List your company
ResearchFormerly WizWand

Mathematical Reasoning on MATH 500 (pass@1 Accuracy)

94.8Accuracy (pass@1)BF16
45.08857.99470.983.806May 20, 2025Jul 26, 2025Oct 2, 2025Dec 8, 2025Feb 14, 2026Apr 22, 2026Jun 29, 2026
Updated 4mo ago

Evaluation results

MethodLinks
BF16Model=DeepSeek-R1-Distill-Qwen-32B, Bit-width=BF162025.1294.8
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.0594.8
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.72026.0594.6
BaseModel backbone=Qwen3-4B-It, Fine-tuning method=Base2026.0594.2
MixKVQModel=DeepSeek-R1-Distill-Qwen-32B, Bit-width=C2.32025.1294
BF16Model=DeepSeek-R1-Distill-Qwen-14B, Bit-width=BF162025.1293.6
KIVIModel=DeepSeek-R1-Distill-Qwen-32B, Bit-width=KV42025.1293.4
RotateKVModel=DeepSeek-R1-Distill-Qwen-32B, Bit-width=KV42025.1293.4
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.0593
KVQuantModel=DeepSeek-R1-Distill-Qwen-32B, Bit-width=KV42025.1292.6
RotateKVModel=DeepSeek-R1-Distill-Qwen-14B, Bit-width=KV42025.1292.4
MixKVQModel=DeepSeek-R1-Distill-Qwen-14B, Bit-width=C2.32025.1292.4
KIVIModel=DeepSeek-R1-Distill-Qwen-14B, Bit-width=KV42025.1292.2
KVTunerModel=DeepSeek-R1-Distill-Qwen-32B, Bit-width=C2.912025.1291.4
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=02026.0590.8
KVQuantModel=DeepSeek-R1-Distill-Qwen-14B, Bit-width=KV42025.1290.4
BF16Model=DeepSeek-R1-Distill-Llama-8B, Bit-width=BF162025.1289.8
KVTunerModel=DeepSeek-R1-Distill-Qwen-14B, Bit-width=C2.902025.1289.8
KIVIModel=DeepSeek-R1-Distill-Qwen-32B, Bit-width=KV22025.1289.8
MixKVQModel=DeepSeek-R1-Distill-Llama-8B, Bit-width=C2.72025.1288.2
RotateKVModel=DeepSeek-R1-Distill-Llama-8B, Bit-width=KV42025.1288
KVTunerModel=DeepSeek-R1-Distill-Llama-8B, Bit-width=C3.252025.1287.8
KIVIModel=DeepSeek-R1-Distill-Qwen-14B, Bit-width=KV22025.1287.8
KVQuantModel=DeepSeek-R1-Distill-Llama-8B, Bit-width=KV42025.1287.4
KIVIModel=DeepSeek-R1-Distill-Llama-8B, Bit-width=KV42025.1286.6
AAPO-1.5BZero-shot=true, Number of parameters=1.5B2025.0586
KIVIModel=DeepSeek-R1-Distill-Llama-8B, Bit-width=KV22025.1285.8
Still-3-1.5B-PreviewZero-shot=true, Number of parameters=1.5B2025.0585.5
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.72026.0585.4
DAPOBase Model=Qwen-2.5 Math-7B2026.0685.2
DAPO (WITH FORKING TOKENS)Base Model=Qwen-2.5 Math-7B2026.0685.1
GPG-1.5BZero-shot=true, Number of parameters=1.5B2025.0585
PS-PPO (OPTIMIZED, B=128)Base Model=Qwen-2.5 Math-7B2026.0685
S-GRPOBase Model=Qwen-2.5 Math-7B2026.0684.7
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.0584.6
PS-PPO (TIME-PRIOR)Base Model=Qwen-2.5 Math-7B2026.0684.6
GRPO-1.5BZero-shot=true, Number of parameters=1.5B2025.0584.4
DeepSeek-R1-Distill-Qwen-1.5BZero-shot=true, Number of parameters=1.5B2025.0583.9
GRPOBase Model=Qwen-2.5 Math-7B2026.0683.9
PS-PPO (HEURISTIC)Base Model=Qwen-2.5 Math-7B2026.0683.7
ISPOBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0683.6
RLOOBase Model=Qwen-2.5 Math-7B2026.0683.6
PS-PPO (UNIFORM)Base Model=Qwen-2.5 Math-7B2026.0683.6
DR.GRPOBase Model=Qwen-2.5 Math-7B2026.0683.5
BaseModel backbone=Qwen3-8B, Fine-tuning method=Base2026.0583.4
FIXED LENGTH (L=512)Base Model=Qwen-2.5 Math-7B2026.0683.3
ISPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0683
BASEBase Model=Qwen-2.5 Math-7B2026.0682.5
AAPO-7BZero-shot=true, Number of parameters=7B2025.0582.4
PACRBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0681.9
Dr. GRPOBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0681.8
ProGRPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0680.5
GTPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0680.2
GRPO-SBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0680.1
GPG-7BZero-shot=true, Number of parameters=7B2025.0580
Oat-Zero-7BZero-shot=true, Number of parameters=7B2025.0580
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.0580
Scaf-GRPOBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0680
ISPOBase Model=Qwen2.5-Math-1.5B, Decoding Strategy=Greedy (T=0)2026.0679.2
SimpleRL-Zero-7BZero-shot=true, Number of parameters=7B2025.0578.2
Eurus-2-7B-PRIMEZero-shot=true, Number of parameters=7B2025.0578.2
PREPOBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0677.9
Scaf-GRPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0677.8
PACRBase Model=Qwen2.5-Math-1.5B, Decoding Strategy=Greedy (T=0)2026.0677.4
Dr. GRPOBase Model=Qwen2.5-Math-1.5B, Decoding Strategy=Greedy (T=0)2026.0676.8
PREPOBase Model=Qwen2.5-Math-1.5B, Decoding Strategy=Greedy (T=0)2026.0676.3
PREPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0676.3
Vanilla GRPOBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0675.8
Vanilla GRPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0675.3
OPSDModel backbone=Qwen3-4B-It, Fine-tuning method=OPSD2026.0574.4
OPSDModel backbone=Qwen3-8B, Fine-tuning method=OPSD2026.0574.2
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=02026.0573.6
Scaf-GRPOBase Model=Qwen2.5-Math-1.5B, Decoding Strategy=Greedy (T=0)2026.0673.4
BaseModel backbone=Qwen3-1.7B, Fine-tuning method=Base2026.0572.4
Vanilla GRPOBase Model=Qwen2.5-Math-1.5B, Decoding Strategy=Greedy (T=0)2026.0672.2
p*LLM Family=Qwen 3 (1.7B and 14B)2026.0464
GRPOModel=Qwen2.5-3B, Training Data=MATH, Inference template=chat inference template, Decoding strategy=Greedy2025.0563.6
GRPO-PVModel=Qwen2.5-3B, Training Data=MATH, Inference template=chat inference template, Decoding strategy=Greedy2025.0563.6
confidence betLLM Family=Qwen 3 (1.7B and 14B), lambda=02026.0461.8
confidence betLLM Family=Qwen 3 (1.7B and 14B), lambda=0.22026.0461.6
INTUITORModel=Qwen2.5-3B, Training Data=MATH, Inference template=chat inference template, Decoding strategy=Greedy2025.0561.2
Imp. rewardLLM Family=Qwen 3 (1.7B and 14B)2026.0460.6
Dual KADLLM Family=Qwen 3 (1.7B and 14B), lambda=0.42026.0460.6
confidence betLLM Family=Qwen 3 (1.7B and 14B), lambda=0.12026.0459.8
Dual KADLLM Family=Qwen 3 (1.7B and 14B), lambda=0.32026.0459.6
NudgingLLM Family=Qwen 3 (1.7B and 14B), lambda=0.42026.0457.8
INTUITOR-CodeModel=Qwen2.5-3B, Training Data=Codeforces, Inference template=chat inference template, Decoding strategy=Greedy2025.0557.2
GRPOModel=Qwen2.5-1.5B, Training Data=MATH, Inference template=chat inference template, Decoding strategy=Greedy2025.0556
Qwen2.5-Math-7BZero-shot=true, Number of parameters=7B2025.0555.4
Base modelBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0654.7
BaseModel=Qwen2.5-3B, Training Data=-, Inference template=chat inference template, Decoding strategy=Greedy2025.0554.4
INTUITORModel=Qwen2.5-1.5B, Training Data=MATH, Inference template=chat inference template, Decoding strategy=Greedy2025.0553
q*LLM Family=Qwen 3 (1.7B and 14B)2026.0453
NudgingLLM Family=Qwen 3 (1.7B and 14B), lambda=0.32026.0452.8
MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.0552.2
pLLM Family=Qwen 3 (1.7B and 14B)2026.0451.8
Base modelBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0650.6
MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.0547.6
PS-PPO (OPTIMIZED, B=128)Base Model=Llama-3.1 8B-Instruct2026.0647.6
DAPO (WITH FORKING TOKENS)Base Model=Llama-3.1 8B-Instruct2026.0647