Browse Teach Curate

Dynamic Attention-Guided Context Decoding for Faithful RAG Generation

Research contribution: DAGCD, a single-pass decoding framework that mitigates context-faithfulness hallucinations in retrieval-augmented LLMs by combining the model's own attention distributions (which encode context utilization) with token-level uncertainty signals to reweight the output distribution, improving faithfulness and robustness on open-book QA without the multi-pass cost of prior contrastive-decoding methods.

Data Text

Dynamic Attention-Guided Context Decoding for Mitigating Context Faithfulness Hallucinations in

This NLP research paper (cs.CL) addresses 'context faithfulness hallucinations', where a retrieval-augmented large language model produces answers that deviate from the retrieved context because it f…