A new multi-turn benchmark shows that sycophancy compounds in LLMs, with accuracy dropping by up to 47 percentage points over seven rounds of biased follow-ups.
Given the diverse range of categories in the MMLU dataset, including engi- neering, physics, and law, we tailored the prompt to reflect the specific category of each question
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
TRUTH DECAY: Quantifying Multi-Turn Sycophancy in Language Models
A new multi-turn benchmark shows that sycophancy compounds in LLMs, with accuracy dropping by up to 47 percentage points over seven rounds of biased follow-ups.