Eric Yoon
I’m interested in post-training, inference, and the data and environments that make models more capable.
Some things I’m thinking about:
- RL environments with meaningful state, long-horizon credit assignment, and verifiers that hold up under optimization.
- Data mixtures, on-policy rollouts, and separating capability gains from benchmark overfitting.
- KV-cache locality, prefill/decode disaggregation, and the latency–throughput frontier of inference.
- Test-time search and distilling useful trajectories back into the policy.
I’m also interested in languages and how people acquire them – I built Skylang, a platform to learn Korean.