首页 > AI前沿 > When Objectives and Evaluators Disagree: Evaluating Transfer for Post-hoc Demographic Framing Mitigation

When Objectives and Evaluators Disagree: Evaluating Transfer for Post-hoc Demographic Framing Mitigation

arXiv自然语言 2026-11-20 13:00 27 阅读 查看原文

研究背景与问题

Post-hoc rewriting is used to reduce demographic disparities in LLM-generated text when the source model is inaccessible. A rewrite can look successful merely because it improved its own target or satisfied its selector. We study this as a transfer problem on 660 gender pairs and 1,980 age contrasts of demographic-conditioned climate messages.

研究方法

A rewriter proposes edits, a selection model chooses among constraint-satisfying candidates, and an independent evaluator uninvolved in either rates pairwise framing disparity.

策略效果对比

A fixed policy that always targets the message with the lower Persuasion Bias Index (PBI) transfers inconsistently across axes and outcome aggregations. An adaptive policy that proposes edits on both sides but accepts at most one significantly improves on that rule, applied to the same candidates, in seven of eight comparisons across two evaluator families, mostly by editing more pairs, not by guided direction choice, at higher preservation risk. A policy editing both messages gives the largest point reductions on fewer pairs.

差异削减与评分表现

These two policies reduce disparity across a preservation-threshold sweep; absolute ratings vary across evaluator families and prompts.

缩小PBI差距的影响

Shrinking the PBI gap increases dimension-mean disparity under both evaluator families.

研究建议

Mitigation should be reported as transfer across objectives, evaluators and their prompts, and editing policies.