RAGEN-2: Reasoning Collapse in Agentic RL
Why agentic RL reasoning collapses into templates, and how to catch it.
AI Research · Engineer · MusicianHi, this is
Hi, this is
Deimos
I train foundation models at Apple.
Welcome to my journey.About
Between the lab and the stage
I'm a research engineer improving foundation models, with a focus on agentic RL post-training. I enjoy doing research at the frontier of human knowledge, and I like it even more when it makes an impact in real-world products.
Curiosity drives me day to day. Humanity and kindness are what I hope is left when the work is done. I believe AI should bring us closer and add warmth to the world.
Outside of work, I'm a musician. I play keys in a pop band called ABT. We were the first Bay Area band to ever perform at Anime Expo, the largest anime convention in North America.
In my spare time, I enjoy reading, playing video games, composing and making piano covers, and hunting for good food.
Research
I'm interested in bridging frontier research and real-world applications at scale, and in pointing that work toward a larger positive social impact. My experience spans post-training infrastructure, RL, recursive self-improvement, and multimodal intelligence.
Selected papers
RAGEN: Self-Evolution in LLM Agents via Multi-Turn RL
One of the first open-source infrastructures for multi-turn RL training.
Projects
This part is under construction. Code agents are reshaping how engineering gets done, and I'm busy building the next big thing.
Music
Ohne Musik wäre das Leben ein Irrtum.Friedrich Nietzsche01
Composing
Soundtracks and re-arrangements I wrote.
Get in touchdeimosgin@gmail.com
Happy to talk about research, startup ideas, music, games, cat videos, or good food :P
I'm always open to meeting new people and helping where I can.