Selected work

Representative Papers

Hereditary pattern-free classes are not always 2-wqo

Jul 24, 2026

This study addresses the open question of whether hereditary graph classes that are pattern-free necessarily possess the 2–well-quasi-ordering (2-wqo) property. By integrating techniques from graph-theoretic pattern theory and well-quasi-ordering theory, we construct the first example of a hereditary graph class that is pattern-free yet fails to be 2-wqo. This counterexample refutes the longstanding conjecture that pattern-freeness implies 2-wqo, thereby demonstrating that the two properties are not inherently linked. Our result resolves a key misconception in the field and offers a new perspective on the relationship between structural constraints in graph classes and their well-quasi-ordering behavior.

0 citationsRead paper

Multi-Source and Cross-Scenario Strategy-Guided Code Optimization

Jul 22, 2026

Existing strategy-guided code optimization approaches struggle to integrate multi-source knowledge and lack cross-language generalization capabilities. This work proposes MoST, a novel framework that, for the first time, enables unified representation of diverse knowledge sources—such as textbooks and web resources—and facilitates transfer of optimization strategies across programming languages. MoST employs self-balancing weighted clustering to identify universal optimization patterns and leverages cross-scenario example transfer to generate high-quality static analysis rules that guide large language models in code optimization. Experimental results demonstrate that MoST significantly outperforms SemOpt on 351 tasks spanning C/C++, Python, and Rust, achieving average performance improvements of 4.44%–258.17% across 15 real-world projects, with a peak gain of 717.42%.

0 citationsRead paper

Benchmarking Resource-Efficient LLMs for Research Topic Ontology Generation in the Biomedical Field

Jul 20, 2026

This study addresses the inefficiency of manual biomedical ontology construction and the unclear potential of small-scale large language models (LLMs) in identifying complex semantic relationships. It presents the first systematic evaluation of five open-source LLMs with parameter counts ≤9 billion on this task, introducing MeSH-Rel-4K—a novel dataset comprising 4,000 expert-annotated semantic relations from Medical Subject Headings (MeSH). The work compares three adaptation strategies: standard prompting, chain-of-thought prompting, and supervised fine-tuning. Experimental results demonstrate that supervised fine-tuning substantially outperforms prompt-based methods, yielding an average F1 score improvement of 34.1 percentage points. These findings confirm that targeted fine-tuning effectively unlocks the capability of small LLMs for automated, domain-specific ontology construction.

0 citationsRead paper

FormIDEAble: Safe and Socially-aware Autonomous Systems

Jun 30, 2026

This work addresses the challenge of simultaneously modeling human social behaviors and providing formal safety guarantees in safety-critical human-robot collaboration scenarios. The authors propose a unified framework that formulates collaborative decision-making as a Priced Timed Markov Decision Process (Priced Timed MDP) and integrates bounded reachability analysis to synthesize policies that are both socially aware and provably safe. By uniquely combining formal verification, social behavior modeling, and cost-constrained decision-making, the approach enables a controllable trade-off between safety assurance and task performance. Its effectiveness is demonstrated in an emergency evacuation scenario, offering a reliable foundation for decision-making in high-risk human-robot systems.

0 citationsRead paper

From Failure to Alignment: A Requirements Engineering Framework for Machine Learning Systems

Jun 30, 2026

This work addresses the challenge of ensuring trustworthiness and stakeholder alignment in machine learning system development, which is often hindered by the absence of systematic requirements engineering. To bridge this gap, the authors propose REAL, a novel framework that uniquely integrates failure mode analysis into the requirements engineering process. REAL establishes a tripartite principle centered on data, model, and holistic system requirements, enabling iterative and traceable requirement refinement. Through a model-driven, stakeholder-oriented design, REAL demonstrates substantial improvements in requirement satisfaction in an autonomous driving case study. The authors further support reproducibility by releasing an open-source implementation toolkit.

