EARAL: Efficient Attribution of Retrieval-Augmented Answer Generation using Logits
Abstract
AbstractLarge Language Models (LLMs) generate grounded answers using Retrieval-Augmented Generation (RAG), reducing hallucinations. Prior work has focused on answer and attribution correctness, with recent studies exploring faithfulness through internal signals that reflect the model’s decision process. However, these approaches often introduce latency and do not tightly align token generation with attribution. We propose EARAL, a novel attribution method for RAG that improves faithfulness through joint answer generation and generation-time attribution with minimal overhead. Documents are marked with special token identifiers, whose logits estimate their contribution during generation and are aggregated into document-level attributions. On the Trust-Align benchmark for trustworthy attributed generation, EARAL outperforms state-of-the-art methods across metrics. Additional analysis shows low latency, robustness, and strong alignment with faithfulness.