MGFiD (Multi-Granularity Guided Fusion-in-Decoder) focuses answer generation on fine-grained evidence instead of passing every retrieved passage to the decoder equally. section 5.6.1
Its processing pipeline is:
- Re-rank the retrieved passages. section 5.6.1
- Select the most important sentence from each top-ranked passage. section 5.6.1
- Aggregate those sentences into an anchor vector. section 5.6.1
- Prune less important passages and provide the decoder with the anchor vector plus the remaining context. section 5.6.1
This multi-granularity design improves exact match on benchmarks while reducing decoder load by concentrating reads on finer-grained evidence. It extends the FiD progression beyond passage-level fusion toward selective, more interpretable context. section 5.6.1
