Localizing knowledge in large language model representations

Large language models (LLMs) are very proficient in NLP tasks. In the first part of this work, we evaluate the performance of LLMs on the task of finding the locations of characters inside a long narrative. The objective of the task is to generate the correct answer when the input is a piece of a narrative followed by a question asking the location of a character. For the evaluation of the task, we generate two new datasets by annotating the characters and their locations in the narratives: Andersen and Persuasion. We show that the LLM performance is not satisfactory on these datasets when compared to the simple baseline we designed that does not use machine learning. We also experiment with in-context learning to improve the performance and report results. Moreover, we address the problem that the LLMs are limited by the bounded context length. We hypothesize that if we localize the character-location relation information among the activations inside an LLM, we can store those activations and inject them into other models that are run with a different prompt so that the LLM can answer the questions about the information that was carried from another prompt, even though the character and location relation is not mentioned explicitly in the current prompt. We develop five different techniques to localize the character-location relation information occurring in the LLMs: Moving and adding LLM activations to other prompts, adding noise to LLM activations, checking cosine similarity between LLM activations, editing LLM activations, and visualizing attention scores during answer generation. We report the observations we made using these techniques.