotated steps. Each task provides a clear naturallanguage goal instruction (e.g., “Open a YouTube link . . . in the Edge browser”). For representation, we treat a step as the minimal interaction unit and represent each execution as a trajectory of steps. Each step records (1) an executable structured action represented as action_type (arguments), for example tap (x=129, y=138). The complete action space and parameter specifications are provided in the Appendix. (2) the visual observation of the current screen; (3) the corresponding UI state in XML format, and when necessary, a reasoning text explaining preference-driven decisions (e.g., a Privacy-first persona enables incognito mode when opening a browser); and (4) a semantic description of the action (e.g., “tap the search bar to enter a query”). This multi-view logging of “action parameters + screenshot + UI structure + semantic annotation” makes the dataset suitable for supervised learning and also facilitates manually auditing the authenticity and executability of trajectories. To ensure data quality and reproducibility, the dataset was annotated by eight annotators, each contributing about 40 hours, for a total of approximately 320 annotation hours. Before formal annotation, all annotators completed a warm-up stage, including a shared pilot annotation and two additional days of trial annotation, to familiarize themselves with the annotation interface, persona-specific rules, and trajectory recording requirements. We used Android Debug Bridge (ADB) to capture on-device screenshots together with executable action traces, so that each step is grounded in verifiable UI evidence. Because the collection process may involve privacy-sensitive scenarios, annotators were strictly prohibited from using any real personal information. All collected data were inspected before inclusion, and only samples verified to contain no private information were retained. For quality assurance, each trajectory was cross-checked by a second annotator, who verified both persona-rule consistency and semantic consistency between the recorded actions and their corresponding step descriptions. When disagreements arose, they were resolved through discussion under the unified annotation guidelines before the trajectory was finalized. In addition, for every task, the paired Privacy-first and Utility-first trajectories were required to contain at least one key persona-differentiating action.
Preference pair construction and alignment. For the same task instance, we manually role-play two user preferences, Privacyfirst and Utility-first, to execute the task and obtain two executable semantic trajectories under the same goal constraint. For a given preference 𝑝, we define the trajectory consistent with the persona as the preferred trajectory 𝑦 + , and the other branch as the dispreferred trajectory 𝑦 − . Therefore, the preference pair (𝑥, 𝑦 +, 𝑦 − ) arises naturally during annotation. Importantly, in mobile tasks, preference differences often induce structural trajectory heterogeneity. For example, Privacy-first trajectories may include additional defensive actions, such as reducing exposure or logging out, which makes the preferred and dispreferred branches systematically different in both trajectory structure and length. As a result, for a preference pair (𝑥, 𝑦 +, 𝑦 − ), the two trajectories are often not directly comparable at the sequence level, i.e., |𝑦 + | ≠ |𝑦 − |. Since DPO-style training compares preferred and dispreferred trajectories under a unified sequence dimension, we first align each pair to a common length 𝑇 = max(|𝑦 + |, |𝑦 − |). To do so, we introduce a semantic placeholder action no_action to fill missing positions in the shorter branch. This placeholder is defined at the trajectory level rather than as a tokenizer-level padding token, and indicates that no corresponding semantic step exists at that position. We use an LLM to assist divergence-point identification between the two trajectories, followed by manual verification, and insert no_action into the shorter trajectory until both branches have equal length. This yields aligned trajectory pairs that remain semantically interpretable and can be used for subsequent preference optimization.
5.2
Experimental Settings
To systematically evaluate the effectiveness of the proposed Trajectory Induced Preference Optimization (TIPO) for personalized trajectory selection in Mobile GUI agents, we conduct experiments from three perspectives: task executability, persona-level preference consistency, and persona distinction ability. Under a unified experimental protocol, we compare TIPO with several representative preference optimization baselines. Task setting. We consider the following task setting: given the same task goal and the same initial UI state, the agent is required to generate an execution trajectory that is consistent with the target user’s privacy persona. We adopt the standard Planner–Executor paradigm. Since our focus is personalized selection at the trajectory level, both training and evaluation are centered on the semantic
Table 1: Comparison of TIPO against baseline methods on the Privacy Preference Dataset. Results are reported in terms of task success rate (SR), persona adherence (PAS-S/PAS-U), compliance, and persona distinction (PD) under Privacy-first (P-f) and Utility-first (U-f) conditions. Otherwise indicated, the higher is better for values in this table.
Methods Frozen SFT DPO [25] ORPO [6] CPO [35] SimPO [21] IPO [1] Ours
SR U-f 34.50% 60.13% 60.37% 61.31% 61.36% 60.62% 60.84% 62.11%
P-f 44.08% 70.21% 70.31% 69.65% 68.63% 67.60% 70.04% 69.08%
Overall 39.29% 65.17% 65.34% 65.48% 65.00% 64.11% 65.44% 65.60%