Extracting Paragraphs from LLM Token Activations
This provides insights into LLMs' contextual understanding for researchers, but it is incremental as it builds on existing token-level analysis.
The study tackled the problem of understanding how LLMs plan paragraph content by analyzing single-token activations, specifically the double newline token, and found that patching these activations can transfer significant contextual information about the following paragraph.
Generative large language models (LLMs) excel in natural language processing tasks, yet their inner workings remain underexplored beyond token-level predictions. This study investigates the degree to which these models decide the content of a paragraph at its onset, shedding light on their contextual understanding. By examining the information encoded in single-token activations, specifically the "\textbackslash n\textbackslash n" double newline token, we demonstrate that patching these activations can transfer significant information about the context of the following paragraph, providing further insights into the model's capacity to plan ahead.