The latest research in artificial intelligence offers ROME (Rank One Model Editing): an extended language model solution for efficiently locating and modifying factual associations in GPT models

Where are the facts kept in a Great Language Model or LLM?

For two reasons, we are curious about how and where a model retains its factual relationships.

• To understand huge opaque neural networks: The internal computations of large language models are poorly understood. Knowing about huge transformer networks first requires understanding how information is processed.
• Make corrections: Since models are often inaccurate, biased, or private, we want to create techniques that will identify and repair specific factual inaccuracies.

In a recently published article, factual associations within GPT have been shown to correspond to a directly editable localized computation.

Large language transformers, such as autoregressive GPT models (Radford et al., 2019; Brown et al., 2020) and masked BERT (Devlin et al., 2019), have been shown to produce predictions consistent with evidence-based knowledge (Petroni et al., 2019; Jiang et al., 2020; Roberts et al., 2020; Brown et al., 2020). While some factual predictions change when restated, others resist paraphrase, according to Elazar et al. (2021a). For example, GPT will accurately predict the fact: “Seattle” when given a prefix like “The Space Needle is located in the city of”.

In order to identify specific modules within a transformer that mediate memory of a fact about a subject, the researchers in this paper first examined the causal effects of hidden states. They found that when processing the last token of the subject name, the MLP feedforward placed at different intermediate levels is essential.

The researchers found two things for GPT-type transformer designs:

1. When processing the final token of the subject, factual associations can be traced along three dimensions to the parameters of the MLP module at various intermediate layers and in particular.

A few states of the causal trace above contain information that can tip the model from one factual prediction to another. These causal traces were used in their experiments, and they found evidence that knowledge retrieval takes place in MLP modules at the initial site; late-site attention processes then pass the information to the point in the computation where the particular word can be anticipated.

2. Small rank one adjustments within a single MLP module may alter some factual correlations. By assessing generalization to alternative formulations of the same information, they can discern between changes in knowledge and merely superficial linguistic changes.

The team developed an entirely new causal tracing technique to identify key calculations influencing factual recall. When processing a factual message, the approach separates the causal consequences of specific states in the neural network. By following this flow of information, it is possible to identify the modules that mainly contribute to the search for factual associations.

The proposed ROME is made to change specific facts inside a GPT model. ROME views a single module as a key-value store where the key encrypts a subject and the value encrypts the knowledge associated with that subject. The model can thus retrieve factual links by obtaining the value corresponding to the key, allowing associations of particular facts to be modified and updated in a specific and generalized way.

The researchers tested ROME using their CounterFact dataset, which contains thousands of counterfactuals and text allowing quantitative assessment of specificity and generalization while learning a counterfactual, as well as the Zero-Shot Relation task. Extraction (zsRE). On the CounterFact dataset, ROME maintained both specificity and generalization while showing competitive results on zsRE in the evaluations.

They potentially increase the transparency of these systems and decrease the energy required to correct errors by describing the internal structure of large autoregressive transformer language models and creating a fast method for modifying stored knowledge.

---

Check paper, projectand GitHub link. All credit for this research goes to the researchers on this project. Also don’t forget to register. our Reddit page and discord channelwhere we share the latest AI research news, cool AI projects, and more.