Selected work

Representative Papers

UI-CUBE: Enterprise-Grade Computer Use Agent Benchmarking Beyond Task Accuracy to Operational Reliability

Nov 21, 2025

Existing CUA (Computer-User-Agent) benchmarks overemphasize functional correctness and fail to assess agent reliability in enterprise production environments. Method: We propose UI-CUBE—a first-of-its-kind, enterprise-ready, systematic diagnostic benchmark comprising 226 graded tasks. It integrates UI perturbation, multi-resolution UI testing, and application-state verification to rigorously evaluate architectural deficiencies in memory management, hierarchical planning, and state coordination. Contribution/Results: Experiments reveal that state-of-the-art agents achieve only 67–85% success on simple tasks but plummet to 9–19% on complex enterprise workflows—far below human novices (61.2%), exposing fundamental architectural bottlenecks. UI-CUBE establishes a reproducible, attributable reliability evaluation paradigm for industrial deployment of CUAs.

0 citationsRead paper

From Characters to Tokens: Dynamic Grouping with Hierarchical BPE

Oct 17, 2025

Existing subword tokenization methods (e.g., BPE) suffer from inefficient rare-word modeling and excessively large vocabularies, while character-level models—though robust—introduce computational bottlenecks in Transformer architectures. To address this, we propose a **language-agnostic dynamic character grouping method**, the first to directly leverage BPE segmentation structure for two-stage hierarchical compression: (1) dynamically aggregating isomorphic character blocks based on BPE boundaries, with explicit block-end tokens inserted; and (2) applying lightweight secondary BPE compression over the resulting block sequence. Our approach requires no whitespace assumptions or auxiliary models, preserving character-level generalization while achieving subword-level efficiency. Experiments across multilingual benchmarks show performance competitive with or superior to entropy- and whitespace-driven baselines, alongside substantial reductions in vocabulary size (−38% on average) and sequence length (−29% on average), thereby lowering computational overhead.

0 citationsRead paper

When Embedding Models Meet: Procrustes Bounds and Applications

Oct 15, 2025

Embedding models trained independently on similar data capture stable semantic meanings but yield inconsistent representation spaces, hindering interoperability across models. This work addresses compatibility challenges in multimodal search and model upgrades via orthogonal transformation-based embedding alignment. Theoretically, we derive the first tight Procrustes alignment error bound, proving the existence of a near-isometric orthogonal transformation that approximately preserves pairwise inner products—establishing rigorous theoretical foundations for alignment. Methodologically, we employ efficient Procrustes analysis as a post-hoc alignment procedure, preserving the intrinsic geometric structure of each embedding space while enabling cross-model alignment. Experiments demonstrate substantial improvements in model retraining compatibility, text retrieval fusion accuracy, and cross-modal search performance; our method achieves state-of-the-art results in hybrid multimodal search.

0 citationsRead paper

wd1: Weighted Policy Optimization for Reasoning in Diffusion Language Models

Jul 07, 2025

Diffusion-based large language models (dLLMs) suffer from high computational overhead and accumulated bias in reinforcement learning (RL)-driven inference optimization, primarily due to repeated approximations of the current, old, and reference policy likelihoods—introducing importance sampling denominator errors and multi-stage approximation bias. Method: We propose Weighted Policy Optimization (WPO), which reformulates the RL objective as a weighted likelihood maximization problem. WPO requires only a single approximation of the current policy likelihood, eliminating dependence on importance sampling and avoiding supervised fine-tuning. It integrates denoising diffusion, zeroth-order optimization, and the R1-Zero training paradigm for end-to-end inference optimization. Contribution/Results: On major reasoning benchmarks, WPO achieves up to a 16% absolute accuracy gain. It significantly reduces the number of gradient-step function evaluations, while improving both training efficiency and stability.

0 citationsRead paper

Toward Autonomous UI Exploration: The UIExplorer Benchmark

Jun 21, 2025

This work addresses the lack of standardized evaluation for autonomous agents’ UI exploration capabilities by introducing UIExplore-Bench—the first dedicated benchmark for this task. It features a three-tier GitLab sandbox environment supporting dual-modality assessment: structured (DOM-based) and purely visual (GUI screenshot-based). A novel metric, hUFO, is proposed to quantify the discovery of interactive UI components. Evaluation integrates DOM parsing, human-like interaction simulation, and functional coverage analysis to systematically compare agent performance against human experts. Experiments show that UIExplore-AlGo achieves 77.2% and 59.0% of human-normalized performance within 2000 steps under structured and visual modalities, respectively—significantly outperforming baseline methods, especially under sparse feedback conditions. All code, datasets, and environments are publicly released, establishing a reproducible infrastructure for UI exploration research.

