Constant-Context Skill Learning for LLM Agents
A novel approach named constant-context skill learning tackles the challenge of balancing privacy, cost, and capability in personal assistants powered by LLMs. While cloud-based models manage complex workflows, they risk revealing sensitive information to external APIs; in contrast, local models maintain privacy but lack reliability. This new method develops reusable strategies within lightweight task-family modules, relying solely on current observations and a concise state block for inference. A deterministic tracker creates this state based on task advancement and offers subgoal rewards for step-level SFT and online RL training. The method underwent testing on ALFWorld, WebShop, and SciWorld.
Key facts
- arXiv:2605.05413v1
- Announce Type: new
- Proposes constant-context skill learning
- Context-to-weights framework for recurring agent workflows
- Reusable procedures learned in lightweight task-family modules
- Inference conditions only on current observation and compact state block
- Deterministic tracker renders state block from task progress
- Tested on ALFWorld, WebShop, and SciWorld
Entities
Institutions
- arXiv