Interpreting Key Mechanisms of Factual Recall in Transformer-Based Language Models
This work provides interpretability insights for factual recall in LLMs, which is crucial for researchers and practitioners aiming to understand and improve model reliability, though it is incremental in building on existing mechanistic analysis.
The paper investigates how Transformer-based language models perform factual recall tasks, identifying a three-step mechanism involving attention heads and MLPs, and proposes an analytic method to decompose MLP outputs for human understanding. The authors used this interpretation to mitigate an anti-overconfidence mechanism in the final layer, improving factual recall confidence across models like GPT-2, OPT, and Llama-2 in zero- and few-shot setups.
In this paper, we delve into several mechanisms employed by Transformer-based language models (LLMs) for factual recall tasks. We outline a pipeline consisting of three major steps: (1) Given a prompt ``The capital of France is,'' task-specific attention heads extract the topic token, such as ``France,'' from the context and pass it to subsequent MLPs. (2) As attention heads' outputs are aggregated with equal weight and added to the residual stream, the subsequent MLP acts as an ``activation,'' which either erases or amplifies the information originating from individual heads. As a result, the topic token ``France'' stands out in the residual stream. (3) A deep MLP takes ``France'' and generates a component that redirects the residual stream towards the direction of the correct answer, i.e., ``Paris.'' This procedure is akin to applying an implicit function such as ``get\_capital($X$),'' and the argument $X$ is the topic token information passed by attention heads. To achieve the above quantitative and qualitative analysis for MLPs, we proposed a novel analytic method aimed at decomposing the outputs of the MLP into components understandable by humans. Additionally, we observed a universal anti-overconfidence mechanism in the final layer of models, which suppresses correct predictions. We mitigate this suppression by leveraging our interpretation to improve factual recall confidence. The above interpretations are evaluated across diverse tasks spanning various domains of factual knowledge, using various language models from the GPT-2 families, 1.3B OPT, up to 7B Llama-2, and in both zero- and few-shot setups.