Can Weight-Space Programming Change a Model's Learned Ethical Priorities?
Can Weight-Space Programming Alter Learned Ethical Priorities?
The authors study whether weight editing can change a model's trade-off between keeping promises and benefiting more people. In a text model with concept supervision, the authors adjust the fitted choice weights so that the benefit required to overturn a formal promise is higher than that required to overturn an ordinary promise. On six independently trained text readers, this edit, under two orderings, selects the side with more people for target ordinary conflicts and target formal conflicts at 100% and 83.3% respectively, while preserving the source model's success rate on smaller conflicts and its factual accuracy; the remaining formal-conflict failures coincide with errors in reading the beneficiary number. In a pilot on YOLO11n on COCO128, bias editing raised small-object matching from 3/43 to 9/43, and an equivalent score calibration reproduced the retained detection results. The authors say these exploratory results demonstrate one path to testing selective changes to learned priorities in models that expose the trade-off mechanism, and whether the method can find such changes in less structured language models remains an open question.