Research shows that providing LLMs with access to irrelevant tools can negatively impact their ability to answer questions from internal knowledge. Testing across 10 domains reveals that tool availability and prior tool interactions alter model behavior even when tools are not required.
HOW THIS AFFECTS YOU
●
builderYou should minimize the toolset exposed to models in specific contexts to prevent performance degradation.
●
researcherThis suggests a new direction for evaluating tool-use interference in agentic workflows.