Core Concepts
Large language models can enhance their controllability and reliability by limiting their knowledge scope and refusing to answer questions they are not confident in addressing.
Abstract
The paper proposes a novel system called "Learn to Refuse" (L2R) to address the issue of hallucination in large language models (LLMs) in the context of question-answering.
The key components of L2R are:
Knowledge Scope Limitation: L2R utilizes an independent, limited, and structured knowledge base to represent the knowledge scope of the LLM, rather than relying on the LLM's internal knowledge. This ensures the knowledge is traceable and controllable.
Refusal Mechanism: L2R incorporates a refusal mechanism that allows the LLM to refuse to answer questions it is not confident in addressing, avoiding potential errors or risks.
The paper also introduces an Automatic Knowledge Enrichment (AKE) method to rapidly expand the knowledge base by leveraging the internal knowledge of LLMs.
Experiments on the TruthfulQA, CommonsenseQA, and MedQA datasets demonstrate that L2R outperforms traditional LLM-based question-answering systems in terms of accuracy, while maintaining a high count of answered questions. The refusal mechanism enables L2R to selectively refuse to answer certain questions, leading to more reliable and controllable responses.
Stats
The capital of the United States is Washington, D.C.
DeepMind was founded in 2010.
Quotes
"Our objective is for the LLM to function solely as a machine that processes input and output data and interacts with users using its language processing ability."
"We presume that the LLM does not possess internal knowledge to avoid the influence of incorrect information and unclear expressions."