Researchers develop method enabling LLMs to answer questions more concisely and accurately

Large language models (LLMs) are machine-learning models designed to understand and generate human language. State-of-the-art LLMs have demonstrated outstanding potential in open-domain question answering (ODQA), where the model is tasked with providing answers to factual questions.

This is particularly useful in fields such as finance, health care, and education. However, LLMs typically rely on their pre-trained knowledge to answer questions that can become outdated in a constantly changing world.

This limitation can be addressed by using Retrieval-Augmented Generation (RAG) with a pre-trained LLM. In this approach, the question is augmented with documents from a knowledge base. Despite these advancements, LLMs often produce lengthy responses, providing contextual information that can make it difficult and time-consuming to identify the exact answer phrase.

Another important aspect of LLMs is their ability to produce confidence scores, which reflect how certain the model is about the correctness of its answer. These scores are especially crucial in high-risk fields such as finance, law, and health care. Although LLMs can generate sequence probabilities for a specific response, this probability is often unreliable in terms of calibration.

This means the predicted confidence may not accurately correlate with the probability of correctness and should not be used as a confidence score. The inability to identify the exact answer phrase and produce a reliable confidence score limits the practical application of LLMs.

To address these limitations, a team of researchers from the Japan Advanced Institute of Science and Technology, led by Professor Nguyen Le Minh and including doctoral students Nguyen-Khang Le and Dieu-Hien Nguyen, introduced a novel method called Answer-prefix Generation (ANSPRE).

“ANSPRE can improve the generation quality of LLMs, allow them to output the exact answer phrase, and produce reliable confidence scores. Additionally, it can be incorporated into any LLM and complex architecture,” says Prof. Nguyen. Their study will be presented at ECAI-2024, the 27th European Conference on Artificial Intelligence, held on October 19–24 in Santiago de Compostela, Spain.

The main idea of ANSPRE is to add a sequence of text to the LLM prompt that leads to the answer phrase. This sequence of text is called the “answer prefix.”

Prof. Nguyen explains, “Consider the example question, ‘What gambling game, requiring two coins to play, was popular in World War I?’ An answer prefix for this question could be, ‘The gambling game requiring two coins to play that was popular in World War I was ___.’ As most LLMs are trained with causal language modeling, using the answer prefix would allow the LLM to generate the exact answer phrase in place of the blank.”

Given a question, ANSPRE first generates an answer prefix using selected few-shot examples. The researchers demonstrated that only a few handcrafted examples were sufficient to generate a high-quality answer prefix. ANSPRE then uses an existing retriever to gather relevant documents from the knowledge base, similar to RAG.

It combines the document, the question, and the answer prefix, and prompts the LLM to generate the answer phrase. Finally, ANSPRE aggregates the answer phrases and confidence scores across different documents used to answer the question, to produce the final answer.

The researchers demonstrated ANSPRE’s versatility by constructing Self-Reflective Answer-Prefix Generation (SELF-ANSPRE), which combines ANSPRE with Self-Reflective RAG (SEFT-RAG).

SEFT-RAG improves LLM generation by introducing reflection tokens to decide when and what to retrieve from the knowledge base and rank the responses based on the utility of the documents and the answer. In SELF-ANSPRE, the confidence scores from ANSPRE and scores from reflection tokens are combined to generate the final ranking score.

The researchers tested ANSPRE on three ODQA benchmarks and various LLM architectures. The results showed that ANSPRE significantly improves pre-trained and instruction-tuned LLMS, producing high-quality answers and confidence scores that strongly correlate with correctness. Moreover, SELF-ANSPRE significantly enhanced SEFT-RAG. Their analysis also highlighted the importance of each ANSPRE component.

“Our method can lead to more concise and accurate question answering in critical fields like medical diagnosis, legal assistance, and education, and improve customer support. Furthermore, in the long term, our research could foster widespread human-artificial intelligence collaboration by increasing trust in AI systems,” says Prof. Nguyen.

Overall, this innovative method marks a significant step forward for LLMs and can lead to their broader application, even in sensitive domains.