From Knowing to Acting: Benchmarking Self-Awareness Capability of LLM Agents

Researchers introduce KAPRO, a framework for evaluating whether LLM agents can accurately determine when to use external tools versus relying on internal knowledge. The study reveals that open-source models suffer from tool overuse due to pattern matching, while proprietary models show better self-awareness, highlighting a critical gap in current AI agent capabilities.

This research addresses a fundamental limitation in autonomous AI agents: the ability to self-assess whether a problem requires external tools or can be solved internally. As LLM agents become increasingly integrated into production systems, distinguishing between genuine tool necessity and unnecessary tool invocation has real implications for cost, latency, and reliability. The KAPRO framework decouples metacognitive judgment from execution behavior, revealing a stark performance gap between model architectures.

The finding that open-source and instruction-following models exhibit significantly higher tool overuse reflects a deeper architectural weakness—these models rely on surface-level pattern matching rather than genuine reasoning about problem requirements. Proprietary models with reasoning capabilities demonstrate more sophisticated gating mechanisms, suggesting that advanced reasoning chains and reinforcement learning from human feedback better calibrate when external resources are genuinely needed.

For developers building AI applications, this matters directly. Excessive tool calls increase operational costs, introduce latency, and create unnecessary dependencies on external APIs. For enterprises deploying agents in production, self-awareness capability becomes a critical quality metric. The research indicates that cheaper, open-source alternatives may require additional fine-tuning or architectural modifications to match proprietary models' judgment reliability.

Looking ahead, this benchmark will likely drive improvements in model training methodologies focused on epistemic boundary recognition. Teams building agent systems should begin evaluating candidate models against self-awareness metrics rather than just task completion rates. The emergence of standardized measurement tools like KAware could accelerate adoption of more sophisticated, cost-efficient agent architectures.

• →LLM agents frequently invoke external tools unnecessarily, with open-source models showing 30-50% higher overuse rates than proprietary alternatives
• →Self-awareness capability—discerning internal vs. external problem-solving needs—correlates strongly with overall task success but remains underbenchmarked
• →Reasoning-oriented models demonstrate superior cognitive gating compared to instruction-following models due to more sophisticated metacognitive processes
• →KAPRO framework enables systematic evaluation of epistemic boundaries across internal, external, and hybrid task categories
• →Cost and latency optimization in agent systems requires evaluating self-awareness metrics alongside traditional performance benchmarks