List your company
ResearchFormerly WizWand

Reward Model Evaluation on RewardBench (Spearman ρ)

0.542Spearman ρprob-weighted EV (M2)
0.228960.310230.39150.47277May 15, 2026
Updated 5mo ago

Evaluation results

MethodLinks
prob-weighted EV (M2)Model=Gemma-3-27B2026.050.542
zero-shot BDAS-1D (M4)Model=Gemma-3-27B2026.050.537
prompted argmax (M1)Model=Gemma-3-27B2026.050.522
supervised residual probe (M3)Model=Gemma-3-27B2026.050.487
prob-weighted EV (M2)Model=Gemma-3-12B2026.050.432
zero-shot BDAS-1D (M4)Model=Gemma-3-12B2026.050.425
prompted argmax (M1)Model=Gemma-3-12B2026.050.422
prob-weighted EV (M2)Model=Qwen2.5-14B2026.050.418
prompted argmax (M1)Model=Qwen2.5-14B2026.050.413
supervised residual probe (M3)Model=Qwen2.5-14B2026.050.398
supervised residual probe (M3)Model=Gemma-3-12B2026.050.375
prob-weighted EV (M2)Model=Qwen2.5-7B2026.050.367
zero-shot BDAS-1D (M4)Model=Qwen2.5-14B2026.050.345
zero-shot BDAS-1D (M4)Model=Qwen2.5-7B2026.050.338
prompted argmax (M1)Model=Qwen2.5-7B2026.050.327
supervised residual probe (M3)Model=Llama-3.1-8B2026.050.32
zero-shot BDAS-1D (M4)Model=Llama-3.1-8B2026.050.298
supervised residual probe (M3)Model=Qwen2.5-7B2026.050.291
prompted argmax (M1)Model=Llama-3.1-8B2026.050.276
prob-weighted EV (M2)Model=Llama-3.1-8B2026.050.241