Zixing Lei

Autonomous model post-training can outperform human-crafted instruct baselines on hard coding benchmarks while completely eliminating reward hacking once experimental exploration is governed by a sandboxed OS and a dynamic peer-review DAG.

NavMCP transforms isolated navigation rollouts into a cohesive, persistent interaction framework, leading to a 14.9 percentage point improvement over traditional episodic interfaces on long-horizon tasks.