List your company
ResearchFormerly WizWand

Science Reasoning on GPQA (Accuracy)

59.22Accuracy (GPQA)GSPO + NSR
15.831227.095638.3649.6244Oct 1, 2025Nov 9, 2025Dec 19, 2025Jan 28, 2026Mar 8, 2026Apr 17, 2026May 27, 2026
Updated 5mo ago

Evaluation results

MethodLinks
GSPO + NSRModel=Qwen3-30B-A3B-Base, Zero-shot=true (k=32, T=1.0)2026.0559.22
GSPOModel=Qwen3-30B-A3B-Base, Zero-shot=true (k=32, T=1.0)2026.0558.96
In-place feedbackModel=Gemma-3-27B, Number of turns=42025.1058.7
In-place feedbackModel=Llama-3.1-70B, Number of turns=42025.1053.2
DAPO + NSRModel=Qwen3-8B-Base, Zero-shot=true (k=32, T=1.0)2026.0553.16
DAPOModel=Qwen3-8B-Base, Zero-shot=true (k=32, T=1.0)2026.0552.02
Multi-turnModel=Gemma-3-27B, Number of turns=42025.1051.6
Multi-turnModel=Llama-3.1-70B, Number of turns=42025.1050.8
MemoryModel=Llama-3.1-70B, Number of turns=42025.1050.8
SoftThinkModel Size=4B*2025.1150.3
OuroModel Size=4B*2025.1150
In-place feedbackModel=Gemma-3-27B, Number of turns=32025.1049.2
TaH+Model Size=4B*2025.1149
TaHModel Size=4B*2025.1148.5
Multi-turnModel=Gemma-3-27B, Number of turns=32025.1048.4
MemoryModel=Gemma-3-27B, Number of turns=42025.1047.6
In-place feedbackModel=Llama-3.1-70B, Number of turns=32025.1046.8
IBTPOBackbone=Qwen3-14B-Base2026.0546.6
MemoryModel=Llama-3.1-70B, Number of turns=32025.1046.6
StandardModel Size=4B*2025.1146.2
Multi-turnModel=Llama-3.1-70B, Number of turns=32025.1046
IBROBackbone=Qwen3-14B-Base2026.0544.7
MemoryModel=Gemma-3-27B, Number of turns=32025.1044.4
GRPO w/ Entropy RegBackbone=Qwen3-14B-Base2026.0544.2
TreeRLBackbone=Qwen3-14B-Base2026.0544.2
Vanilla GRPOBackbone=Qwen3-14B-Base2026.0543.7
In-place feedbackModel=Gemma-3-27B, Number of turns=22025.1043.7
DAPO + NSRModel=Qwen2.5-Math-7B, Zero-shot=true (k=32, T=1.0)2026.0542.99
In-place feedbackModel=Llama-3.1-70B, Number of turns=22025.1042.9
Multi-turnModel=Gemma-3-27B, Number of turns=22025.1042.9
Multi-turnModel=Llama-3.1-70B, Number of turns=22025.1042.1
MemoryModel=Llama-3.1-70B, Number of turns=22025.1040.5
MemoryModel=Gemma-3-27B, Number of turns=22025.1039.7
TaH+Model Size=1.7B2025.1139.4
DAPOModel=Qwen2.5-Math-7B, Zero-shot=true (k=32, T=1.0)2026.0539.27
Initial ModelBackbone=Qwen3-14B-Base2026.0538.6
OuroModel Size=0.6B2025.1135.4
In-place feedbackModel=Gemma-3-27B, Number of turns=12025.1034.9
Qwen3.5 2BN_act=1.9B, N_total=1.9B2026.0534.3
Multi-turnModel=Gemma-3-27B, Number of turns=12025.1034.1
MobileMoE-LN_act=922M, N_total=5.3B2026.0533.8
SoftThinkModel Size=1.7B2025.1133.3
TaHModel Size=1.7B2025.1133.3
MemoryModel=Gemma-3-27B, Number of turns=12025.1033.3
OuroModel Size=1.7B2025.1132.8
In-place feedbackModel=Llama-3.1-70B, Number of turns=12025.1031.8
TaH+Model Size=0.6B2025.1131.3
StandardModel Size=0.6B2025.1131.1
AlwaysThinkModel Size=0.6B2025.1130.8
AlwaysThinkModel Size=1.7B2025.1130.5
StandardModel Size=1.7B2025.1130.3
MemoryModel=Llama-3.1-70B, Number of turns=12025.1030.2
OLMo 2 1BN_act=1.5B, N_total=1.5B2026.0529.8
SmolLM2 1.7BN_act=1.7B, N_total=1.7B2026.0529.8
Multi-turnModel=Llama-3.1-70B, Number of turns=12025.1029.4
TaHModel Size=0.6B2025.1129
Llama 3.2 1BN_act=1.2B, N_total=1.2B2026.0528.8
MobileMoE-SN_act=272M, N_total=1.3B2026.0527.8
Qwen3.5 0.8BN_act=749M, N_total=749M2026.0526.3
Gemma 3 270MN_act=270M, N_total=270M2026.0525.8
SmolLM2 360MN_act=362M, N_total=362M2026.0525.8
Gemma 3 1BN_act=1.0B, N_total=1.0B2026.0525.3
MobileMoE-MN_act=528M, N_total=2.8B2026.0524.8
SoftThinkModel Size=0.6B2025.1124.7
OLMoE-1B-7BN_act=1.3B, N_total=6.9B2026.0524.2
MobileLLM-ProN_act=1.1B, N_total=1.1B2026.0523.2
Multi-turnModel=Llama-3.1-70B, Number of turns=02025.1018.3
MemoryModel=Llama-3.1-70B, Number of turns=02025.1018.3
In-place feedbackModel=Llama-3.1-70B, Number of turns=02025.1018.3
Multi-turnModel=Gemma-3-27B, Number of turns=02025.1017.5
MemoryModel=Gemma-3-27B, Number of turns=02025.1017.5
In-place feedbackModel=Gemma-3-27B, Number of turns=02025.1017.5