Memory Injections: Correcting Multi-Hop Reasoning Failures during Inference in Transformer-Based Language Models

TL;DR

This paper introduces a method to improve multi-hop reasoning in transformer-based language models by injecting targeted memories into attention layers, significantly enhancing their reasoning accuracy during inference.

Contribution

It presents a novel memory injection technique that corrects reasoning failures in LLMs by adding prompt-specific information at critical attention points.

Findings

Memory injections can increase correct token prediction probability by up to 424%.

Targeted memory injections improve multi-hop reasoning performance.

Analysis of GPT-2 activations guides effective memory placement.

Abstract

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%.