When to Retrieve: Teaching LLMs to Utilize Information Retrieval Effectively

Overview

• This paper explores how Large Language Models (LLMs) can effectively use an off-the-shelf information retrieval (IR) system when additional context is required to answer a given question.
• The researchers propose a tailored training approach for LLMs, leveraging existing open-domain question answering datasets, where LLMs are trained to generate a special token when they do not know the answer to a question.
• The evaluation of the Adaptive Retrieval LLM (Adapt-LLM) on the PopQA dataset shows improvements over the same LLM under different configurations, demonstrating Adapt-LLM's ability to determine when to rely on its parametric memory or when to use an IR system.

Plain English Explanation

Large Language Models (LLMs) are powerful AI systems that can understand and generate human-like text. However, there are times when LLMs may not have all the information needed to answer a specific question. In these cases, the LLM could benefit from accessing additional information from an external source, such as an information retrieval (IR) system.

The researchers in this paper have developed a new approach to train LLMs to recognize when they don't know the answer to a question and should use an IR system to find the necessary information. They do this by training the LLM to generate a special token when it doesn't know the answer, rather than just guessing or making something up.

When the LLM generates this special token, it's a signal that the system should then use an IR tool to search for and retrieve the relevant information to answer the question. This adaptive approach allows the LLM to leverage its own knowledge when possible, but also know when to seek out additional information to provide a more accurate and complete answer.

The researchers tested this Adaptive Retrieval LLM (Adapt-LLM) on a dataset called PopQA, which contains questions of varying popularity. Their results showed that Adapt-LLM was able to achieve high accuracy when relying on its own knowledge, while also effectively identifying when it needed to use an IR system to find the answer. This suggests that this approach could be a useful way to improve the performance of LLMs on open-domain question answering tasks.

Technical Explanation

The key idea behind this research is that the optimal strategy for question answering does not always involve external information retrieval. In many cases, the LLM's own parametric memory can be leveraged to effectively answer questions, as demonstrated in the PopQA dataset. However, for less popular questions, the LLM may need to utilize an IR system to retrieve additional context.

To address this, the researchers propose a tailored training approach for LLMs. They leverage existing open-domain question answering datasets to train the LLMs to generate a special token, <UNKNOWN>, when they do not know the answer to a question. This signal indicates that the system should then use an IR tool to find the relevant information.

The researchers evaluate their Adaptive Retrieval LLM (Adapt-LLM) on the PopQA dataset, comparing its performance to the same LLM under three different configurations: (i) retrieving information for all the questions, (ii) always using the LLM's parametric memory, and (iii) using a popularity threshold to decide when to use a retriever. The results show that Adapt-LLM is able to generate the <UNKNOWN> token when it determines that it does not know the answer, while achieving high accuracy when relying on its own parametric memory.

This adaptive approach addresses the limitations of previous work that either always used an IR system or always relied on the LLM's own knowledge. By training the LLM to recognize when it needs additional information, the Adapt-LLM can leverage the strengths of both the LLM's parametric memory and the IR system, as demonstrated in the Generative Information Retrieval Evaluation and Optimization Methods for Personalizing Large Language Models Through approaches.

Critical Analysis

The paper presents a promising approach to improving the performance of LLMs on open-domain question answering tasks. By training the LLM to recognize when it doesn't know the answer and needs to use an IR system, the researchers have addressed a key limitation of previous work that relied solely on either the LLM's memory or external information retrieval.

However, the paper does not discuss the potential challenges or limitations of this approach. For example, it's unclear how well the Adapt-LLM would perform on questions that require complex reasoning or integration of information from multiple sources. Additionally, the paper does not explore the impact of the quality and coverage of the underlying IR system on the overall performance of the Adapt-LLM.

Further research could investigate the generalizability of this approach to different types of questions and datasets, as well as the impact of the IR system's capabilities on the Adapt-LLM's performance. Exploring ways to enhance question answering using enterprise knowledge bases could also be a promising direction.

Conclusion

This paper presents an innovative approach to training Large Language Models (LLMs) to effectively utilize an information retrieval (IR) system when additional context is required to answer a given question. By training the LLM to generate a special token when it does not know the answer, the Adaptive Retrieval LLM (Adapt-LLM) can leverage the strengths of both the LLM's parametric memory and the IR system, leading to improved performance on open-domain question answering tasks.

While the paper provides a promising solution, further research is needed to address potential challenges and limitations, such as the impact of the IR system's capabilities and the generalizability of the approach to different types of questions and datasets. Nonetheless, this work represents an important step towards developing more robust and reliable language models that can seamlessly integrate external information to provide accurate and comprehensive answers.