A rationale signal tells a decoder which retrieved passages contain justification for an answer. Rather than treating every passage equally, it supplies confidence scores that guide attention toward supporting evidence and away from related noise. section 5.6.1
Ten passages, three useful sources
A standard FiD model may receive ten passages when only three contain the answer’s evidence. RFiD trains the model to identify those rationale-bearing passages, then passes their scores to the decoder so generation relies more heavily on them. section 5.6.1
The signal supports two linked goals:
- Faithfulness: ground the generated answer in passages judged to contain its justification. section 5.6.1
- Interpretability: expose which passages the model trusted most. section 5.6.1
- Joint learning: optimize answer generation and rationale classification through a multi-task loss. section 5.6.1
