Discovering Language Model Behaviors with Model-Written Evaluations

Our summary

The paper that first documented sycophancy as an inverse-scaling effect: across 154 model-written evaluation datasets, the authors found larger, more RLHF-tuned models more readily echo a user's stated political or ideological leaning back to them, rather than converging on a single stance.

Why it matters

The historical origin of the whole concept — establishes that sycophancy isn't a minor side-effect but a direct, worsening product of the exact optimization (RLHF on human preference) used to train most deployed assistants.

Cited by these methods

Cite this