Search-Equipped LLM Agents Exhibit Cognitive Biases Based on Evidence Order
November 5, 2026
An analysis of 10 models across 1,000 synthetic search environments shows that initial evidence order significantly influences final answers during active search. The study finds that dead-end search histories reduce strategy switching and that agents are more sensitive to retrieval order than to context length alone.
HOW THIS AFFECTS YOU
●
builderYou should implement robust retrieval re-ranking to prevent agent bias from early search results.
●
researcherThe findings suggest search-based agent evaluations must control for document ordering and history effects.