Tag
This paper introduces YOPO, a method that combines steering probes and sufficiency directions in a single forward pass of frozen language models to improve reasoning accuracy and enable abstention when information is insufficient, demonstrating enhancements across various benchmarks and model scales.