A*STAR Career Development Fund (CDF) Day 2025 - Poster Award Winner
A*STAR supports the next generation of scientific talent through its Career Development Fund (CDF), benefitting over 200 early-career researchers. The A*STAR CDF Day celebrated the achievements of these researchers, spotlighting their innovative work across diverse disciplines. The event served as a vibrant platform for knowledge exchange, networking, and collaboration among the next generation of scientific leaders.
Congratulations to Dr Yin Haiyan, Early Career Principal Investigator from A*STAR Centre for Frontier AI Research (ASTAR CFAR), on being named the A*STAR CDF Day 2025 Poster Award Winner for the project:
Towards Meta Continuous Reinforcement Learning Agents for Solving Text-based Games
This project explores how text-based agents can continually improve in long-horizon environments, where delayed feedback, evolving objectives, and changing tool availability require disciplined approaches to learning, adaptation, and reliable decision-making.
It frames long-horizon decision-making through step-resolved decomposition, leveraging agentic workflows that make reasoning and tool, or code use transparent and inspectable - revealing where errors occur and improvements can be made. The project introduces MermaidFlow, a verifiable workflow optimisation framework that encodes workflows in Mermaid for LLM-based code generation, ensuring that continuously evolving workflows remain valid and safety-constrained throughout optimisation.
Continual improvement is further enabled by treating interaction traces as structured evidence. The project develops an embodied robotic code-generation system with experience-driven symbolic constraint induction, automatically distilling failures into constraints that guide future decisions. It also investigates agentic memory, grounding memory-driven world modelling through counterfactual refinement of embodied multi-agent plans.
Together, this research contributes toward a scalable foundation for next-generation agentic and embodied AI, advancing long-horizon reasoning, reliable tool and code use, continual adaptation, and grounded decision-making. Looking ahead, the project sets the stage for agents that improve continuously as their scope and interfaces expand - safely extending tool and code use, accumulating richer memory-grounded world models, and progressing from individual competence to self-evolving, embodied multi-agent coordination in open-ended environments.
References
[1] InstructFlow: Adaptive Symbolic Constraint-Guided Code Generation for Long-Horizon Planning. https://openreview.net/pdf?id=nzwjvpCO4F
[2] MermaidFlow: Redefining Agentic Workflow Generation via Safety-Constrained Evolutionary Programming. https://openreview.net/pdf?id=ocpFXA7Fwr