List your company
ResearchFormerly WizWand

Logical Reasoning on LogiQA (Accuracy %)

78.9LogiQA AccuracyDenser
41.30451.064560.82570.5855Jul 20, 2025Sep 8, 2025Oct 28, 2025Dec 18, 2025Feb 6, 2026Mar 28, 2026May 18, 2026
Updated 5mo ago

Evaluation results

MethodLinks
DenserBackbone=Qwen3-32B-think2025.1278.9
Process SupervisionBackbone=Qwen3-32B-think2025.1278.6
Reflection-CoTBackbone=Qwen3-32B-think2025.1278.1
DenserBackbone=Qwen3-32B2025.1277.3
Self-VerificationBackbone=Qwen3-32B-think2025.1277.2
Tree-of-ThoughtBackbone=Qwen3-32B-think2025.1276.9
Process SupervisionBackbone=Qwen3-32B2025.1276.8
Reflection-CoTBackbone=Qwen3-32B2025.1276.3
Self-ConsistencyBackbone=Qwen3-32B-think2025.1276.3
Self-VerificationBackbone=Qwen3-32B2025.1275.6
Think-to-ThinkBackbone=Qwen3-32B-think2025.1275.6
Tree-of-ThoughtBackbone=Qwen3-32B2025.1275.1
Self-ConsistencyBackbone=Qwen3-32B2025.1274.5
Think-to-ThinkBackbone=Qwen3-32B2025.1273.8
Chain-of-ThoughtBackbone=Qwen3-32B-think2025.1273.5
DenserBackbone=Qwen3-8B-think2025.1272.6
DenserBackbone=Qwen3-14B-no-think2025.1272.5
Chain-of-ThoughtBackbone=Qwen3-32B2025.1271.8
Process SupervisionBackbone=Qwen3-8B-think2025.1271.7
Process SupervisionBackbone=Qwen3-14B-no-think2025.1271.4
Reflection-CoTBackbone=Qwen3-8B-think2025.1271.2
DenserBackbone=Qwen3-8B2025.1270.9
Reflection-CoTBackbone=Qwen3-14B-no-think2025.1270.9
Self-VerificationBackbone=Qwen3-8B-think2025.1270.5
Self-VerificationBackbone=Qwen3-14B-no-think2025.1270.3
Process SupervisionBackbone=Qwen3-8B2025.1269.9
Tree-of-ThoughtBackbone=Qwen3-8B-think2025.1269.9
Tree-of-ThoughtBackbone=Qwen3-14B-no-think2025.1269.7
DenserBackbone=Qwen3-4B-think2025.1269.4
Reflection-CoTBackbone=Qwen3-8B2025.1269.4
Self-ConsistencyBackbone=Qwen3-8B-think2025.1269.4
Self-ConsistencyBackbone=Qwen3-14B-no-think2025.1269.2
Self-VerificationBackbone=Qwen3-8B2025.1268.8
Think-to-ThinkBackbone=Qwen3-8B-think2025.1268.7
Think-to-ThinkBackbone=Qwen3-14B-no-think2025.1268.5
Process SupervisionBackbone=Qwen3-4B-think2025.1268.3
Tree-of-ThoughtBackbone=Qwen3-8B2025.1268.1
Reflection-CoTBackbone=Qwen3-4B-think2025.1267.8
Self-ConsistencyBackbone=Qwen3-8B2025.1267.6
Self-VerificationBackbone=Qwen3-4B-think2025.1267.1
Think-to-ThinkBackbone=Qwen3-8B2025.1266.9
Chain-of-ThoughtBackbone=Qwen3-8B-think2025.1266.9
DenserBackbone=Qwen3-4B2025.1266.8
Chain-of-ThoughtBackbone=Qwen3-14B-no-think2025.1266.7
IH-GRPOBackbone=Qwen3-8B2026.0566.67
Tree-of-ThoughtBackbone=Qwen3-4B-think2025.1266.4
Process SupervisionBackbone=Qwen3-4B2025.1266
Self-ConsistencyBackbone=Qwen3-4B-think2025.1265.9
Reflection-CoTBackbone=Qwen3-4B2025.1265.5
Think-to-ThinkBackbone=Qwen3-4B-think2025.1265.2
Chain-of-ThoughtBackbone=Qwen3-8B2025.1265.2
Self-VerificationBackbone=Qwen3-4B2025.1264.9
