IC-189Most LLMs show degraded instruction satisfaction when user instructions conflict with system messages, indicating difficulty in prioritizing system message constraints
Yanzhao Qin, Tao Zhang, Yanjun Shen, Wenjing Luo, sunhaoze, Yan Zhang, Yujing Qiao, weipeng chen, Zenan Zhou, Wentao Zhang, Bin CUI
When user instructions are misaligned (conflicting) with system messages, most models show lower ISR than on aligned instructions. The gap is largest for weaker models: GLM-4-9b drops from 48.3% (aligned) to 28.4% (misaligned), a 19.9-point drop. GPT-4o shows a smaller gap (77.8% vs. 71.4%). Notably, GPT-3.5-turbo is an exception, scoring higher on misaligned (47.7%) than aligned (41.9%) instructions, suggesting acute awareness of system message priority. The authors attribute the misaligned degradation to insufficient recognition of system message priority.
Evidence
correlational
Key metric
GPT-4o aligned 77.8% vs. misaligned 71.4%; GLM-4-9b aligned 48.3% vs. misaligned 28.4%; GPT-3.5 aligned 41.9% vs. misaligned 47.7%
Caveat
The alignment classification is determined by the dataset design, not by the model; the paper does not control for the specific nature of the conflict.