Beacon: Single-Turn Diagnosis and Mitigation of Latent Sycophancy in Large Language Models
This addresses the issue of alignment drift in large-scale generative systems for AI safety and reliability, providing a reproducible foundation for study and mitigation.
The paper tackles the problem of latent sycophancy in large language models, where models prioritize user agreement over truthfulness, and introduces Beacon, a benchmark that measures this bias across twelve state-of-the-art models, revealing it decomposes into sub-biases that scale with model capacity.
Large language models internalize a structural trade-off between truthfulness and obsequious flattery, emerging from reward optimization that conflates helpfulness with polite submission. This latent bias, known as sycophancy, manifests as a preference for user agreement over principled reasoning. We introduce Beacon, a single-turn forced-choice benchmark that isolates this bias independent of conversational context, enabling precise measurement of the tension between factual accuracy and submissive bias. Evaluations across twelve state-of-the-art models reveal that sycophancy decomposes into stable linguistic and affective sub-biases, each scaling with model capacity. We further propose prompt-level and activation-level interventions that modulate these biases in opposing directions, exposing the internal geometry of alignment as a dynamic manifold between truthfulness and socially compliant judgment. Beacon reframes sycophancy as a measurable form of normative misgeneralization, providing a reproducible foundation for studying and mitigating alignment drift in large-scale generative systems.