A multi-document synthesis task requires generating one answer by combining evidence scattered across several documents rather than extracting a single fact. It appears when no passage contains the complete answer, as in summarizing the drivers of Apple's revenue growth from evidence about iPhone sales, Services growth, and logistics. section 5.5
Fusion-in-Decoder (FiD) addresses this setting by independently encoding each passage and giving the decoder access to their separate representations. This can isolate evidence effectively, but synthesis still fails when one passage produces a disproportionately strong signal and dominates generation. section 5.5
Use the task structure to distinguish synthesis from single-fact retrieval:
- Choose multi-document synthesis when the answer must aggregate partial evidence across documents, such as a news or research summary. section 5.6
- Prefer a basic retrieval-augmented generation system for simple, single-fact lookups such as identifying a capital city. section 5.6
- Expect FiD to help with many passages and complex question answering, while recognizing the cost of independently encoding each passage. section 5.6
Apple revenue-growth question
Given passages about strong iPhone sales, 20% Services growth, and resilient logistics, a successful synthesis should integrate all three drivers. General-purpose readers instead returned only “strong iPhone sales,” showing how extraction can replace synthesis when one signal dominates. section 5.5
