Selected work

Representative Papers

ExperienceIndex: Artifact-Grounded Memory

Oct 07, 2026

This study addresses the limitation that AI agents lack artifact-based experiential memory, which results in low quality and high costs for knowledge-intensive tasks. To this end, this work proposes a pioneering artifact-oriented lightweight experience architecture. The method extracts structured knowledge from reasoning trajectories to construct an experience layer storing individual artifacts and inter-artifact relationships, while integrating a retrieval mechanism as middleware to guide agent decision-making. Its core contributions lie in enabling cross-task experience transfer and knowledge distillation between strong and weak models. Experimental results demonstrate that this architecture improves response quality by up to 11.0 points and reduces online costs by 50.5%, further validating its superior generalization and collaborative capabilities.

0 citationsRead paper

Deep Reinforcement Learning for Reliability Based Bi-Objective Portfolio Optimization

Jul 07, 2026

This study addresses the limitations of traditional static portfolio optimization in handling sequential decision-making, tail risk, and market frictions such as transaction costs. To this end, it proposes a deep reinforcement learning–based bi-objective dynamic optimization framework that jointly maximizes expected return and minimizes downside risk by integrating three risk measures—variance, Conditional Value-at-Risk (CVaR), and Entropic Value-at-Risk (EVaR)—while explicitly incorporating transaction costs and position constraints. The approach innovatively applies deep reinforcement learning to multi-objective portfolio optimization, modeling asset return uncertainty through a combination of GARCH(1,1), extreme value theory, and t-copula. Scenario generation employs quasi-Monte Carlo simulation, and the policy is trained using the Proximal Policy Optimization (PPO) algorithm. Empirical validation on equity index data from ten countries across pre-, mid-, and post-pandemic periods demonstrates that the proposed method significantly outperforms benchmarks such as NSGA-II in terms of risk–return trade-off, control of extreme downside risk, and scalability to high-dimensional portfolios.

0 citationsRead paper

Predicting effect of novel treatments using molecular pathways and real-world data

Sep 08, 2025

Predicting preclinical efficacy of novel drugs remains challenging due to limited experimental data and poor generalizability across drug–disease pairs. Method: This paper proposes a generalizable computational framework integrating molecular pathway mechanisms with real-world patient data. It introduces a drug–pathway interaction weight scoring model—implemented via two complementary algorithms—and couples it with clinical outcome association analysis to estimate therapeutic efficacy for untested drug–disease combinations. The framework is modular, extensible, and explicitly defines its domain of applicability. Results: Validation on real-world datasets demonstrates significantly improved cross-drug generalization performance. Key determinants of high predictive accuracy are identified, including pathway perturbation magnitude and consistency between molecular perturbations and clinical phenotypes. The framework provides an iterative, mechanistically interpretable foundation for preclinical efficacy assessment, enabling hypothesis-driven prioritization of novel therapeutics.

0 citationsRead paper

Can Machines Philosophize?

Jul 01, 2025

This study investigates whether AI agent populations can reflect human philosophical stances, focusing on scientific realism. We propose a three-step empirical framework: (1) constructing an AI population using large generative language models; (2) administering standardized philosophical questionnaires to both human participants and AI agents; and (3) conducting statistical analyses to compare distributional properties and internal consistency between human and AI responses. To our knowledge, this is the first work to simulate and quantitatively characterize human philosophical positions via AI systems. Results show that the AI population exhibits a similar aggregate tendency toward anti-realism as humans do, yet demonstrates higher response stability and lower inter-agent variability. This approach establishes a reproducible, scalable paradigm for experimental philosophy and empirically validates AI agents as viable, complementary proxies for philosophical inquiry—offering distinct advantages in controllability, scalability, and measurement precision.

0 citationsRead paper

Recent publications

Latest Papers

ExperienceIndex: Artifact-Grounded Memory

Oct 07, 2026

This study addresses the limitation that AI agents lack artifact-based experiential memory, which results in low quality and high costs for knowledge-intensive tasks. To this end, this work proposes a pioneering artifact-oriented lightweight experience architecture. The method extracts structured knowledge from reasoning trajectories to construct an experience layer storing individual artifacts and inter-artifact relationships, while integrating a retrieval mechanism as middleware to guide agent decision-making. Its core contributions lie in enabling cross-task experience transfer and knowledge distillation between strong and weak models. Experimental results demonstrate that this architecture improves response quality by up to 11.0 points and reduces online costs by 50.5%, further validating its superior generalization and collaborative capabilities.

0 citationsRead paper

Deep Reinforcement Learning for Reliability Based Bi-Objective Portfolio Optimization

Jul 07, 2026

This study addresses the limitations of traditional static portfolio optimization in handling sequential decision-making, tail risk, and market frictions such as transaction costs. To this end, it proposes a deep reinforcement learning–based bi-objective dynamic optimization framework that jointly maximizes expected return and minimizes downside risk by integrating three risk measures—variance, Conditional Value-at-Risk (CVaR), and Entropic Value-at-Risk (EVaR)—while explicitly incorporating transaction costs and position constraints. The approach innovatively applies deep reinforcement learning to multi-objective portfolio optimization, modeling asset return uncertainty through a combination of GARCH(1,1), extreme value theory, and t-copula. Scenario generation employs quasi-Monte Carlo simulation, and the policy is trained using the Proximal Policy Optimization (PPO) algorithm. Empirical validation on equity index data from ten countries across pre-, mid-, and post-pandemic periods demonstrates that the proposed method significantly outperforms benchmarks such as NSGA-II in terms of risk–return trade-off, control of extreme downside risk, and scalability to high-dimensional portfolios.

0 citationsRead paper

Predicting effect of novel treatments using molecular pathways and real-world data

Sep 08, 2025

Predicting preclinical efficacy of novel drugs remains challenging due to limited experimental data and poor generalizability across drug–disease pairs. Method: This paper proposes a generalizable computational framework integrating molecular pathway mechanisms with real-world patient data. It introduces a drug–pathway interaction weight scoring model—implemented via two complementary algorithms—and couples it with clinical outcome association analysis to estimate therapeutic efficacy for untested drug–disease combinations. The framework is modular, extensible, and explicitly defines its domain of applicability. Results: Validation on real-world datasets demonstrates significantly improved cross-drug generalization performance. Key determinants of high predictive accuracy are identified, including pathway perturbation magnitude and consistency between molecular perturbations and clinical phenotypes. The framework provides an iterative, mechanistically interpretable foundation for preclinical efficacy assessment, enabling hypothesis-driven prioritization of novel therapeutics.

0 citationsRead paper

Can Machines Philosophize?

Jul 01, 2025

This study investigates whether AI agent populations can reflect human philosophical stances, focusing on scientific realism. We propose a three-step empirical framework: (1) constructing an AI population using large generative language models; (2) administering standardized philosophical questionnaires to both human participants and AI agents; and (3) conducting statistical analyses to compare distributional properties and internal consistency between human and AI responses. To our knowledge, this is the first work to simulate and quantitatively characterize human philosophical positions via AI systems. Results show that the AI population exhibits a similar aggregate tendency toward anti-realism as humans do, yet demonstrates higher response stability and lower inter-agent variability. This approach establishes a reproducible, scalable paradigm for experimental philosophy and empirically validates AI agents as viable, complementary proxies for philosophical inquiry—offering distinct advantages in controllability, scalability, and measurement precision.

0 citationsRead paper