Harness Continual Learning: Continual Adaptation Beyond Model Parameters
The paper argues that agents can forget old behavior even when their model weights never change.
It defines “Harness Continual Learning” for systems that adapt through prompts, memories, tools, skills, and routing rules around a frozen foundation model. The authors call the resulting regressions “harness-level forgetting” and propose guarded updates, where a candidate harness is evaluated before being committed. Their experiments report capability accumulation and failure recovery across reasoning, multimodal perception, and open-world interaction, with relative gains above 10% over baselines in multiple settings. ArXiv · AI/CL/LG's note
It defines “Harness Continual Learning” for systems that adapt through prompts, memories, tools, skills, and routing rules around a frozen foundation model. The authors call the resulting regressions “harness-level forgetting” and propose guarded updates, where a candidate harness is evaluated before being committed. Their experiments report capability accumulation and failure recovery across reasoning, multimodal perception, and open-world interaction, with relative gains above 10% over baselines in multiple settings. ArXiv · AI/CL/LG's note
score 5