Failure of contextual invariance in gender inference with large language models

1. 1Evaluate large language models on controlled pronoun selection tasks to assess contextual invariance
2. 2Analyze model outputs for systematic shifts induced by minimal discourse context changes
3. 3Investigate correlations between model outputs and cultural gender stereotypes

AI engineers and ML researchers benefit from understanding the limitations of large language models in gender inference tasks, as it affects the development of fair and unbiased AI systems

💡 Large language models' outputs are not stable under contextually equivalent formulations of a task, which can perpetuate cultural gender stereotypes

💡 LLMs' gender inference outputs can shift significantly with minimal context changes