List your company

Comprehensive Reasoning on ScienceQA

84.2AccuracyUnsilencing Latent Reasoning
68.28872.41976.5580.681May 4, 2026

Evaluation results

MethodLinks
Unsilencing Latent ReasoningBackbone=Qwen2.5VL-7B2026.0584.2
MCoTBackbone=Qwen2.5VL-7B2026.0583.9
CCoTBackbone=Qwen2.5VL-7B2026.0583.8
CoVTBackbone=Qwen2.5VL-7B2026.0583.8
DMLRBackbone=Qwen2.5VL-7B2026.0583.4
LVRRFBackbone=Qwen2.5VL-7B2026.0583.1
LVRBackbone=Qwen2.5VL-7B2026.0582.8
VanillaBackbone=Qwen2.5VL-7B2026.0582.3
MonetBackbone=Qwen2.5VL-7B2026.0578.8
ICoTBackbone=Qwen2.5VL-7B2026.0578.4
Unsilencing Latent ReasoningBackbone=Qwen2.5VL-3B2026.0574.3
MCoTBackbone=Qwen2.5VL-3B2026.0574.1
CCoTBackbone=Qwen2.5VL-3B2026.0573.9
DMLRBackbone=Qwen2.5VL-3B2026.0573.8
LVRRFBackbone=Qwen2.5VL-3B2026.0573.7
VanillaBackbone=Qwen2.5VL-3B2026.0573.5
LVRBackbone=Qwen2.5VL-3B2026.0573.2
ICoTBackbone=Qwen2.5VL-3B2026.0568.9