Recent advances in large language models (LLMs) like ChatGPT have demonstrated impressive capabilities in generating human-like text. However, a key limitation of current LLMs is their tendency to make factual errors in the absence of external knowledge.
To address this, researchers from the University of Washington, the Allen Institute for AI and IBM Research AI have proposed a new framework called Self-Reflective Retrieval-Augmented Generation (Self-RAG). The key idea is to train the AI system to supplement its own knowledge by retrieving relevant information from documents, and then reflect on its generations to ensure consistency with the retrieved facts.
The Self-RAG system consists of three main components:
- Retrieval module: retrieves relevant passages from a document collection on demand based on the input prompt and current generation context.
- Generator module: a standard LLM that is trained to concurrently process multiple retrieved passages and generate text augmented with special “reflection tokens.”
- Reflection tokens: are special tokens indicating whether retrieval is needed, if a passage is relevant, whether the generation is supported by the evidence, and an overall utility score.
By generating these reflective tokens, the system can criticize its own outputs and select the best response grounded in the retrieved evidence. The researchers trained Self-RAG models with 7B and 13B parameters on a diverse dataset of 150k instruction-response pairs.
In evaluations across question answering, reasoning, and long-form generation tasks, Self-RAG significantly outperformed both standard LLMs like GPT-3, and retrieval-augmented methods like RAG and REALM.
The reflective tokens also make SELF-RAG highly customizable during inference. By tuning the weights given to different critique aspects like relevance vs. evidence support, users can control the retrieval frequency and trade-offs between completeness vs. factual precision.
Overall, Self-RAG demonstrates that augmenting LLMs with the abilities to retrieve knowledge and self-reflect can greatly enhance their factuality without sacrificing too much fluency. If these improvements can be scaled up to models with hundreds of billions of parameters, it could be an important step towards more robust and trustworthy AI systems.
Human evaluations also reveal that Self-RAG’s self-critiques are well-aligned with human judgments. The researchers plan to further expand the capabilities of Self-RAG, and also study potential negative societal impacts of factually-ground AI systems.