RL Training Increases Spurious Tool Use by 39 Percent
September 16, 2026
Reinforcement learning can cause LLM agents to develop shortcut policies, invoking tools based on superficial prompt cues rather than task necessity. In controlled environments, agents showed spurious tool invocation rates increasing by up to 39 percent when training data contained correlations between cues and specific tools.
HOW THIS AFFECTS YOU
●
builderYour agent's reliability may degrade in production if it relies on superficial patterns rather than logic.
●
researcherYou must account for causal irrelevance in your RL training data to prevent shortcut learning.