0 citationsRead paper

Recent publications

Latest Papers

Hereditary pattern-free classes are not always 2-wqo

Jul 24, 2026

This study addresses the open question of whether hereditary graph classes that are pattern-free necessarily possess the 2–well-quasi-ordering (2-wqo) property. By integrating techniques from graph-theoretic pattern theory and well-quasi-ordering theory, we construct the first example of a hereditary graph class that is pattern-free yet fails to be 2-wqo. This counterexample refutes the longstanding conjecture that pattern-freeness implies 2-wqo, thereby demonstrating that the two properties are not inherently linked. Our result resolves a key misconception in the field and offers a new perspective on the relationship between structural constraints in graph classes and their well-quasi-ordering behavior.

0 citationsRead paper

Multi-Source and Cross-Scenario Strategy-Guided Code Optimization

Jul 22, 2026

Existing strategy-guided code optimization approaches struggle to integrate multi-source knowledge and lack cross-language generalization capabilities. This work proposes MoST, a novel framework that, for the first time, enables unified representation of diverse knowledge sources—such as textbooks and web resources—and facilitates transfer of optimization strategies across programming languages. MoST employs self-balancing weighted clustering to identify universal optimization patterns and leverages cross-scenario example transfer to generate high-quality static analysis rules that guide large language models in code optimization. Experimental results demonstrate that MoST significantly outperforms SemOpt on 351 tasks spanning C/C++, Python, and Rust, achieving average performance improvements of 4.44%–258.17% across 15 real-world projects, with a peak gain of 717.42%.

0 citationsRead paper

Benchmarking Resource-Efficient LLMs for Research Topic Ontology Generation in the Biomedical Field

Jul 20, 2026

This study addresses the inefficiency of manual biomedical ontology construction and the unclear potential of small-scale large language models (LLMs) in identifying complex semantic relationships. It presents the first systematic evaluation of five open-source LLMs with parameter counts ≤9 billion on this task, introducing MeSH-Rel-4K—a novel dataset comprising 4,000 expert-annotated semantic relations from Medical Subject Headings (MeSH). The work compares three adaptation strategies: standard prompting, chain-of-thought prompting, and supervised fine-tuning. Experimental results demonstrate that supervised fine-tuning substantially outperforms prompt-based methods, yielding an average F1 score improvement of 34.1 percentage points. These findings confirm that targeted fine-tuning effectively unlocks the capability of small LLMs for automated, domain-specific ontology construction.

0 citationsRead paper

FormIDEAble: Safe and Socially-aware Autonomous Systems

Jun 30, 2026

This work addresses the challenge of simultaneously modeling human social behaviors and providing formal safety guarantees in safety-critical human-robot collaboration scenarios. The authors propose a unified framework that formulates collaborative decision-making as a Priced Timed Markov Decision Process (Priced Timed MDP) and integrates bounded reachability analysis to synthesize policies that are both socially aware and provably safe. By uniquely combining formal verification, social behavior modeling, and cost-constrained decision-making, the approach enables a controllable trade-off between safety assurance and task performance. Its effectiveness is demonstrated in an emergency evacuation scenario, offering a reliable foundation for decision-making in high-risk human-robot systems.

0 citationsRead paper

From Failure to Alignment: A Requirements Engineering Framework for Machine Learning Systems

Jun 30, 2026

This work addresses the challenge of ensuring trustworthiness and stakeholder alignment in machine learning system development, which is often hindered by the absence of systematic requirements engineering. To bridge this gap, the authors propose REAL, a novel framework that uniquely integrates failure mode analysis into the requirements engineering process. REAL establishes a tripartite principle centered on data, model, and holistic system requirements, enabling iterative and traceable requirement refinement. Through a model-driven, stakeholder-oriented design, REAL demonstrates substantial improvements in requirement satisfaction in an autonomous driving case study. The authors further support reproducibility by releasing an open-source implementation toolkit.

0 citationsRead paper