List your company
ResearchFormerly WizWand

Logical Reasoning on Sudoku

94.3AccuracyEGSPO-SA
11.25632.815554.37575.9345Oct 2, 2025Nov 9, 2025Dec 18, 2025Jan 25, 2026Mar 5, 2026Apr 12, 2026May 21, 2026
Updated 4mo ago

Evaluation results

MethodLinks
EGSPO-SASeq. Len.=256, n-shot=0-shot2026.0394.3
EGSPO-SASeq. Len.=Best, n-shot=0-shot2026.0394.3
SPGSeq. Len.=256, n-shot=3-shot2026.0394
SPGSeq. Len.=Best, n-shot=3-shot2026.0394
SPG w/ mixture (ours)Seq Len=256, Evaluation Protocol=3-shot2025.1094
EGSPO-SASeq. Len.=128, n-shot=0-shot2026.0393.7
EGSPOSeq. Len.=256, n-shot=0-shot2026.0393.6
EGSPOSeq. Len.=Best, n-shot=0-shot2026.0393.6
EGSPO-SASeq. Len.=512, n-shot=0-shot2026.0393.4
OracleBackbone=LLaDA-8B-Base, Shots=4-shot2026.0493.4
EGSPOSeq. Len.=128, n-shot=0-shot2026.0393.3
SPGSeq. Len.=512, n-shot=3-shot2026.0393.1
SPG w/ mixture (ours)Seq Len=512, Evaluation Protocol=3-shot2025.1093.1
d2Seq. Len.=Best, n-shot=0-shot2026.0391.9
OracleBackbone=Dream-7B-Base, Shots=4-shot2026.0491.2
SPG w/ EUBO (ours)Seq Len=512, Evaluation Protocol=3-shot2025.1089.9
RL (BF16)Model=Qwen3.5-9B, Bitwidth=BF162026.0589.84
AISModel=Qwen3.5-9B, Bitwidth=FP82026.0589.45
Auto-ICLBackbone=Dream-7B-Base, Shots=4-shot, Latency (s)=0.702026.0489.3
LLaDA-8B-Instruct w/ DAMSeq Len=1282026.0289.2
EGSPOSeq. Len.=512, n-shot=0-shot2026.0389.1
FlashRL (TIS)Model=Qwen3.5-9B, Bitwidth=FP82026.0588.67
LLaDA-8B-Instruct w/ DAMSeq Len=2562026.0288.12
SPG w/ EUBO (ours)Seq Len=256, Evaluation Protocol=3-shot2025.1087.1
PrefixBackbone=Dream-7B-Base, Shots=4-shot, Latency (s)=0.682026.0486.4
RL (FP8 Rollout)Model=Qwen3.5-9B, Bitwidth=FP82026.0585.94
Auto-ICLBackbone=LLaDA-8B-Base, Shots=4-shot, Latency (s)=0.752026.0484.4
PrefixBackbone=LLaDA-8B-Base, Shots=4-shot, Latency (s)=0.772026.0483
SPGSeq. Len.=128, n-shot=3-shot2026.0382.9
SPG w/ mixture (ours)Seq Len=128, Evaluation Protocol=3-shot2025.1082.9
SPG w/ EUBO (ours)Seq Len=128, Evaluation Protocol=3-shot2025.1081.2
VanillaBackbone=LLaDA-8B-Base, Shots=4-shot, Latency (s)=0.762026.0479
wd1Seq. Len.=256, n-shot=0-shot2026.0376.4
wd1Seq. Len.=Best, n-shot=0-shot2026.0376.4
+ PlannerSampling Method=+ Planner, Timestep Budget (T)=1282026.0475.2
BothSampling Method=Both, Timestep Budget (T)=1282026.0474.3
+ EOS Temp.Sampling Method=+ EOS Temp., Timestep Budget (T)=1282026.0474.2
+ EOS Temp.Sampling Method=+ EOS Temp., Timestep Budget (T)=642026.0473.9
Prob MarginSampling Method=Prob Margin, Timestep Budget (T)=1282026.0473.8
Prob MarginSampling Method=Prob Margin, Timestep Budget (T)=642026.0471.3
Top1 ProbSampling Method=Top1 Prob, Timestep Budget (T)=1282026.0470.5
+ PlannerSampling Method=+ Planner, Timestep Budget (T)=642026.0470.5
TemperatureSampling Method=Temperature, Timestep Budget (T)=1282026.0470.4
BothSampling Method=Both, Timestep Budget (T)=642026.0469.9
