Here are the key points about SELF-RAG:
- SELF-RAG is a framework to enhance large language models (LLMs) through retrieval and self-reflection. It trains an LM to learn to retrieve relevant passages on demand, generate text segments, and critique its own generations using special tokens called reflection tokens.
- Reflection tokens indicate whether retrieval is needed, if a retrieved passage is relevant, whether the generated text is supported by evidence, and the overall quality/usefulness of the generation.
- At inference time, SELF-RAG first decides if retrieval is needed. If so, it retrieves passages, generates text segments in parallel while processing passages, and critiques each segment using reflection tokens. It can rank outputs using the critique tokens.
- Reflection tokens enable customization of the model’s behavior at test time to meet different requirements (e.g. emphasizing factual accuracy vs creativity).
- SELF-RAG is trained on an LM objective over text augmented with reflection tokens predicted by a critic model. The critic is trained on demonstrations from GPT-4.
- Experiments on 6 tasks show SELF-RAG outperforms LLMs and retrieval-augmented models like ChatGPT in accuracy, factuality, citation precision, etc. The self-reflection mechanism is shown to be critical for gains.
In summary, SELF-RAG improves LLMs via on-demand retrieval and self-critiquing generations using reflection tokens, enabling customization and significantly outperforming existing models. The key innovation is training the LM to self-reflect.
