Reproduce “Extracting Paragraphs from LLM Token Activations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nicholas Pochinkov, Angelo Benoit, Lovkush Agarwal, Zainab Ali Majid, Lucile Ter-Minassian
The study demonstrated that Large Language Models encode substantial information about upcoming paragraphs within the activations of the double newline token. Activation patching and semantic similarity analysis revealed that a single token's activation can effectively reconstruct the semantic content of an entire subsequent paragraph.
Key insights
- Large Language Models (LLMs) exhibit "paragraph-level planning," encoding information about upcoming content and structure within specific token activations.
- Text structure, particularly paragraph boundaries, is consistently encoded in the model's attention patterns across all layers, demonstrating structural awareness.
- The internal representations of LLMs extend beyond token-by-token prediction, approximating human-like organization of thoughts into coherent textual units.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.