Memory Injections: Correcting Multi-Hop Reasoning Failures during Inference in Transformer-Based Language Models
This addresses inconsistent reasoning in large language models for tasks requiring information synthesis, offering a targeted correction method that is incremental in nature.
The paper tackles the problem of multi-hop reasoning failures in transformer-based language models by proposing targeted memory injections into attention heads during inference, resulting in up to a 424% increase in the probability of desired next tokens.
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%.