On the Monitorability of Natural Language Autoencoders: Identification vs Discrimination
The paper argues that NLA explanations may point to the right activation without reliably ruling out nearby false accounts.
It frames training and oversight as different referential games: one teaches generation, the other requires discrimination. In the reported checkpoint test, true accounts beat accounts from other activations in all 320 comparisons, but beat fluent single-claim rivals only 60.2% of the time. The author says this is not explained by private shorthand between the verbaliser and reconstructor, because meaning-preserving rewrites barely changed distance. ArXiv · AI/CL/LG's note
It frames training and oversight as different referential games: one teaches generation, the other requires discrimination. In the reported checkpoint test, true accounts beat accounts from other activations in all 320 comparisons, but beat fluent single-claim rivals only 60.2% of the time. The author says this is not explained by private shorthand between the verbaliser and reconstructor, because meaning-preserving rewrites barely changed distance. ArXiv · AI/CL/LG's note
score 4