BMJ digital health & AI · 2026
Exposing the fragility of LLM reasoning through bias-inducing prompts: evidence from BiasMedQA.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Kim Su Hwan, Ziegelmayer Sebastian, Busch Felix, Mertens Christian J, Keicher Matthias, Adams Lisa C, Bressem Keno K, Braren Rickmer, Makowski Marcus R, Kirschke Jan S, Hedderich Dennis M, Wiestler Benedikt
Аннотация
OBJECTIVES: To evaluate the impact of large language model (LLM) reasoning on model susceptibility to cognitive bias-inducing prompts.
METHODS AND ANALYSIS: The performance of Llama-3.3-70B, Qwen3-32B and Gemini-2.5-Flash, along with their reasoning-enhanced variants, was evaluated in the public BiasMedQA dataset developed to evaluate seven established cognitive biases in 1273 clinical case vignettes. Each model was tested using a base prompt, a debiasing prompt with the instruction to actively mitigate cognitive bias and a few-shot prompt with additional sample cases of biased responses. Beyond the seven biases from BiasMedQA, Gemini-2.5-Flash was additionally tested using four unpublished bias-inducing prompts to unveil signs of potential data contamination and actively investigate brittleness. For each model pair, two mixed-effects logistic regression models were fitted to determine the impact of biases and mitigation strategies on performance.
RESULTS: In all three models, the reasoning-enhanced variant achieved higher rates of correct responses (Llama-3.3-70B: 72.5-82.1% vs 61.0-73.4%, Qwen3-32B: 71.7-78.7% vs 55.5-64.1%, Gemini-2.5-Flash: 81.8-88.6% vs 80.0-83.7%). The performance of Gemini-2.5-Flash dropped considerably when exposing it to four additional unpublished bias-inducing prompts (from 80.0-88.6% to 47.4-86.1%), hinting at potential contamination of its training data and exposing underlying brittleness.In Llama-3.3-70B and Gemini-2.5-Flash, reasoning amplified model vulnerability to several bias-inducing prompts, while reasoning reduced susceptibility of Qwen3-32B to one of the seven biases. The debiasing and few-shot prompting approaches demonstrated statistically significant reductions in biased responses across all three model architectures.
CONCLUSION: In none of the three LLMs, reasoning was able to consistently reduce vulnerability to bias-inducing prompts, revealing the fragility of the reasoning capabilities purported by the model developers.
METHODS AND ANALYSIS: The performance of Llama-3.3-70B, Qwen3-32B and Gemini-2.5-Flash, along with their reasoning-enhanced variants, was evaluated in the public BiasMedQA dataset developed to evaluate seven established cognitive biases in 1273 clinical case vignettes. Each model was tested using a base prompt, a debiasing prompt with the instruction to actively mitigate cognitive bias and a few-shot prompt with additional sample cases of biased responses. Beyond the seven biases from BiasMedQA, Gemini-2.5-Flash was additionally tested using four unpublished bias-inducing prompts to unveil signs of potential data contamination and actively investigate brittleness. For each model pair, two mixed-effects logistic regression models were fitted to determine the impact of biases and mitigation strategies on performance.
RESULTS: In all three models, the reasoning-enhanced variant achieved higher rates of correct responses (Llama-3.3-70B: 72.5-82.1% vs 61.0-73.4%, Qwen3-32B: 71.7-78.7% vs 55.5-64.1%, Gemini-2.5-Flash: 81.8-88.6% vs 80.0-83.7%). The performance of Gemini-2.5-Flash dropped considerably when exposing it to four additional unpublished bias-inducing prompts (from 80.0-88.6% to 47.4-86.1%), hinting at potential contamination of its training data and exposing underlying brittleness.In Llama-3.3-70B and Gemini-2.5-Flash, reasoning amplified model vulnerability to several bias-inducing prompts, while reasoning reduced susceptibility of Qwen3-32B to one of the seven biases. The debiasing and few-shot prompting approaches demonstrated statistically significant reductions in biased responses across all three model architectures.
CONCLUSION: In none of the three LLMs, reasoning was able to consistently reduce vulnerability to bias-inducing prompts, revealing the fragility of the reasoning capabilities purported by the model developers.