Beacon: Single-Turn Diagnosis and Mitigation of Latent Sycophancy in Large Language Models

Sanskar Pandey, Ruhaan Chopra, Angkul Puniya +1

Large language models internalize a structural trade-off between truthfulness and obsequious flattery, emerging from reward optimization that conflates helpfulness with polite subm…