Skip to content

R0043/2026-04-01/Q001/SRC06/E01

Research R0043 — Sycophancy Vocabulary
Run 2026-04-01
Query Q001
Source SRC06
Evidence SRC06-E01
Type Statistical

Counter-finding: LLMs exhibit opposite of acquiescence bias

URL: https://arxiv.org/abs/2509.08480

Extract

Contrary to the hypothesis that LLMs would exhibit acquiescence bias (tendency to agree with statements), the study found the opposite: across English tasks, models showed "a bias towards replying 'no', independent of whether that indicates agreement or disagreement."

Study scope: 37,975 question variations across 5 LLM models (Llama-3.1-8B, Mistral-Small-24B, Gemma-2-27b, Llama-3.3-70B, GPT-4o), 9 tasks, 3 languages (English, German, Polish).

The author concludes: "it is questionable whether [LLMs] can be used to simulate human responses" — the bias patterns do not match human acquiescence patterns.

JUDGMENT: This is a significant finding for the vocabulary mapping because it demonstrates that borrowing terms from survey methodology ("acquiescence bias") and applying them to LLMs can be misleading. The same term describes different phenomena in different contexts.

Relevance to Hypotheses

Hypothesis Relationship Strength
H1 N/A Neither supports nor contradicts the existence of vocabulary
H2 N/A Not about domain gaps
H3 Strongly supports Demonstrates that terms borrowed across domains can describe different (even opposite) phenomena

Context

This finding complicates the vocabulary mapping by showing that terminological borrowing can be misleading. "Acquiescence bias" in survey methodology means tendency to agree; when tested in LLMs, the opposite was found. This validates H3's caution that cross-domain terms may not be true synonyms.

Notes

The Lim & Lee (2024) terminological shift from "sycophancy" to "agreeableness bias" between preprint versions (noted in other sources) further illustrates the instability of this vocabulary.