Same Probabilities, Different Advice: Gain-Loss Framing Effects in High-Stakes LLM Decision Support

[arXiv] Yuyao Zhang, Yakun Cui, Han Zhu, Jinsong Lu, Yike Guo and Sirui Han published a paper stating that it remains unclear whether large models' high-risk advice changes when the same probability is expressed in complementary gain and loss frames. The authors introduce a controlled benchmark, FrameBench-HPS, derived from 240 high-risk scenarios into 2,880 prompts, covering healthcare, finance, education/career and law/consumer rights. Across seven large models, the authors report that gain/loss framing changed the advice category in 32.4% of the full set, mostly involving uncertainty rather than direct reversals; human evaluation confirmed that perceived risk and prescriptiveness shifted accordingly. Robustness experiments show that the framing effect persists under dialogue rewrites and exceeds ordinary repeated-generation fluctuations. The paper states that probability normalization can significantly reduce framing sensitivity, and that probability representation is a feasible lever for improving the consistency of high-risk decision support.