value-consistency

#value-consistency

VALCON: Benchmarking and Enhancing Value Consistency in Large Language Models via Dual-Role Variation

arXiv cs.CL ↗ · 2d ago Cached

该论文提出VALCON框架，通过生成者/评判者双重角色变体来基准测试并提升大语言模型的价值一致性，实验揭示28个LLM普遍存在跨角色不一致问题，并提出基于表征的干预方法加以缓解。

0 favorites 0 likes