Core Concepts
Retrieval-augmented language models (RAG) can reduce generation risks with certified guarantees.
Abstract
Abstract:
Large language models (LLMs) face trustworthiness issues.
Retrieval-augmented language models (RAG) aim to enhance credibility.
C-RAG framework certifies generation risks for RAG models.
Introduction:
LLMs exhibit emergent abilities but suffer from unreliable generations.
RAG models aim to mitigate generation risks through in-context learning.
Theoretical analysis of generation risks for RAG models is explored.
Conformal Generation Risks of RAG Models:
Constrained generation protocol for RAG models proposed.
Conformal risk analysis for RAG models under the protocol.
Certifies generation risks based on test statistics from calibration samples.
Theoretical Analysis of C-RAG:
RAG achieves lower conformal generation risk compared to LLMs.
Benefits correlated with quality of retrieval model and transformer.
Theoretical results presented in a certification framework.
Out-of-Distribution Test Samples:
Conformal risk analysis under distribution shifts for general bounded risk functions.
Evaluation:
C-RAG evaluated on four NLP datasets with different retrieval models.
Conformal generation risks validated and compared across models.
Soundness and tightness of conformal generation risks under distribution shifts demonstrated.
Stats
RAG achieves a lower conformal generation risk than a single LLM when the quality of the retrieval model and transformer is non-trivial.