Forward KL consistently exhibits higher CE-KL gradient cosine alignment than Reverse KL across pre-training and mid-training, with the gap widening at higher alpha and later training steps.

Forward KL consistently exhibits higher CE-KL gradient cosine alignment than Reverse KL across pre-training and mid-training, with the gap widening at higher alpha and later training steps. · CiteArk

Forward KL consistently exhibits higher CE-KL gradient cosine alignment than Reverse KL across pre-training and mid-training, with the gap widening at higher alpha and later training steps.

报告指标与观测值

Assessment（0）

关联运行（0）