Linear Probe Penalties Reduce LLM Sycophancy

Henry Papadatos, Rachel Freedman

Large language models (LLMs) are often sycophantic, prioritizing agreement with their users over accurate or objective statements. This problematic behavior becomes more pronounced…