GASLIGHTBENCH: Quantifying LLM Susceptibility to Social Prompting

Lening Nick Cui, Sahil Ghosh, Gareth Lee, Xuanzhe Yao, Swarit Srivastava, William H. Logian, Michael Li, Kevin Zhu, Sunishchal Dev, Michael Saxon, Aaron Sandoval, Sean O'Brien, Ellie Podoshev

Keywords: benchmark, sycophancy, llms, dataset, ai safety, prompt manipulation, social cues, multi-turn dialogue, LLM-as-a-judge

TL;DR: GaslightBench: a plug-and-play benchmark (24,160 single-turn; 720 multi-turn) using social/linguistic modifiers to probe sycophancy, comprehensive analyses of ten modifier families across nine statement domains.

Abstract: Large language models (LLMs) can be manipulated by simple social and logistic cues, producing sycophancy. We introduce GaslightBench, a plug-and-play benchmark that systematically applies socio-psychological and linguistic modifiers (e.g. flattery, false citations, assumptive language) to trivially verifiable facts to test model sycophancy. The dataset comprises a single-turn prompting section of 24,160 prompts spanning nine domains and ten modifiers families, and a multi-turn prompting section of 720 four-turn dialogue sequences, evaluated via LLM-as-a-judge. We find that state-of-the-art models consistently score highly in single-turn prompting (92%-98% accuracy) while multi-turn prompting results in highly varied accuracies ranging from ~60%-98%. We find that injecting bias into the model via a descriptive background induces the most sycophancy, up to 23% in naive single-turn prompting. Across almost all the models we analyze, we also find a statistically significant difference in verbosity between sycophantic and non-sycophantic responses. GaslightBench standardizes stress tests of prompt-style susceptibility and identifies which social cues most undermine factual reliability. We will release all code and data upon publication.

Submission Number: 226