Per-Option Scoring Beats Joint Scoring for Small Decision Encoders
October 23, 2026
A 32M parameter encoder using per-option scoring outperforms joint scoring by 5.8 points of chance-corrected skill on a 37-benchmark decision suite. While per-option scoring requires 5 to 13 times more encoder passes, it maintains manageable latency and higher accuracy for closed-set tasks like tool selection.
HOW THIS AFFECTS YOU
●
builderYou can improve tool-selection accuracy by implementing per-option scoring with small, specialized encoders.
●
researcherThe advantage of per-option scoring appears to depend on reranker initialization rather than the raw encoder alone.