HASE co-evolves model weights, harness
Harness-Aware Self-Evolving (HASE) is an agentic reinforcement learning framework that allows a single model to co-evolve its policy weights, task solutions, and environment harness in a unified multi-turn action space. By enabling the model to dynamically modify harness components like prompts, memory formatting, and validators, HASE allows smaller models like Qwen3-8B to match or beat the performance of models as large as 120B parameters in domains like text classification and alpha factor mining.
Co-evolving the evaluation harness alongside model policy weights is a critical paradigm shift that shows parameter scale is no longer the sole bottleneck to agentic performance.
- –Allowing models to modify guidance and validation components in a unified action space breaks the limits of traditional reinforcement learning.
- –The dramatic performance boost of a Qwen3-8B matching a 120B model highlights massive potential compute and parameter efficiency gains.
- –Empowering models to edit their own validators raises significant safety and alignment questions, particularly around potential reward hacking.
DISCOVERED
1d ago
2026-07-08
PUBLISHED
1d ago
2026-07-08
RELEVANCE
AUTHOR
Discover AI