arXiv subjects

Zhongmin Cai

Publications and source records attributed to Zhongmin Cai.

7 recordsLinked to original sources

arXiv2026

Alignment Is Local: A Paired Diagnostic for GUI Agents under User-Side Persuasion

Trustworthy deployment of GUI agents in ubiquitous computing settings requires alignment that survives dynamic interaction and precise threat conditions, not just single-turn refusal of explicit harmful requests. We argue that prompt-level alignment, the dominant lightweight defense in current mobile agents, is a local phenomenon: it works reliably only in the narrow evaluation slice where it is typically measured, namely single-turn, explicitly-verbalized intent, and degrades systematically along two axes that any real user can traverse. Using a paired diagnostic on three frontier GUI agents, screen-grounded, user-side persuasion, with no environment injection, we show that a one-line guardrail achieves large single-shot ASR reductions, up to roughly 40 points, at near-zero over-refusal cost. Nevertheless, moving from independent probes to four-turn escalation chains raises guarded ASR by approximately 20 points on every model. Relative to the neutral baselines, this increase reflects substantial guardrail erosion for Qwen but a largely defense-orthogonal dynamic risk for Claude and GPT. The sign of the salience gap flips under the guardrail: concealed requests are not systematically more successful than explicit ones without a guardrail, but are more successful with one, indicating that the defense engages primarily when intent is named. Static single-turn ASR therefore overstates deployed robustness by a systematic and predictable margin.

cs.CR↗

arXiv2026

DroidRetriever: A Transparent and Steerable Automation System for Collaborative Mobile Information Seeking

Information seeking on mobile devices is often fragmented, trapping users in repetitive cycles of context switching and data re-entry, which increases cognitive load and disrupts workflow. Existing mobile agents provide limited cross-source integration and are largely opaque, presenting progress as a linear feed with few opportunities to intervene, steer, or take control. We present DroidRetriever, a transparent, steerable system for cross-source mobile information seeking. It accepts voice or typed input and the multi-LLM system decomposes the task, navigates to target pages, takes screenshots, and synthesizes a concise report with citation-linked screenshots. We make the process transparent through a progress dashboard combining sub-task progress and real-time exploration maps for seamless takeover. DroidRetriever also pauses on detected privacy or high-risk screens and prompts intervention. Across 35 tasks over 24 apps, experiments and user studies demonstrate improvements in coverage, transparency, and reduced workload. We release our code at https://github.com/AkimotoAyako/DroidRetriever.

cs.HC↗

arXiv2025

Predicting User Behavior in Smart Spaces with LLM-Enhanced Logs and Personalized Prompts

Enhancing the intelligence of smart systems, such as smart home, and smart vehicle, and smart grids, critically depends on developing sophisticated planning capabilities that can anticipate the next desired function based on historical interactions. While existing methods view user behaviors as sequential data and apply models like RNNs and Transformers to predict future actions, they often fail to incorporate domain knowledge and capture personalized user preferences. In this paper, we propose a novel approach that incorporates LLM-enhanced logs and personalized prompts. Our approach first constructs a graph that captures individual behavior preferences derived from their interaction histories. This graph effectively transforms into a soft continuous prompt that precedes the sequence of user behaviors. Then our approach leverages the vast general knowledge and robust reasoning capabilities of a pretrained LLM to enrich the oversimplified and incomplete log records. By enhancing these logs semantically, our approach better understands the user's actions and intentions, especially for those rare events in the dataset. We evaluate the method across four real-world datasets from both smart vehicle and smart home settings. The findings validate the effectiveness of our LLM-enhanced description and personalized prompt, shedding light on potential ways to advance the intelligence of smart space.

