LLMs Show Systematic Biases When Arbitrating Conflicting Text and Numerical Evidence
August 21, 2026
A new benchmark reveals that LLMs exhibit systematic preferences when text, numbers, and tools provide conflicting evidence. Models consistently favor temporal recency over explicit reliability cues and show distinct modality-based preferences during arbitration.
HOW THIS AFFECTS YOU
●
builderYou must implement explicit arbitration logic if your application requires weighing conflicting numerical and textual inputs.
●
researcherYou should account for systematic modality-based biases when evaluating agentic reasoning.