List your company

Multi-answer Question Answering on MAQA-ΔK−1

-0.149KL DivergenceP(True)
-0.680762.908626.49810.08738Jan 29, 2026

Evaluation results

MethodLinks
P(True)Backbone=Gemma-3-4B2026.01-0.149—
P(True)Backbone=Gemma-3-12B2026.010.083—
P(True)Backbone=Qwen3-30B-A3B2026.010.125—
P(True)Backbone=Qwen3-4B2026.010.148—
KLEBackbone=Qwen3-4B2026.010.365—
SEBackbone=Qwen3-4B2026.010.381—
COCOABackbone=Qwen3-4B2026.010.411—
SARBackbone=Qwen3-4B2026.010.416—
MSPBackbone=Qwen3-4B2026.010.43—
R_BayesBackbone=Qwen3-4B2026.010.467—
MSPBackbone=Gemma-3-12B2026.010.54—
SEBackbone=Gemma-3-12B2026.010.544—
KLEBackbone=Gemma-3-12B2026.010.55—
SARBackbone=Gemma-3-12B2026.010.573—
COCOABackbone=Gemma-3-12B2026.010.573—
MSPBackbone=Qwen3-30B-A3B2026.010.573—
KLEBackbone=Qwen3-30B-A3B2026.010.58—
SARBackbone=Qwen3-30B-A3B2026.010.581—
SEBackbone=Qwen3-30B-A3B2026.010.587—
R_BayesBackbone=Qwen3-30B-A3B2026.010.601—
COCOABackbone=Qwen3-30B-A3B2026.010.619—
R_BayesBackbone=Gemma-3-12B2026.010.634—
MSPBackbone=Gemma-3-4B2026.010.635—
COCOABackbone=Gemma-3-4B2026.010.679—
R_BayesBackbone=Gemma-3-4B2026.010.696—
SARBackbone=Gemma-3-4B2026.010.699—
KLEBackbone=Gemma-3-4B2026.010.699—
SEBackbone=Gemma-3-4B2026.010.705—
Bayes (ours)Model=Qwen3-30B-A3B, Decoding Strategy=Bayes (ours)2026.015.032—
Bayes (ours)Model=Gemma-3-12B, Decoding Strategy=Bayes (ours)2026.015.632—
BeamModel=Qwen3-30B-A3B, Decoding Strategy=Beam2026.016.759—
ℓMAPModel=Qwen3-30B-A3B, Decoding Strategy=ℓMAP2026.016.884—
Self-ConsistencyModel=Qwen3-30B-A3B, Decoding Strategy=Self-Consistency2026.016.962—
ℓSampleModel=Qwen3-30B-A3B, Decoding Strategy=ℓSample2026.017.283—
BeamModel=Gemma-3-12B, Decoding Strategy=Beam2026.017.792—
Bayes (ours)Model=Gemma-3-4B, Decoding Strategy=Bayes (ours)2026.017.94—
ℓMAPModel=Gemma-3-12B, Decoding Strategy=ℓMAP2026.018.051—
Self-ConsistencyModel=Gemma-3-12B, Decoding Strategy=Self-Consistency2026.018.232—
ℓSampleModel=Gemma-3-12B, Decoding Strategy=ℓSample2026.018.42—
Self-ConsistencyModel=Gemma-3-4B, Decoding Strategy=Self-Consistency2026.0110.315—
BeamModel=Gemma-3-4B, Decoding Strategy=Beam2026.0110.33—
Bayes (ours)Model=Qwen3-4B, Decoding Strategy=Bayes (ours)2026.0110.415—
ℓSampleModel=Gemma-3-4B, Decoding Strategy=ℓSample2026.0110.514—
ℓMAPModel=Gemma-3-4B, Decoding Strategy=ℓMAP2026.0110.545—
BeamModel=Qwen3-4B, Decoding Strategy=Beam2026.0112.786—
ℓMAPModel=Qwen3-4B, Decoding Strategy=ℓMAP2026.0112.885—
Self-ConsistencyModel=Qwen3-4B, Decoding Strategy=Self-Consistency2026.0113.103—
ℓSampleModel=Qwen3-4B, Decoding Strategy=ℓSample2026.0113.145—
COCOAModel=Gemma-3-4B2026.01—0.698
COCOAModel=Gemma-3-12B2026.01—0.689
COCOAModel=Qwen3-4B2026.01—0.639
COCOAModel=Qwen3-30B-A3B2026.01—0.705
KLEModel=Gemma-3-4B2026.01—0.708
KLEModel=Gemma-3-12B2026.01—0.701
KLEModel=Qwen3-4B2026.01—0.62
KLEModel=Qwen3-30B-A3B2026.01—0.688
MSPModel=Gemma-3-4B2026.01—0.685
MSPModel=Gemma-3-12B2026.01—0.668
MSPModel=Qwen3-4B2026.01—0.639
MSPModel=Qwen3-30B-A3B2026.01—0.681
P(True)Model=Gemma-3-4B2026.01—0.465
P(True)Model=Gemma-3-12B2026.01—0.539
P(True)Model=Qwen3-4B2026.01—0.555
P(True)Model=Qwen3-30B-A3B2026.01—0.529
RBayesModel=Gemma-3-4B2026.01—0.744
RBayesModel=Gemma-3-12B2026.01—0.753
RBayesModel=Qwen3-4B2026.01—0.639
RBayesModel=Qwen3-30B-A3B2026.01—0.709
SARModel=Gemma-3-4B2026.01—0.709
SARModel=Gemma-3-12B2026.01—0.704
SARModel=Qwen3-4B2026.01—0.639
SARModel=Qwen3-30B-A3B2026.01—0.689
SEModel=Gemma-3-4B2026.01—0.724
SEModel=Gemma-3-12B2026.01—0.708
SEModel=Qwen3-4B2026.01—0.624
SEModel=Qwen3-30B-A3B2026.01—0.7