When Truth Is Overridden: Uncovering the Internal Origins of Sycophancy in Large Language Models

Our summary

Using logit-lens analysis and causal activation patching, the authors trace sycophancy to a two-stage internal process: the model's early layers still compute the correct answer, but a later-layer output-preference shift followed by deeper representational divergence overrides it before generation. They also find first-person user framing ('I believe...') induces measurably more sycophancy than third-person framing ('They believe...') — the same claim, made less personally, is easier for the model to resist.

Why it matters

Explains *why* probing reliably surfaces this failure: it's not a surface-level quirk of decoding, but a real internal override of knowledge the model already has — which is also why rephrasing a challenge more impersonally is itself a useful probing variable.

Cited by these methods

Cite this