List your company

Mathematical Reasoning on GSM8K (test) (Accuracy, Best Baseline, Correct Count)

92AccuracyDiffMean
-0.87223.23947.3571.461Apr 13, 2026Apr 25, 2026May 7, 2026May 20, 2026Jun 1, 2026Jun 13, 2026Jun 26, 2026

Evaluation results

MethodLinks
DiffMeanIntervention Level=Node-level, Intervention Target=Critic2026.0592——
DiffMeanIntervention Level=Node-level, Intervention Target=Refiner2026.0591.2——
DiffMeanIntervention Level=Node-level, Intervention Target=Planner2026.0590.3——
DiffMeanIntervention Level=Edge-level, Intervention Target=C→R2026.0588.5——
PCAIntervention Level=Node-level, Intervention Target=Refiner2026.0587.3——
CleanIntervention Level=Baseline2026.0587——
PCAIntervention Level=Node-level, Intervention Target=Critic2026.0586.7——
SMMDModel=SmolLM3, Params=3B2026.0670.74——
NTILModel=SmolLM3, Params=3B2026.0670.13——
PCAIntervention Level=Node-level, Intervention Target=Planner2026.0569.3——
CEModel=SmolLM3, Params=3B2026.0667.78——
Dynamic Persona RoutingTotal=1319, Sampled=1312026.0467.17—88
GCEModel=SmolLM3, Params=3B2026.0666.56——
NTLModel=SmolLM3, Params=3B2026.0664.13——
DiffMeanIntervention Level=Edge-level, Intervention Target=P→C2026.0561.1——
GCEModel=Llama3, Params=8B2026.0658.15——
SMMDModel=Qwen2.5, Params=1.5B2026.0657.77——
SMMDModel=Llama3, Params=8B2026.0657.7——
NTILModel=Llama3, Params=8B2026.0656.63——
NTLModel=Qwen2.5, Params=1.5B2026.0656.17——
CEModel=Llama3, Params=8B2026.0655.72——
NTILModel=Qwen2.5, Params=1.5B2026.0655.19——
CEModel=Qwen2.5, Params=1.5B2026.0654.99——
GCEModel=Qwen2.5, Params=1.5B2026.0652.69——
PCAIntervention Level=Edge-level, Intervention Target=R→J2026.0549.6——
PCAIntervention Level=Edge-level, Intervention Target=C→R2026.0548.7——
NTLModel=Llama3, Params=8B2026.0648.12——
PCAIntervention Level=Edge-level, Intervention Target=P→C2026.0543.4——
DiffMeanIntervention Level=Edge-level, Intervention Target=R→J2026.0537.2——
Direct attackIntervention Level=Text-level, Intervention Target=Critic2026.0535——
SMMDModel=Qwen2.5, Params=0.5B2026.0631.31——
NTILModel=Qwen2.5, Params=0.5B2026.0630.78——
NTLModel=Qwen2.5, Params=0.5B2026.0630.25——
CEModel=Qwen2.5, Params=0.5B2026.0629.34——
RePSIntervention Level=Node-level, Intervention Target=Planner2026.0529.2——
GCEModel=Qwen2.5, Params=0.5B2026.0628.2——
Direct attackIntervention Level=Text-level, Intervention Target=Refiner2026.0526.7——
RePSIntervention Level=Node-level, Intervention Target=Critic2026.0525.7——
Direct attackIntervention Level=Text-level, Intervention Target=Planner2026.0521.3——
RePSIntervention Level=Node-level, Intervention Target=Refiner2026.0519.5——
RePSIntervention Level=Edge-level, Intervention Target=R→J2026.0519.5——
RePSIntervention Level=Edge-level, Intervention Target=P→C2026.0513.3——
RePSIntervention Level=Edge-level, Intervention Target=C→R2026.052.7——
Best Static BaselineTotal=1319, Sampled=1312026.04—72.18—