Editing Arbitrary Propositions in LLMs without Subject Labels

TL;DR

This paper introduces a new dataset, Factual Accuracy Classification Test (FACT), which includes non-binary propositions and for which subject labels are not generally applicable, and therefore is beyond the scope of existing L\&E methods, and shows that with the method editing is possible on FACT.

Abstract

Large Language Model (LLM) editing modifies factual information in LLMs. Locate-and-Edit (L\&E) methods accomplish this by finding where relevant information is stored within the neural network, and editing the weights at that location. The goal of editing is to modify the response of an LLM to a proposition independently of its phrasing, while not modifying its response to other related propositions. Existing methods are limited to binary propositions, which represent straightforward binary relations between a subject and an object. Furthermore, existing methods rely on semantic subject labels, which may not be available or even be well-defined in practice. In this paper, we show that both of these issues can be effectively skirted with a simple and fast localization method called Gradient Tracing (GT). This localization method allows editing arbitrary propositions instead of just binary ones, and does so without the need for subject labels. As propositions always have a truth value, our experiments prompt an LLM as a boolean classifier, and edit its T/F response to propositions. Our method applies GT for location tracing, and then edit the model at that location using a mild variant of Rank-One Model Editing (ROME). On datasets of binary propositions derived from the CounterFact dataset, we show that our method -- without access to subject labels -- performs close to state-of-the-art L\&E methods which has access subject labels. We then introduce a new dataset, Factual Accuracy Classification Test (FACT), which includes non-binary propositions and for which subject labels are not generally applicable, and therefore is beyond the scope of existing L\&E methods. Nevertheless, we show that with our method editing is possible on FACT.

Figures (1)

• Figure 1: Mean gradient norm for CFF. The horizontal axis spans through Vicuna-7b's 32 layers. The vertical axis partitions the (non T/F formatting) prompt tokens into buckets: all tokens before the first subject token, all subject tokens except the last, the last subject token, all tokens after last subject token except the last token in proposition, and the last (non T/F formatting) token. For each bucket, we record the maximum gradient norm among the tokens in the bucket. We then average over all prompts, where if a bucket is empty on a particular prompt, we drop that prompt from that bucket's mean calculation (for example, when the subject appears first, there are no tokens before the first subject token).