RA-DIT (Retrieval-Augmented Dual Instruction Tuning) retrofits cooperation between an off-the-shelf retriever and language model through two lightweight fine-tuning passes, avoiding Atlas-style joint pre-training. section 6.6 section 6.6.1
The method separates the two training jobs:
- Fine-tune only the query encoder with LM-Supervised Retrieval. A KL-divergence loss matches retriever scores to a distribution derived from each chunk’s improvement to the language model’s probability of the correct output. section 6.6.1
- Freeze the document encoder so the existing index remains usable and never needs re-embedding. section 6.6.1
- Fine-tune the language model on instruction examples with retrieved chunks prepended as background, using the ordinary next-token objective. section 6.6.1
- Teach the reader to rely on retrieved passages rather than treating them as irrelevant prompt text. section 6.6.1
A miniature reader pass
Six training questions receive contexts where a distractor sentence precedes the answer passage. Fifteen optimization steps move two held-out answers from log-probabilities of -3.2 and -0.4 to approximately 0.0 and 0.0, showing the reader learning to use the supplied context. section 6.6.1
RA-DIT assumes access to both components for fine-tuning. When the language model is closed behind an API, REPLUG preserves the likelihood-based retrieval supervision but trains only the retriever against the frozen model. section 6.6.2
