Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
1
0
2
4
Ablating just 50 retrieval-specific heads identified by LOCOS can reduce model performance from a ROUGE-L score of 0.401 to virtually zero, highlighting their critical role in non-literal answer synthesis.