Self-evolving AI
Models and agents that adapt from feedback, trajectories, tool use, and experience.
Research
Research
We study AI systems that learn from experience, reason in language, and stay both efficient and responsible under real constraints.
BREATHE AI Lab's central question is how language models can move beyond static prediction toward systems that improve through interaction with tools and feedback, stay efficient enough to deploy at scale, and earn trust through evidence rather than impression.
Language agents
Systems that plan, use tools, collaborate, and complete long-horizon tasks.
Efficient language models
Compact models that remain capable, controllable, and useful under deployment constraints.
Responsible evaluation
Measurements that expose capability, robustness, failure modes, and usefulness.
Jun 2025
EMNLP 2025
Through the Valley: Path to Effective Long CoT Training for Small Language Models
Analysis and training guidance for long chain-of-thought supervision in small language models, including when long CoT can degrade performance.