List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME26 (pass@1)

40Pass@1Qwen3-4B +CAST
-1.69.22030.8May 12, 2026May 14, 2026May 17, 2026May 20, 2026May 23, 2026May 26, 2026May 29, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Qwen3-4B +CASTModel Scale=4B, Method=+CAST2026.0540
Baseline ModelRollout Length=16K tokens2026.0539.6
Qwen3-8B +CASTModel Scale=8B, Method=+CAST2026.0536.7
Qwen3-4B BaseModel Scale=4B, Method=Base2026.0523.3
Qwen3-4B +GRPOModel Scale=4B, Method=+GRPO2026.0523.3
Qwen3-4B +RLSDModel Scale=4B, Method=+RLSD2026.0523.3
Qwen3-4B +RLRTModel Scale=4B, Method=+RLRT2026.0520
Qwen3-8B +RLSDModel Scale=8B, Method=+RLSD2026.0520
FP8 Rollout P3ORollout Length=16K tokens, Training Iteration=15, Rollout Precision=FP8, Training Precision=BF162026.0517.9
FP8 Rollout P3ORollout Length=16K tokens, Training Iteration=30, Rollout Precision=FP8, Training Precision=BF162026.0517.5
Qwen3-4B +GRPO+OPSDModel Scale=4B, Method=+GRPO+OPSD2026.0516.7
Qwen3-8B +GRPOModel Scale=8B, Method=+GRPO2026.0516.7
P3ORollout Length=4K tokens, Clipping Parameter (epsilon)=∈ {0.2, 0.4, 0.6}2026.0516
FP8 Rollout GRPORollout Length=16K tokens, Training Iteration=15, Rollout Precision=FP8, Training Precision=BF162026.0516
Qwen3-8B BaseModel Scale=8B, Method=Base2026.0513.3
Qwen3-8B +OPSDModel Scale=8B, Method=+OPSD2026.0513.3
Qwen3-8B +GRPO+OPSDModel Scale=8B, Method=+GRPO+OPSD2026.0513.3
Qwen3-8B +RLRTModel Scale=8B, Method=+RLRT2026.0513.3
GRPO (clip avg)Rollout Length=4K tokens, Clipping Parameter (epsilon)=∈ {0.2, 0.4, 0.6}2026.0512.6
Qwen3-1.7B +GRPO+OPSDModel Scale=1.7B, Method=+GRPO+OPSD2026.0510
Qwen3-1.7B +CASTModel Scale=1.7B, Method=+CAST2026.0510
Qwen3-4B +OPSDModel Scale=4B, Method=+OPSD2026.0510
Qwen3-1.7B +GRPOModel Scale=1.7B, Method=+GRPO2026.056.7
Qwen3-1.7B +RLSDModel Scale=1.7B, Method=+RLSD2026.056.7
Qwen3-1.7B +OPSDModel Scale=1.7B, Method=+OPSD2026.053.3
Qwen3-1.7B +RLRTModel Scale=1.7B, Method=+RLRT2026.053.3
Baseline ModelRollout Length=4K tokens2026.050.6
FP8 Rollout GRPORollout Length=16K tokens, Training Iteration=30, Rollout Precision=FP8, Training Precision=BF162026.050.2
Qwen3-1.7B BaseModel Scale=1.7B, Method=Base2026.050