List your company

Mathematical Reasoning on GSM8K (Drop in Utility (%))

-7.8Drop in UtilityLatent Similarity
-9.3040.8481121.152Apr 1, 2026

Evaluation results

MethodLinks
Latent SimilaritySelection Criterion=Latent Similarity, Layer=Best Layer2026.04-7.8
Self CertaintySelection Criterion=Self Certainty2026.04-6
PerplexitySelection Criterion=Perplexity2026.04-5.1
RandomSelection Criterion=Random2026.041.6
Latent SimilaritySelection Criterion=Latent Similarity, Layer=122026.047.6
KLSelection Criterion=KL2026.0429.8