The Behavioral Ratchet: How Conversational History Shapes LLM Sycophancy

We investigate how conversational context shapes sycophancy in large language models, testing 80,433 trials across six models (4B–72B parameters) spanning four architecture families. Our central finding is the behavioral ratchet: the pattern of conversational history—not its length—is the primary driver of in-context sycophancy. Agreement-pattern filler roughly doubles sycophancy compared to correction-pattern filler (p < 10⁻¹⁴ in every model tested). This effect is universal across all model sizes and architectures. Code, data, and full analysis available at: github.com/thtskaran/context_window_research