Fusion-in-Decoder
An open-domain question answering architecture proposed by Izacard and Grave in 2021. It independently encodes multiple retrieved text passages, each concatenated with the question, then fuses all passage encodings and feeds them into a single decoder to generate an answer. The architecture is designed for structured retrieval-augmented generation pipelines.
Timeline (last 90 days)
For a published 406M-parameter Fusion-in-Decoder specialized model, the ROUGE-1 score dropped sharply by 6.30 points when the input was switched from truncated long text to 2,000-word retrieved segments.
Fusion-in-Decoder (FiD) was proposed by Izacard and Grave in 2021 as an open-domain question answering architecture that independently encodes each retrieved segment concatenated with the question, then concatenates the encoded representations and feeds them into a unified decoder to generate the answer.
All Facts (2)
For a published 406M-parameter Fusion-in-Decoder specialized model, the ROUGE-1 score dropped sharply by 6.30 points when the input was switched from truncated long text to 2,000-word retrieved segments.
50%UnverifiedFusion-in-Decoder (FiD) was proposed by Izacard and Grave in 2021 as an open-domain question answering architecture that independently encodes each retrieved segment concatenated with the question, then concatenates the encoded representations and feeds them into a unified decoder to generate the answer.
50%