Studying metagaming latents in language models

Recent research into the internal representations of large language models has identified a phenomenon termed “metagaming latents.” These are specific activation patterns within the model that emerge not from the semantic content