Selected work

Representative Papers

Associative Constructive Evolution: Enhancing Metaheuristics through Hebbian-Learned Generative Guidance

Mar 31, 2026

Traditional metaheuristic algorithms struggle to accumulate and reuse successful experiences from their search processes. This work proposes the Associative Constructive Evolution (ACE) framework, which leverages a Generative Construction Automaton (GCA) to extract co-occurring successful operation patterns from the search trajectories of metaheuristics such as evolutionary algorithms (EA) and particle swarm optimization (PSO). By integrating Hebbian learning, guided sampling, and symbolic abstraction, ACE autonomously generates reusable macro-operations, enabling knowledge-driven intelligent search. Experimental results demonstrate that ACE-PSO achieves a 27.5% higher success rate and 49.6% faster convergence in maze navigation tasks, while ACE-EA improves fitness by 10.1% in molecular design and automatically discovers 126 chemically interpretable macro-operations.

0 citationsRead paper

Multi-Constrained Evolutionary Molecular Design Framework: An Interpretable Drug Design Method Combining Rule-Based Evolution and Molecular Crossover

Jan 15, 2026

This work proposes MCEMOL, a novel framework addressing the limitations of conventional deep learning–based drug design approaches—namely their heavy reliance on large datasets, high computational cost, and poor interpretability. MCEMOL introduces a dual-layer evolutionary mechanism that uniquely integrates interpretable rule evolution with molecular structural crossover operations. By combining message-passing neural networks, rule-level evolution, and molecular crossover and mutation strategies—while embedding chemical and pharmacophoric constraints—the method efficiently generates novel compounds from only a small set of initial molecules. The generated molecules are 100% chemically valid, exhibit high structural diversity and favorable drug-like properties, and demonstrate superior performance in symmetry, pharmacophore matching, and stereochemical integrity, thereby ensuring both scientific credibility and practical utility.

0 citationsRead paper

Evaluating Large Language Models for IUCN Red List Species Information

Oct 03, 2025

Large language models (LLMs) are increasingly deployed in biodiversity conservation, yet their reliability across core dimensions of the IUCN Red List—taxonomy, conservation status, distribution, and threats—remains unassessed. Method: We systematically evaluated five state-of-the-art LLMs against the IUCN framework using a species-level, multidimensional verification protocol. Contribution/Results: We identify a critical “knowledge–reasoning gap”: while taxonomic classification achieves 94.9% accuracy, conservation status assessment drops to 27.2%. Models exhibit systematic bias toward charismatic vertebrates, potentially exacerbating conservation inequity. Crucially, we attribute this bias to inherent architectural limitations—not merely training data deficiencies—thereby clarifying the operational boundaries of LLMs in conservation decision-making. We propose an expert-validated human–AI collaboration paradigm, establishing a methodological benchmark and practical guideline for deploying AI in biodiversity conservation.

0 citationsRead paper

Recent publications

Latest Papers

Associative Constructive Evolution: Enhancing Metaheuristics through Hebbian-Learned Generative Guidance

Mar 31, 2026

Traditional metaheuristic algorithms struggle to accumulate and reuse successful experiences from their search processes. This work proposes the Associative Constructive Evolution (ACE) framework, which leverages a Generative Construction Automaton (GCA) to extract co-occurring successful operation patterns from the search trajectories of metaheuristics such as evolutionary algorithms (EA) and particle swarm optimization (PSO). By integrating Hebbian learning, guided sampling, and symbolic abstraction, ACE autonomously generates reusable macro-operations, enabling knowledge-driven intelligent search. Experimental results demonstrate that ACE-PSO achieves a 27.5% higher success rate and 49.6% faster convergence in maze navigation tasks, while ACE-EA improves fitness by 10.1% in molecular design and automatically discovers 126 chemically interpretable macro-operations.

0 citationsRead paper

Multi-Constrained Evolutionary Molecular Design Framework: An Interpretable Drug Design Method Combining Rule-Based Evolution and Molecular Crossover

Jan 15, 2026

This work proposes MCEMOL, a novel framework addressing the limitations of conventional deep learning–based drug design approaches—namely their heavy reliance on large datasets, high computational cost, and poor interpretability. MCEMOL introduces a dual-layer evolutionary mechanism that uniquely integrates interpretable rule evolution with molecular structural crossover operations. By combining message-passing neural networks, rule-level evolution, and molecular crossover and mutation strategies—while embedding chemical and pharmacophoric constraints—the method efficiently generates novel compounds from only a small set of initial molecules. The generated molecules are 100% chemically valid, exhibit high structural diversity and favorable drug-like properties, and demonstrate superior performance in symmetry, pharmacophore matching, and stereochemical integrity, thereby ensuring both scientific credibility and practical utility.

0 citationsRead paper

Evaluating Large Language Models for IUCN Red List Species Information

Oct 03, 2025

Large language models (LLMs) are increasingly deployed in biodiversity conservation, yet their reliability across core dimensions of the IUCN Red List—taxonomy, conservation status, distribution, and threats—remains unassessed. Method: We systematically evaluated five state-of-the-art LLMs against the IUCN framework using a species-level, multidimensional verification protocol. Contribution/Results: We identify a critical “knowledge–reasoning gap”: while taxonomic classification achieves 94.9% accuracy, conservation status assessment drops to 27.2%. Models exhibit systematic bias toward charismatic vertebrates, potentially exacerbating conservation inequity. Crucially, we attribute this bias to inherent architectural limitations—not merely training data deficiencies—thereby clarifying the operational boundaries of LLMs in conservation decision-making. We propose an expert-validated human–AI collaboration paradigm, establishing a methodological benchmark and practical guideline for deploying AI in biodiversity conservation.

0 citationsRead paper