0 citationsRead paper

Recent publications

Latest Papers

UI-CUBE: Enterprise-Grade Computer Use Agent Benchmarking Beyond Task Accuracy to Operational Reliability

Nov 21, 2025

Existing CUA (Computer-User-Agent) benchmarks overemphasize functional correctness and fail to assess agent reliability in enterprise production environments. Method: We propose UI-CUBE—a first-of-its-kind, enterprise-ready, systematic diagnostic benchmark comprising 226 graded tasks. It integrates UI perturbation, multi-resolution UI testing, and application-state verification to rigorously evaluate architectural deficiencies in memory management, hierarchical planning, and state coordination. Contribution/Results: Experiments reveal that state-of-the-art agents achieve only 67–85% success on simple tasks but plummet to 9–19% on complex enterprise workflows—far below human novices (61.2%), exposing fundamental architectural bottlenecks. UI-CUBE establishes a reproducible, attributable reliability evaluation paradigm for industrial deployment of CUAs.

0 citationsRead paper

From Characters to Tokens: Dynamic Grouping with Hierarchical BPE

Oct 17, 2025

Existing subword tokenization methods (e.g., BPE) suffer from inefficient rare-word modeling and excessively large vocabularies, while character-level models—though robust—introduce computational bottlenecks in Transformer architectures. To address this, we propose a **language-agnostic dynamic character grouping method**, the first to directly leverage BPE segmentation structure for two-stage hierarchical compression: (1) dynamically aggregating isomorphic character blocks based on BPE boundaries, with explicit block-end tokens inserted; and (2) applying lightweight secondary BPE compression over the resulting block sequence. Our approach requires no whitespace assumptions or auxiliary models, preserving character-level generalization while achieving subword-level efficiency. Experiments across multilingual benchmarks show performance competitive with or superior to entropy- and whitespace-driven baselines, alongside substantial reductions in vocabulary size (−38% on average) and sequence length (−29% on average), thereby lowering computational overhead.

0 citationsRead paper

When Embedding Models Meet: Procrustes Bounds and Applications

Oct 15, 2025

Embedding models trained independently on similar data capture stable semantic meanings but yield inconsistent representation spaces, hindering interoperability across models. This work addresses compatibility challenges in multimodal search and model upgrades via orthogonal transformation-based embedding alignment. Theoretically, we derive the first tight Procrustes alignment error bound, proving the existence of a near-isometric orthogonal transformation that approximately preserves pairwise inner products—establishing rigorous theoretical foundations for alignment. Methodologically, we employ efficient Procrustes analysis as a post-hoc alignment procedure, preserving the intrinsic geometric structure of each embedding space while enabling cross-model alignment. Experiments demonstrate substantial improvements in model retraining compatibility, text retrieval fusion accuracy, and cross-modal search performance; our method achieves state-of-the-art results in hybrid multimodal search.

0 citationsRead paper

wd1: Weighted Policy Optimization for Reasoning in Diffusion Language Models

Jul 07, 2025

Diffusion-based large language models (dLLMs) suffer from high computational overhead and accumulated bias in reinforcement learning (RL)-driven inference optimization, primarily due to repeated approximations of the current, old, and reference policy likelihoods—introducing importance sampling denominator errors and multi-stage approximation bias. Method: We propose Weighted Policy Optimization (WPO), which reformulates the RL objective as a weighted likelihood maximization problem. WPO requires only a single approximation of the current policy likelihood, eliminating dependence on importance sampling and avoiding supervised fine-tuning. It integrates denoising diffusion, zeroth-order optimization, and the R1-Zero training paradigm for end-to-end inference optimization. Contribution/Results: On major reasoning benchmarks, WPO achieves up to a 16% absolute accuracy gain. It significantly reduces the number of gradient-step function evaluations, while improving both training efficiency and stability.

0 citationsRead paper

Toward Autonomous UI Exploration: The UIExplorer Benchmark

Jun 21, 2025

This work addresses the lack of standardized evaluation for autonomous agents’ UI exploration capabilities by introducing UIExplore-Bench—the first dedicated benchmark for this task. It features a three-tier GitLab sandbox environment supporting dual-modality assessment: structured (DOM-based) and purely visual (GUI screenshot-based). A novel metric, hUFO, is proposed to quantify the discovery of interactive UI components. Evaluation integrates DOM parsing, human-like interaction simulation, and functional coverage analysis to systematically compare agent performance against human experts. Experiments show that UIExplore-AlGo achieves 77.2% and 59.0% of human-normalized performance within 2000 steps under structured and visual modalities, respectively—significantly outperforming baseline methods, especially under sparse feedback conditions. All code, datasets, and environments are publicly released, establishing a reproducible infrastructure for UI exploration research.

0 citationsRead paper