A new prompting technique can measurably reduce gender bias in large language models without touching their weights.
Researchers built DR.GAP (Decoupled Reasoning for Gender-Aware Prompting), an automated pipeline that generates gender-neutral reasoning traces and feeds them back into a model as in-context examples during inference. The goal is to separate gender information from the task itself, since earlier fixes tend to overcorrect: steering prompts push models to fixate on gender cues, while reasoning-based prompts end up baking bias into the chain of thought. The team tested DR.GAP on coreference-resolution and question-answering tasks across six different LLMs and report it held up across models without degrading performance. They also extended the method to vision-language models, where it cut bias substantially as well.
Most bias fixes require retraining or fine-tuning, which costs time and money and often has to be redone per model. A prompting-only approach that works across six different LLMs without parameter changes is a more practical fix, if it holds up beyond the paper's own benchmarks. That distinction matters most for coreference and QA tasks, where gendered pronoun assumptions (the doctor is a he, the nurse is a she) are a well-documented and stubborn failure mode.
Prompt-level fixes are cheap and easy to swap in, which is also why they are fragile: change the task template or the demonstration set, and there is no guarantee the bias does not creep back.