Tree-of-ThoughtBackbone=Qwen3-4B2025.1264.2
Self-ConsistencyBackbone=Qwen3-4B2025.1263.8
Chain-of-ThoughtBackbone=Qwen3-4B-think2025.1263.5
IH-GRPOBackbone=Qwen3-4B2026.0563.29
Think-to-ThinkBackbone=Qwen3-4B2025.1263.1
Chain-of-ThoughtBackbone=Qwen3-4B2025.1261.7
BaseBackbone=Qwen3-4B2026.0557.76
IH-GRPOBackbone=Qwen3-1.7B2026.0556.99
BaseBackbone=Qwen3-8B2026.0551.15
RandomIteration=2, Selection Ratio=0.2, Model Scale=3b2026.0449.77
EVOSELECTIteration=1, Selection Ratio=0.5, Model Scale=3b2026.0449.62
AttributionIteration=2, Selection Ratio=0.5, Model Scale=3b2026.0449.16
RL+SAEModel=Qwen3-30B-A3B2026.0548.75
EVOSELECTIteration=2, Selection Ratio=0.5, Model Scale=3b2026.0448.54
TSDSIteration=1, Selection Ratio=0.5, Model Scale=3b2026.0448.39
Vanilla RLModel=Qwen3-30B-A3B2026.0548.38
Before RLModel=Qwen3-30B-A3B2026.0548.12
TSDSIteration=2, Selection Ratio=0.2, Model Scale=3b2026.0448.08
Before RLModel=Qwen3-8B2026.0548
EVOSELECTIteration=1, Selection Ratio=0.2, Model Scale=3b2026.0447.93
RandomIteration=1, Selection Ratio=0.2, Model Scale=3b2026.0447.62
DiversityIteration=1, Selection Ratio=0.5, Model Scale=3b2026.0447.62
DiversityIteration=1, Selection Ratio=0.2, Model Scale=3b2026.0447.47
AllIteration=2, Model Scale=3b2026.0447.31
Vanilla RLModel=Qwen3-8B2026.0547.12
RL+SAEModel=Qwen3-8B2026.0547.12
BaseIteration=Base, Model Scale=3b2026.0446.7
DiversityIteration=2, Selection Ratio=0.5, Model Scale=3b2026.0446.39
Attr-DivIteration=2, Selection Ratio=0.2, Model Scale=3b2026.0446.24
Attr-DivIteration=2, Selection Ratio=0.5, Model Scale=3b2026.0446.24
TSDSIteration=1, Selection Ratio=0.2, Model Scale=3b2026.0446.08
EVOSELECTIteration=2, Selection Ratio=0.2, Model Scale=3b2026.0446.08
AttributionIteration=1, Selection Ratio=0.5, Model Scale=3b2026.0445.62
AttributionIteration=2, Selection Ratio=0.2, Model Scale=3b2026.0445.47
TSDSIteration=2, Selection Ratio=0.5, Model Scale=3b2026.0445.47
AttributionIteration=1, Selection Ratio=0.2, Model Scale=3b2026.0445.31
Attr-DivIteration=1, Selection Ratio=0.5, Model Scale=3b2026.0445.16
DiversityIteration=2, Selection Ratio=0.2, Model Scale=3b2026.0445.16
BaseBackbone=Qwen3-1.7B2026.0545.01
Attr-DivIteration=1, Selection Ratio=0.2, Model Scale=3b2026.0444.85
AllIteration=1, Model Scale=3b2026.0444.55
RandomIteration=2, Selection Ratio=0.5, Model Scale=3b2026.0444.55
SASFTModel=Qwen3-8B-Base, Dataset=Russian 110k2025.0744.5
RandomIteration=1, Selection Ratio=0.5, Model Scale=3b2026.0443.78
SASFTModel=Qwen3-8B-Base, Training Set=Chinese 210k2025.0743.38
SASFTBackbone=Qwen3-8B-Base, Training dataset setting=Korean 210k2025.0742.88
SASFTModel=Qwen3-8B-Base2025.0742.75
SASFTModel=Qwen3-8B-Base, Training Dataset=Chinese 110k2026.0542.75