Per-Option Scoring Outperforms Joint Encoding for Small Models
October 21, 2026
A study on decision-making encoders shows that per-option scoring outperforms joint scoring by 5.8 points in chance-corrected skill. For small encoders (32M parameters), reading options separately is more efficient for tool selection than reading them jointly.
HOW THIS AFFECTS YOU
●
builderYou can achieve better tool-calling accuracy by using per-option scoring with smaller, faster encoders.
●
researcherThis demonstrates that architectural choices like joint vs. separate reading are critical for low-parameter deciders.