5-35B-A3B with mini-SWE-agent, an external harness that controls what the model observes, how commands are executed, what feedback is returned, and when work stops.
Coding agents improve more when their model and harness co-evolve
Self-HPO alternates harness redesign with reinforcement learning, targeting new failure modes after each policy update.
Big Tech
Zhepei Wei · Wenlin Yao · Baolin Peng · Danqing Wang · Qianhui Wu · Hao Cheng · +2 more
Microsoft Research · University of Virginia · Carnegie Mellon University
Research Digest··3 min read
Wei et al.
§