The authors propose ParametricSkills, a framework that converts free-form textual skills into parameters at test time by training a hypernetwork to generate LoRA adapters. This approach enables context-free skill exploitation, addressing the difficulty of adhering to instructions in complex scenarios.

  • The method constructs a large-scale skill library and synthesizes single-turn and multi-turn exploitation trajectories using OpenCode.
  • A hypernetwork receives textual skills to parameterize both skill content and exploitation methodology as LoRA adapters.
  • Experimental results on six software engineering subtasks show an average improvement of 6.44 points over in-context learning when judged by DeepSeek-V4-Flash.
  • The framework achieves significantly higher BERT Score and F1 scores compared to baseline methods.

Parametric skills are inherently accumulative, offering a promising avenue toward test-time continual learning for agentic capabilities.