Dual-View Training for Instruction-Following Information Retrieval

Researchers proposed a dual-view training method using polarity reversal to generate complementary instructions and teach a retriever to separate topical relevance from instruction compliance.

The approach improved a 305M-parameter encoder by 45% on the FollowIR benchmark.

It goes beyond simple topic matching by helping retrieval models follow user constraints more faithfully.

This points to better search quality and precision for instruction-based queries.