Poster
in
Workshop: Foundation Model Interventions
GPT-2 Small Fine-Tuned on Logical Reasoning Summarizes Information on Punctuation Tokens
Sonakshi Chauhan · Atticus Geiger
Keywords: [ Reasoning ] [ Interpretability ] [ Causality ]
How is information stored and aggregated within a language model performing inference? Preliminary evidence suggests that representations of punctuation tokens might serve as summary points'' for information about preceding text. We add to this body of evidence by demonstrating that GPT-2 small fine-tuned on the RuleTaker logical inference dataset aggregates crucial information about rules and sentences above period tokens.