UID-Conj. Metric Improves Small Language Model Self-Improvement
October 29, 2026
The UID-Conj. metric uses global entropy variation and local smoothness to provide auxiliary trajectory rewards. On AIME problems, this metric improves online macro-average performance by 2.08 percentage points over standard methods.
HOW THIS AFFECTS YOU
●
researcherThis shows that trajectory-level preferences are more effective than simple answer-level rewards for SLM training.