Research
RAIL Lab studies AI systems that reason reliably, act autonomously in digital and physical environments, align with human goals, and continually improve through experience.
Reliable Reasoning
Reasoning is the foundation of trustworthy intelligence. We study how large models solve tasks that require multi-step inference, structured thinking, and verifiable decisions across language, code, mathematics, vision, and interactive environments.
- Mathematical, code, and structured problem solving
- Long-chain reasoning and process-level verification
- Reward modeling and reinforcement learning for reasoning
- Multimodal and spatial reasoning in vision-language models
- Evaluation of reliability, robustness, and failure modes
Agentic AI
We build agents that can plan, use tools, interact with interfaces, collaborate with other agents, and complete complex tasks. Our research connects software agents, embodied agents, and multi-agent systems so that AI can operate beyond static dialogue.
- GUI, mobile, desktop, and web automation agents
- Tool-use agents, code agents, and task automation
- Embodied agents for spatial reasoning, planning, and action
- Multi-agent collaboration, communication, and reflection
- Retrieval-augmented and memory-equipped agents
Alignment
Intelligent systems should remain aligned with human intent, social context, safety requirements, and real user needs. We study alignment not only as preference tuning, but also as a broader problem of making AI systems useful, controllable, transparent, and responsible.
- Human preference alignment and feedback-driven optimization
- Safety, personalization, and trustworthy model behavior
- Alignment for education, tutoring, and learner-centered AI
- Interpretability, evaluation, and controllability of AI systems
- Robust alignment under distribution shift and long-tail cases
Learning and Self-Evolution
The laboratory is where reasoning, agency, and alignment are integrated into systems that can improve after deployment. We study how models learn from interaction, feedback, tool use, and self-generated experience so that AI capabilities can grow continuously.
- Self-evolving agents and iterative improvement loops
- Generator-verifier co-evolution and policy/reward co-optimization
- Continual adaptation for deployed AI systems
- World models and feedback loops for physical and digital tasks
- Benchmarks that measure progress across reasoning, action, and alignment