Researchers develop cognitive test to measure how AI agents adapt to hidden tool reliability changes

When tools a language model agent relies on silently become less effective, how does it respond? Researchers studying tool-use behavior in LLMs have developed a novel test inspired by cognitive psychology to measure exactly that—and the results show significant inconsistency across models.

The Set-Shifting Test#

The team borrowed the concept of "set-shifting" from cognitive science, a principle used to understand how humans adapt when established patterns change. Their framework gives LLM agents access to multiple redundant tools—many performing the same task but with varying levels of hidden reliability. As the conversation progresses, the researchers shift which tools actually work best, allowing them to observe whether agents notice and adjust their strategy or continue with outdated routines.

What the Models revealed#

The findings expose stark differences across the LLM landscape:

• Rigid behavior: Some models quickly lock into a fixed pattern and stick with it for the remainder of the session, even after conditions shift
• Continued exploration: Others keep calling different tool groups throughout, never settling on a consistent approach
• Capability gap: Less capable models frequently skip the most reliable tool entirely, while state-of-the-art models tend to invoke it alongside other options

The instability appears even when researchers prompt agents with explicit policies designed to improve behavior.

Why This Matters#

The findings highlight a fundamental brittleness in current LLM agents. When important contextual information changes in ways that aren't directly obvious—such as a tool's underlying reliability—agents often fail to adapt smoothly. The researchers developed a suite of new metrics to quantify exactly how these behavioral shifts occur, providing a foundation for future improvements in agent robustness and reliability.