Backbone-Adaptive Evidence Routing (BAER) optimizes pairwise LLM judging by adapting evidence mechanisms to specific model backbones. Testing on 8B judge backbones showed accuracy gains of 0.87 to 7.32 points across four benchmarks.
HOW THIS AFFECTS YOU
●
builderYou can implement BAER to improve the reliability and accuracy of LLM-as-a-judge evaluation pipelines.
●
researcherThis method preserves candidate symmetry while allowing for adaptive evidence stacking and routing.