TemperatureSampling Method=Temperature, Timestep Budget (T)=642026.0469.7
Top1 ProbSampling Method=Top1 Prob, Timestep Budget (T)=642026.0468.4
UniGRPOSeq Len=256, Evaluation Protocol=3-shot2025.1067
Init. PositionSampling Method=Init. Position, Timestep Budget (T)=1282026.0466.9
RandomBackbone=LLaDA-8B-Base, Shots=4-shot, Latency (s)=0.772026.0463.1
RandomBackbone=Dream-7B-Base, Shots=4-shot, Latency (s)=0.692026.0463.1
UniGRPOSeq Len=512, Evaluation Protocol=3-shot2025.1062.9
wd1Seq. Len.=512, n-shot=0-shot2026.0362.8
UniGRPOSeq Len=128, Evaluation Protocol=3-shot2025.1059
Init. PositionSampling Method=Init. Position, Timestep Budget (T)=642026.0456.4
AISModel=Qwen3-8B, Bitwidth=FP82026.0539.1
RL (BF16)Model=Qwen3-8B, Bitwidth=BF162026.0538.28
FlashRL (TIS)Model=Qwen3-8B, Bitwidth=FP82026.0537.5
RL (FP8 Rollout)Model=Qwen3-8B, Bitwidth=FP82026.0536.33
WD1 (reproduced)Seq Len=128, Evaluation Protocol=3-shot2025.1033.1
DMPO-LLaDASequence Length=1282025.1032.76
D1 (reproduced)Seq Len=256, Evaluation Protocol=3-shot2025.1032.5
WD1 (reproduced)Seq Len=256, Evaluation Protocol=3-shot2025.1032.1
VanillaBackbone=Dream-7B-Base, Shots=4-shot, Latency (s)=0.672026.0430
D1 (reproduced)Seq Len=512, Evaluation Protocol=3-shot2025.1029.3
LLaDA-1.5Seq Len=512, Evaluation Protocol=3-shot2025.1029
LLaDA-8B-InstructSeq Len=256, Evaluation Protocol=3-shot2025.1027.7
LLaDA-1.5Seq Len=256, Evaluation Protocol=3-shot2025.1026.9
LLaDA-8B-InstructSeq Len=512, Evaluation Protocol=3-shot2025.1026.2
DMPO-LLaDA-SFTSequence Length=2562025.1025.73
DMPO-LLaDA-1.5Sequence Length=1282025.1025.34
DMPO-LLaDA-SFTSequence Length=1282025.1025.2
DMPO-LLaDASequence Length=2562025.1024.56
DMPO-LLaDA-1.5Sequence Length=2562025.1024.51
cGRPO-LLaDASequence Length=1282025.1024.17
cGRPO-LLaDASequence Length=2562025.1024.17
LLaDA-8B-Instruct w/ D1Seq Len=1282026.0223.82
DMPO-LLaDA-SFTSequence Length=5122025.1023.78
DMPO-LLaDA-1.5Sequence Length=5122025.1023.34
WD1 (reproduced)Seq Len=512, Evaluation Protocol=3-shot2025.1022.5
LIFT2Backbone=Dream-7B, H=22026.0522.5
OursSeq Len=1282025.1022.4
d1Seq. Len.=128, n-shot=0-shot2026.0322.1
d1Seq. Len.=Best, n-shot=0-shot2026.0322.1
d1-LLaDASequence Length=1282025.1021.97
cGRPO-LLaDASequence Length=5122025.1021.97
OursSeq Len=2562025.1020.3
DMPO-LLaDASequence Length=5122025.1019.97
AncestralSampling Method=Ancestral, Timestep Budget (T)=642026.0419.1
AncestralSampling Method=Ancestral, Timestep Budget (T)=1282026.0418.5
diffu-GRPOSeq Len=1282025.1018.4
LIFT3Backbone=Dream-7B, H=32026.0517.5
CARTBackbone=Dream-7B2026.0517.2
d1Seq. Len.=256, n-shot=0-shot2026.0316.7
Dream-Instruct (7B)Sequence Length=2562025.1016.41
OursSeq Len=5122025.1016.1
GIFTBackbone=Dream-7B2026.0516
WINOSeq Len=2562025.1015.2
LLaDA-1.5 (8B)Sequence Length=1282025.1015.19
VanillaBackbone=Dream-7B2026.0514.8
Dream-Instruct (7B)Sequence Length=1282025.1014.45