Parallel Key-Value Cache Fusion for Position Invariant RAG
This addresses a critical issue for users of RAG pipelines in LLMs by improving reliability when relevant information is not at the beginning or end of contexts.
The paper tackles the 'Lost in the Middle' problem in Retrieval Augmented Generation (RAG) by introducing a framework that ensures decoder-only models generate consistent outputs regardless of input context order, achieving position invariance and superior robustness to irrelevant passages in open domain question answering tasks.
Recent advancements in Large Language Models (LLMs) underscore the necessity of Retrieval Augmented Generation (RAG) to leverage external information. However, LLMs are sensitive to the position of relevant information within contexts and tend to generate incorrect responses when such information is placed in the middle, known as `Lost in the Middle' phenomenon. In this paper, we introduce a framework that generates consistent outputs for decoder-only models, irrespective of the input context order. Experimental results for three open domain question answering tasks demonstrate position invariance, where the model is not sensitive to input context order, and superior robustness to irrelevent passages compared to prevailing approaches for RAG pipelines.