Selected work

Representative Papers

Concatenated Matrix SVD: Compression Bounds, Incremental Approximation, and Error-Constrained Clustering

Jan 12, 2026arXiv.org

This work addresses the issue of uncontrolled reconstruction errors in SVD-based compression of large collections of matrices when heuristic grouping is employed prior to concatenation. To overcome this limitation, the authors propose a theory-driven compressive clustering framework grounded in spectral analysis of horizontally concatenated matrices. They establish, for the first time, a globally provable upper bound on SVD reconstruction error and derive two novel spectral bounds based on a lower bound for singular value growth. Building upon these theoretical guarantees, they design three clustering algorithms with explicit error control, integrated with incremental approximate SVD to efficiently estimate compression error without explicitly forming the full concatenated matrix. The resulting approach achieves a favorable balance among speed, accuracy, and scalability, significantly enhancing the reliability and practicality of SVD compression in applications such as multi-view learning, signal processing, and neural network compression.

1 citationsRead paper

MacArena: Benchmarking Computer Use Agents on an Online macOS Environment

Jun 04, 2026

Current agent evaluation benchmarks lack coverage of the macOS platform, particularly Apple Silicon architecture and native GUI tasks. This work proposes MacArena—the first large-scale online macOS evaluation environment built upon Apple’s native Virtualization framework—which integrates ported OSWorld tasks with newly designed native tasks spanning 50 applications and 421 human-validated scenarios. Experimental results reveal that state-of-the-art models suffer an average performance drop exceeding 26% on native tasks compared to ported ones, and their relative rankings invert between task types. These findings highlight fundamental limitations in cross-platform GUI understanding and demonstrate the effectiveness of macOS as a high-difficulty benchmark environment for evaluating embodied AI agents.

0 citationsRead paper

GUIrilla: A Scalable Framework for Automated Desktop UI Exploration

Oct 16, 2025

High-quality interaction data for GUI automation on macOS is scarce, and representative UI datasets are lacking. Method: We propose GUIrilla—the first large-scale, extensible UI exploration framework tailored to the macOS ecosystem. Built upon native Accessibility APIs, it constructs hierarchical GUI graph structures, incorporates a dedicated interaction processor, and introduces a semantic action trajectory generation mechanism to enable deep, cross-application interface exploration. Contribution/Results: We release GUIrilla-Task, an open-source dataset comprising 1,108 macOS applications and 27,000 function-annotated tasks, along with a corresponding evaluation benchmark. Experiments show that LLM agents fine-tuned on GUIrilla-Task achieve superior performance on ScreenSpot Pro—outperforming synthetic-data baselines using only 3% of the data volume—demonstrating substantial gains in UI intelligence and agent capability.

0 citationsRead paper

Screen2AX: Vision-Based Approach for Automatic macOS Accessibility Generation

Jul 22, 2025

A significant accessibility metadata gap exists across macOS applications, with only 33% providing full accessibility support—severely hindering AI agent interaction and accessibility for visually impaired users. To address this, we propose the first method for real-time hierarchical accessibility tree (AX Tree) generation from a single application screenshot, bypassing reliance on native system APIs. Our approach synergistically integrates vision-language models with UI object detection to jointly perform element identification, semantic description, and hierarchical structure reconstruction. We introduce Screen2AX-Task, the first benchmark for macOS desktop task accessibility evaluation, and release a large-scale, open-source dataset comprising 112 macOS applications. Our method achieves 77% F1 score on AX tree reconstruction—2.2× higher than the native macOS accessibility API—and outperforms OmniParser V2 on the ScreenSpot benchmark.

0 citationsRead paper

SwiftEval: Developing a Language-Specific Benchmark for LLM-generated Code Evaluation

May 30, 2025

Existing LLM code-generation benchmarks are heavily skewed toward Python, with scarce high-quality evaluation suites for Swift; HumanEval-XL and MultiPL-E suffer from translation inaccuracies and inadequate coverage of Swift-specific features (e.g., Optionals, Protocols), limiting their ability to assess models’ mastery of Swift’s core semantics. Method: We introduce SwiftEval—the first high-fidelity, human-authored benchmark tailored for Swift—comprising 28 original programming tasks, each accompanied by manually crafted, functionally rigorous test cases emphasizing language-specific challenges. Contribution/Results: Evaluating 44 state-of-the-art code LLMs reveals pronounced performance degradation on Swift-specific tasks—especially among smaller models—validating SwiftEval’s diagnostic sensitivity and benchmark rigor. SwiftEval fills a critical gap in Swift code competency assessment and establishes a new paradigm for multi-language LLM evaluation grounded in deep language semantics.

