Retrieval In Decoder benefits generative models for explainable complex question answering.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Large-scale Language Models (LLMs) show promise but suffer from factual hallucinations.
• Current retrieval-augmented methods have limitations due to separate retriever and generator components.
• Supervised training restricts generator capabilities in existing retrieval-augmented approaches.

Purpose of the Study:

• To propose an unsupervised framework, Retrieval In Decoder (RID), for multi-granularity decoding.
• To integrate retrieval directly into the decoding process of generative models.
• To enhance adaptive explanation generation for Small-scale Language Models (SLMs) using reinforcement learning.

Main Methods:

• Developed the unsupervised Retrieval In Decoder (RID) framework.
• Implemented dynamic adjustment of decoding granularity based on retrieval outcomes.
• Introduced reinforcement learning-driven knowledge distillation for adaptive explanation generation.

Main Results:

• RID framework demonstrated superior performance across six public benchmarks.
• Outperformed popular LLMs and existing retrieval-augmented methods.
• Validated effectiveness and scalability for both LLMs and SLMs.

Conclusions:

• RID effectively mitigates factual hallucinations in generative models.
• The framework shows significant improvements in model performance and applicability.
• RID offers a scalable and effective solution for enhancing language model generation.