Simon Yu

I am a 2nd year PhD student at Northeastern University, advised by Weiyan Shi. In summer 2025, I interned at Orby AI, mentored by Peng Qi. I am a main organizer of the NeurIPS 2025 Workshop on Multi-Turn Interactions in LLMs. I was a M.Res. and BSc student at the University of Edinburgh, also part of the EdinburghNLP. I am part of the Cohere For AI community working with Marzieh Fadaee. I have done research about alignment, instruction tuning and safety.

My core research interest is self-play as the path for AI agents’ self-improvement: agents that get better by generating their own tasks, environments, and feedback, with no human supervision in the loop. I work on this from three angles:

1. Agents managing agents: Shepherd (a collaboration with Stanford NLP) turns an agent’s execution into a reversible, Git-like trace, so meta-agents can inspect, fork, replay, and revert other agents’ runs to supervise, optimize, and train them. Code at shepherd-agents/shepherd.
2. Self-Play Learning: SPIRAL shows that self-play on zero-sum games improves reasoning, with zero human-curated data.
3. Environment Scaling: Scaling the environments models learn from, including TextArena for multi-agent environments and evaluation and GEM for unified, scalable environment generation.

I also work on continual learning for agents, e.g. PolySkill on compositional skill learning.

One of the most influential lessons to me is from The Bitter Lesson by Richard Sutton and The Era of Experience by David Silver and Richard Sutton. The idea is not just limited to AI but can be applied to any choice in life. Always choose the path that benefits in the long run, instead of the path that might be easier in the short run.