Influence-Guided Response Rewriting Study
Rewriting identified training examples produces stronger behavioral changes than reweighting, offering more effective intervention methods for AI safety and alignment.
Rewriting identified training examples produces stronger behavioral changes than reweighting, offering more effective intervention methods for AI safety and alignment.
Clairy 让 AI 决策过程透明化,解决了模型输出错误时只能靠猜的痛点,做 AI 安全、审计或模型调试的团队值得关注,可以直接用它来诊断和修正模型偏见。
STRIDE 解决了 LLM 训练数据归因计算成本高、依赖局部近似的痛点,做模型可解释性、数据质量分析的团队可以直接用这个新框架。