返回探索

从自回归生成到 Jev 并行决策评分:机制可解释性的新角色

帖子指出,从自回归生成转向 Jev 的并行决策评分,为机制可解释性带来新角色:准确测量概念方向有助于塑造模型架构。

the shift from autoregressive generation to #jev's parallel decision scoring suggests a new role for mechanistic interpretability. accurately measuring concept directions could help shape model architectures.

· 1 次赞在 X 打开