It's Not Always Sycophancy: Measuring LLM Conformity as a Function of Epistemic Uncertainty

Summary

Disentangles two distinct mechanisms behind an LLM abandoning its position - sycophantic conformity, which occurs even at absolute certainty, and uncertainty-driven conformity, which scales with the model's epistemic uncertainty - via the MUSE two-stage framework. It also shows both grow with the model's perceived expertise of the interlocutor and with the plausibility of the pushback, which maps directly onto hierarchical structures where a designated senior or manager agent pushes back. Useful for arguing that consensus collapse under hierarchy is not reducible to RLHF sycophancy, and for specifying which conformity channel a given structure activates.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-09-07 literature sweep as a medium-priority candidate for the Architecture as Epistemology (ISR) paper (angle: correlated errors / homogeneity across LLM agents). Verified against: Opened https://arxiv.org/abs/2605.27288 (arXiv abstract page; confirmed title, 8 authors, cs.CL, v1 26 May 2026, full abstract read)

Critique / open questions

Queued — not yet assessed.

Key quotes

Large language models (LLMs) are known to abandon their initial stance to conform to user pushback. (abstract)

While prior research largely attributes this behavior to sycophancy learned during reinforcement learning from human feedback, we hypothesize that conformity is also driven by a model's epistemic uncertainty at inference time. (abstract)