List your company
ResearchFormerly WizWand

Multilingual Mathematical Reasoning on MGSM Thai (test)

41.6AccuracyOne Step + KMM
30.36833.28436.239.116May 15, 2026
Updated 5mo ago

Evaluation results

MethodLinks
One Step + KMMk (auxiliary subset size)=4, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0541.6—
TV + KMMk (auxiliary subset size)=1, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0540.4—
TVk (auxiliary subset size)=2, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0539.6—
One Stepk (auxiliary subset size)=3, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0539.2—
Randomk (auxiliary subset size)=4, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0538.4—
TV + KMMk (auxiliary subset size)=2, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0538—
One Step + KMMk (auxiliary subset size)=1, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0537.6—
One Step + KMMk (auxiliary subset size)=2, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0537.6—
TVk (auxiliary subset size)=3, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0537.6—
One Step + KMMk (auxiliary subset size)=5, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0537.2—
TV + KMMk (auxiliary subset size)=3, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0537.2—
TV + KMMk (auxiliary subset size)=4, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0537.2—
TVk (auxiliary subset size)=1, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0536.8—
Randomk (auxiliary subset size)=5, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0536.8—
One Stepk (auxiliary subset size)=1, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0535.6—
One Step + KMMk (auxiliary subset size)=3, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0535.6—
One Stepk (auxiliary subset size)=4, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0534.8—
TVk (auxiliary subset size)=5, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0534.8—
Randomk (auxiliary subset size)=2, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0534.8—
TVk (auxiliary subset size)=4, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0534.4—
TV + KMMk (auxiliary subset size)=5, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0534.4—
One Stepk (auxiliary subset size)=2, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0532.8—
Randomk (auxiliary subset size)=3, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0532—
One Stepk (auxiliary subset size)=5, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0531.2—
Randomk (auxiliary subset size)=1, Pre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.0530.8—
One StepPre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.05—6
One Step + KMMPre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.05—14
RandomPre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.05—7
TVPre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.05—11
TV + KMMPre-trained model=Qwen2.5-1.5B-Instruct, Algorithm=GRPO2026.05—12