Mansi Sakarvadia

Answering multi-hop reasoning questions requires retrieving and synthesizing information from diverse sources. Large Language Models (LLMs) struggle to perform such reasoning consistently. Here we propose an approach to pinpoint and rectify multi-hop reasoning failures through targeted memory injections on LLM attention heads. First, we analyze the per-layer activations of GPT-2 models in response to single and multi-hop prompts. We then propose a mechanism that allows users to inject pertinent prompt-specific information, which we refer to as "memories," at critical LLM locations during inference. By thus enabling the LLM to incorporate additional relevant information during inference, we enhance the quality of multi-hop prompt completions. We show empirically that a simple, efficient, and targeted memory injection into a key attention layer can often increase the probability of the desired next token in multi-hop tasks, by up to 424%.

Transformer-based Large Language Models (LLMs) are the state-of-the-art for natural language tasks. Much recent work has attempted to decode the internal mechanisms by which LLMs arrive at their final predictions for text completion tasks, including by reverse-engineering the role of linear layers. Yet little is known about the role of attention heads in producing the final token prediction. We propose the Attention Lens, a tool that enables researchers to translate the outputs of attention heads into vocabulary tokens via learned attention head-specific transformations called lenses. Preliminary findings from our trained lenses indicate that attention heads play highly specialized and specific roles in language models.