cs.HC↗

arXiv2024

QGEval: Benchmarking Multi-dimensional Evaluation for Question Generation

Automatically generated questions often suffer from problems such as unclear expression or factual inaccuracies, requiring a reliable and comprehensive evaluation of their quality. Human evaluation is widely used in the field of question generation (QG) and serves as the gold standard for automatic metrics. However, there is a lack of unified human evaluation criteria, which hampers consistent and reliable evaluations of both QG models and automatic metrics. To address this, we propose QGEval, a multi-dimensional Evaluation benchmark for Question Generation, which evaluates both generated questions and existing automatic metrics across 7 dimensions: fluency, clarity, conciseness, relevance, consistency, answerability, and answer consistency. We demonstrate the appropriateness of these dimensions by examining their correlations and distinctions. Through consistent evaluations of QG models and automatic metrics with QGEval, we find that 1) most QG models perform unsatisfactorily in terms of answerability and answer consistency, and 2) existing metrics fail to align well with human judgments when evaluating generated questions across the 7 dimensions. We expect this work to foster the development of both QG technologies and their evaluation.

cs.CL↗

arXiv2024

In-Situ Mode: Generative AI-Driven Characters Transforming Art Engagement Through Anthropomorphic Narratives

Art appreciation serves as a crucial medium for emotional communication and sociocultural dialogue. In the digital era, fostering deep user engagement on online art appreciation platforms remains a challenge. Leveraging generative AI technologies, we present EyeSee, a system designed to engage users through anthropomorphic characters. We implemented and evaluated three modes (Narrator, Artist, and In-Situ) acting as a third-person narrator, a first-person creator, and first-person created objects, respectively, across two sessions: Narrative and Recommendation. We conducted a within-subject study with 24 participants. In the Narrative session, we found that the In-Situ and Artist modes had higher aesthetic appeal than the Narrator mode, although the Artist mode showed lower perceived usability. Additionally, from the Narrative to Recommendation session, we found that user-perceived relatability and believability within each interaction mode were sustained, but the user-perceived consistency and stereotypicality changed. Our findings suggest novel implications for applying anthropomorphic in-situ narratives to other educational settings.

cs.HC↗

arXiv2024

VisionTasker: Mobile Task Automation Using Vision Based UI Understanding and LLM Task Planning

Mobile task automation is an emerging field that leverages AI to streamline and optimize the execution of routine tasks on mobile devices, thereby enhancing efficiency and productivity. Traditional methods, such as Programming By Demonstration (PBD), are limited due to their dependence on predefined tasks and susceptibility to app updates. Recent advancements have utilized the view hierarchy to collect UI information and employed Large Language Models (LLM) to enhance task automation. However, view hierarchies have accessibility issues and face potential problems like missing object descriptions or misaligned structures. This paper introduces VisionTasker, a two-stage framework combining vision-based UI understanding and LLM task planning, for mobile task automation in a step-by-step manner. VisionTasker firstly converts a UI screenshot into natural language interpretations using a vision-based UI understanding approach, eliminating the need for view hierarchies. Secondly, it adopts a step-by-step task planning method, presenting one interface at a time to the LLM. The LLM then identifies relevant elements within the interface and determines the next action, enhancing accuracy and practicality. Extensive experiments show that VisionTasker outperforms previous methods, providing effective UI representations across four datasets. Additionally, in automating 147 real-world tasks on an Android smartphone, VisionTasker demonstrates advantages over humans in tasks where humans show unfamiliarity and shows significant improvements when integrated with the PBD mechanism. VisionTasker is open-source and available at https://github.com/AkimotoAyako/VisionTasker.

cs.HC↗

arXiv2015

Re-scale AdaBoost for Attack Detection in Collaborative Filtering Recommender Systems

Collaborative filtering recommender systems (CFRSs) are the key components of successful e-commerce systems. Actually, CFRSs are highly vulnerable to attacks since its openness. However, since attack size is far smaller than that of genuine users, conventional supervised learning based detection methods could be too "dull" to handle such imbalanced classification. In this paper, we improve detection performance from following two aspects. First, we extract well-designed features from user profiles based on the statistical properties of the diverse attack models, making hard classification task becomes easier to perform. Then, refer to the general idea of re-scale Boosting (RBoosting) and AdaBoost, we apply a variant of AdaBoost, called the re-scale AdaBoost (RAdaBoost) as our detection method based on extracted features. RAdaBoost is comparable to the optimal Boosting-type algorithm and can effectively improve the performance in some hard scenarios. Finally, a series of experiments on the MovieLens-100K data set are conducted to demonstrate the outperformance of RAdaBoost comparing with some classical techniques such as SVM, kNN and AdaBoost.

cs.IR↗