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

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

AKBE uses dual rollouts with and without tools to identify a per-question knowledge boundary and create targeted training signals.

On seven QA benchmarks, it improved average accuracy by 1.85 points and cut tool calls by 18% versus standard agentic RL.

The approach aims to make tool-using agents more accurate and more efficient by reducing redundant tool use.