0 citationsRead paper

Recent publications

Latest Papers

MacArena: Benchmarking Computer Use Agents on an Online macOS Environment

Jun 04, 2026

Current agent evaluation benchmarks lack coverage of the macOS platform, particularly Apple Silicon architecture and native GUI tasks. This work proposes MacArena—the first large-scale online macOS evaluation environment built upon Apple’s native Virtualization framework—which integrates ported OSWorld tasks with newly designed native tasks spanning 50 applications and 421 human-validated scenarios. Experimental results reveal that state-of-the-art models suffer an average performance drop exceeding 26% on native tasks compared to ported ones, and their relative rankings invert between task types. These findings highlight fundamental limitations in cross-platform GUI understanding and demonstrate the effectiveness of macOS as a high-difficulty benchmark environment for evaluating embodied AI agents.

0 citationsRead paper

Concatenated Matrix SVD: Compression Bounds, Incremental Approximation, and Error-Constrained Clustering

Jan 12, 2026arXiv.org

This work addresses the issue of uncontrolled reconstruction errors in SVD-based compression of large collections of matrices when heuristic grouping is employed prior to concatenation. To overcome this limitation, the authors propose a theory-driven compressive clustering framework grounded in spectral analysis of horizontally concatenated matrices. They establish, for the first time, a globally provable upper bound on SVD reconstruction error and derive two novel spectral bounds based on a lower bound for singular value growth. Building upon these theoretical guarantees, they design three clustering algorithms with explicit error control, integrated with incremental approximate SVD to efficiently estimate compression error without explicitly forming the full concatenated matrix. The resulting approach achieves a favorable balance among speed, accuracy, and scalability, significantly enhancing the reliability and practicality of SVD compression in applications such as multi-view learning, signal processing, and neural network compression.

1 citationsRead paper

GUIrilla: A Scalable Framework for Automated Desktop UI Exploration

Oct 16, 2025

High-quality interaction data for GUI automation on macOS is scarce, and representative UI datasets are lacking. Method: We propose GUIrilla—the first large-scale, extensible UI exploration framework tailored to the macOS ecosystem. Built upon native Accessibility APIs, it constructs hierarchical GUI graph structures, incorporates a dedicated interaction processor, and introduces a semantic action trajectory generation mechanism to enable deep, cross-application interface exploration. Contribution/Results: We release GUIrilla-Task, an open-source dataset comprising 1,108 macOS applications and 27,000 function-annotated tasks, along with a corresponding evaluation benchmark. Experiments show that LLM agents fine-tuned on GUIrilla-Task achieve superior performance on ScreenSpot Pro—outperforming synthetic-data baselines using only 3% of the data volume—demonstrating substantial gains in UI intelligence and agent capability.

0 citationsRead paper

Screen2AX: Vision-Based Approach for Automatic macOS Accessibility Generation

Jul 22, 2025

A significant accessibility metadata gap exists across macOS applications, with only 33% providing full accessibility support—severely hindering AI agent interaction and accessibility for visually impaired users. To address this, we propose the first method for real-time hierarchical accessibility tree (AX Tree) generation from a single application screenshot, bypassing reliance on native system APIs. Our approach synergistically integrates vision-language models with UI object detection to jointly perform element identification, semantic description, and hierarchical structure reconstruction. We introduce Screen2AX-Task, the first benchmark for macOS desktop task accessibility evaluation, and release a large-scale, open-source dataset comprising 112 macOS applications. Our method achieves 77% F1 score on AX tree reconstruction—2.2× higher than the native macOS accessibility API—and outperforms OmniParser V2 on the ScreenSpot benchmark.

0 citationsRead paper

SwiftEval: Developing a Language-Specific Benchmark for LLM-generated Code Evaluation

May 30, 2025

Existing LLM code-generation benchmarks are heavily skewed toward Python, with scarce high-quality evaluation suites for Swift; HumanEval-XL and MultiPL-E suffer from translation inaccuracies and inadequate coverage of Swift-specific features (e.g., Optionals, Protocols), limiting their ability to assess models’ mastery of Swift’s core semantics. Method: We introduce SwiftEval—the first high-fidelity, human-authored benchmark tailored for Swift—comprising 28 original programming tasks, each accompanied by manually crafted, functionally rigorous test cases emphasizing language-specific challenges. Contribution/Results: Evaluating 44 state-of-the-art code LLMs reveals pronounced performance degradation on Swift-specific tasks—especially among smaller models—validating SwiftEval’s diagnostic sensitivity and benchmark rigor. SwiftEval fills a critical gap in Swift code competency assessment and establishes a new paradigm for multi-language LLM evaluation grounded in deep language semantics.

0 citationsRead paper