Dynamic Weight Grafting: Localizing Finetuned Factual Knowledge in Transformers

TL;DR

This paper introduces dynamic weight grafting, a new interpretability method that localizes where and how finetuned factual knowledge is stored and retrieved in transformer models, revealing multiple pathways for knowledge access.

Contribution

It proposes dynamic weight grafting to analyze finetuned knowledge localization, identifying enrichment and recall pathways and their mechanisms within transformer models.

Findings

Knowledge is stored via enrichment and recall pathways.

Recall involves task-specific attention and entity extraction.

Model components responsible for knowledge retrieval are localized.

Abstract

When an LLM learns a new fact during finetuning (e.g., new movie releases, newly elected pope, etc.), where does this information go? Are entities enriched with relation information immediately, or do models recall information just-in-time before a prediction? Or, are "all of the above" true, with LLMs implementing multiple redundant heuristics? Existing localization approaches (e.g., activation patching) are ill-suited for this analysis because they usually replace parts of the residual stream, thus overriding previous information. To fill this interpretability gap, we propose dynamic weight grafting, an analysis technique that selectively grafts subsets of weights from a finetuned model onto a pretrained model. Using this technique, we show two separate pathways for retrieving finetuned relation information: 1) "enriching" the residual stream with relation information while processing the tokens that correspond to an entity (e.g., "Zendaya" in "Zendaya co-starred with Timoth\'ee Chalamet" and 2) "recalling" this information at the final token position before generating a target fact. In some cases, models need information from both of these pathways to correctly generate finetuned facts while, in other cases, either the "enrichment" or "recall" pathway alone is sufficient. We localize the "recall" pathway to model components -- finding that "recall" occurs via both task-specific attention mechanisms and an entity-specific extraction step in the feedforward networks of the final layers before prediction. By targeting model components and parameters, as opposed to just activations, we are able to understand the mechanisms by which finetuned knowledge is retrieved during generation.