Inspecting and Editing Knowledge Representations in Language Models
This work addresses the need for fine-grained inspection and control of knowledge in language models, which is crucial for improving reliability and interpretability in AI applications, though it is incremental in linking existing techniques.
The authors tackled the problem of understanding and modifying knowledge representations in language models by developing REMEDI, a method that maps natural language statements to fact encodings, enabling editing of LM outputs to align with new facts and probing to predict conflicts with background knowledge.
Neural language models (LMs) represent facts about the world described by text. Sometimes these facts derive from training data (in most LMs, a representation of the word "banana" encodes the fact that bananas are fruits). Sometimes facts derive from input text itself (a representation of the sentence "I poured out the bottle" encodes the fact that the bottle became empty). We describe REMEDI, a method for learning to map statements in natural language to fact encodings in an LM's internal representation system. REMEDI encodings can be used as knowledge editors: when added to LM hidden representations, they modify downstream generation to be consistent with new facts. REMEDI encodings may also be used as probes: when compared to LM representations, they reveal which properties LMs already attribute to mentioned entities, in some cases making it possible to predict when LMs will generate outputs that conflict with background knowledge or input text. REMEDI thus links work on probing, prompting, and LM editing, and offers steps toward general tools for fine-grained inspection and control of knowledge in LMs.