RFiD is a Fusion-in-Decoder architecture for noisy multi-document question answering. When ten retrieved passages contain only three pieces of actual evidence, it learns to identify and prioritize the passages that justify the answer. section 5.6.1
Unlike standard FiD, which may treat related but unsupported passages as useful context, RFiD assigns rationale confidence scores and passes them to the decoder as an explicit signal. A multi-task loss trains answer generation and rationale classification together. section 5.6.1
RFiD is useful when retrieval produces plausible noise alongside supporting evidence:
- Weights high-rationale passages more heavily during answer generation. section 5.6.1
- Improves answer faithfulness while exposing which passages the model trusted as supporting evidence. section 5.6.1
