我们怎么知道，AI是真的“懂了”？

[LG] Latent Concept Disentanglement in Transformer-based Language Models

[Purdue University & University of Southern California]

arxiv.org