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

2026-05-26

AI Summary

This work addresses the challenge that large language model agents in reinforcement learning often redundantly invoke external tools due to ambiguous knowledge boundaries, struggling to decide whether to rely on internal parametric knowledge or external resources. The authors propose AKBE, a novel method that explicitly defines instance-level knowledge boundaries by dynamically probing them through dual-path reasoning—with and without tool access—and generates in-policy supervision signals based on correctness comparisons between the two paths. This approach avoids reward hacking without modifying the reward function and is compatible with diverse reinforcement learning algorithms, enabling plug-and-play optimization of tool usage. Evaluated across seven question-answering benchmarks, AKBE improves average accuracy by 1.85%, reduces tool calls by 18%, and enhances usage efficiency by 25%, all without requiring trade-offs between accuracy and efficiency.

Abstract

Agentic reinforcement learning (RL) has proven effective for training LLM-based agents with external tool-use capabilities. However, we identify that agentic RL training induces increasing redundant tool calls and blurs the model's intrinsic knowledge boundary, where the model fails to distinguish when tools are needed versus when parametric knowledge suffices. Existing solutions based on reward shaping create coarse-grained optimization targets that tend to incentivize indiscriminate tool-call suppression, leading to reward hacking. In this paper, we propose AKBE (Agentic Knowledge Boundary Enhancement), an on-policy method that dynamically probes the model's intrinsic knowledge boundary through dual-path (with-tool and no-tool) rollouts during training. We define the knowledge boundary as the per-instance determination of whether tools are required and the minimum tool calls necessary. By comparing correctness across paths, AKBE categorizes trajectories and constructs targeted supervisory signals that guide efficient tool-use patterns for each question. These signals are integrated seamlessly into the agentic RL training loop. Experiments on seven QA benchmarks demonstrate that AKBE improves task accuracy by +1.85 on average and reduces tool calls by 18% over standard agentic RL, yielding 25% higher tool productivity without any accuracy-efficiency trade-off. Further analysis suggests its plug-and-play compatibility across different RL algorithms and the mechanism of each signal category. Our code is available at https://github.com/CuSO4-Chen/AKBE.

Problem

• agentic reinforcement learning
• intrinsic knowledge boundary
• tool-use efficiency
• redundant tool calls
• knowledge-tool distinction

Innovation

• Agentic Reinforcement Learning
• Intrinsic Knowledge Boundary
• On-Policy Learning
• Tool-Use Efficiency
• Dual-Path Rollout