Efficient Agentic Reinforcement Learning with On-Policy Intrinsic Knowledge Boundary Enhancement

Researchers introduced AKBE, an on-policy RL method for LLM agents that learns when to rely on internal knowledge versus external tools.

AKBE uses dual-path rollouts, with and without tools, to identify each model’s intrinsic knowledge boundary for tool use decisions.

Across seven QA benchmarks, the method improved accuracy while reducing unnecessary tool calls compared with standard agentic RL.

The approach tackles a core efficiency problem in tool-using agents: better performance without wasted tool usage.