Skip to content

R0020/2026-03-25/Q002/SRC02/E01
Research R0020 — Prompt Engineering Gaps
Run 2026-03-25
Query Q002
Source SRC02
Evidence SRC02-E01
Type Statistical

Question reframing technique reduces sycophancy by 24 percentage points across frontier models

URL: https://arxiv.org/html/2602.23971v2

Extract

Three techniques tested across GPT-4o, GPT-5, and Sonnet-4.5 using 440 controlled prompts:

  1. Question reframing (most effective): Convert assertions into questions before answering. Example: "I believe pineapple belongs on pizza" becomes "Does pineapple belong on pizza?"
  2. 24 percentage-point reduction in sycophancy scores (0-15 scale)

  3. Outperformed explicit "don't be sycophantic" instructions

  4. Perspective reframing (less effective): Convert first-person to third-person. "I believe..." becomes "The user believes..."

  5. Up to 63.8% improvement in debate settings with third-person persona

  6. No-sycophancy baseline: Explicit instruction not to be sycophantic

  7. Less effective than question reframing

Additional finding: Sycophancy increased monotonically with epistemic certainty (convictions > beliefs > statements). I-perspective framing amplified sycophancy relative to user-perspective.

Relevance to Hypotheses

Hypothesis Relationship Strength
H1 Supports Rigorous research demonstrates effective prompt-level techniques
H2 Contradicts Techniques with quantitative evidence exist
H3 Supports These techniques are in academic papers, not mainstream prompt guides

Context

This is the strongest evidence for actionable prompt-level sycophancy mitigation. The finding that question reframing outperforms explicit "don't be sycophantic" instructions is highly relevant — it suggests that structural prompt changes are more effective than directive approaches. However, this technique is published in a February 2026 arXiv paper and has not yet appeared in mainstream vendor documentation.

Notes

The finding that epistemic certainty amplifies sycophancy (convictions > beliefs > statements) has practical implications: prompts that express strong opinions will elicit more sycophantic responses than prompts that frame the same content as questions.