List your company
ResearchFormerly WizWand

Math & Code Reasoning on SciQ

66.5ScoreDiReCT
42.16448.48254.861.118May 29, 2026
Updated 5mo ago

Evaluation results

MethodLinks
DiReCTBackbone=Llama-1.1B2026.0566.5
InfoBatchBackbone=Llama-1.1B2026.0565.6
Perplexity-basedBackbone=Llama-1.1B2026.0563.1
GradNorm (IS)Backbone=Llama-1.1B2026.0562.8
Uniform SamplingBackbone=Llama-1.1B2026.0561.2
Loss-basedBackbone=Llama-1.1B2026.0560.5
DiReCTBackbone=GPT-2-Medium (355M)2026.0548.5
InfoBatchBackbone=GPT-2-Medium (355M)2026.0546
GradNorm (IS)Backbone=GPT-2-Medium (355M)2026.0545.2
Perplexity-basedBackbone=GPT-2-Medium (355M)2026.0544.6
Loss-basedBackbone=GPT-2-Medium (355M)2026.0543.7
Uniform SamplingBackbone=GPT-2-Medium (355M)2026.0543.1