List your company

Scientific Question Answering on GPQA D (Accuracy (mean@4, %) & Speedup)

38.5Accuracy (mean@4)GRPO
19.7824.6429.534.36May 8, 2026

Evaluation results

MethodLinks
GRPOBackbone=Qwen3-4B-base2026.0538.51
DUETBackbone=Qwen3-4B-base, Budget=Full2026.0538.31.43
DUETBackbone=Qwen3-4B-base, Budget=50%2026.0538.12.38
DAPOBackbone=Qwen3-4B-base2026.0537.40.63
VIPBackbone=Qwen3-4B-base2026.0536.80.75
ARRoLBackbone=Qwen3-4B-base2026.0534.51.19
DAPOBackbone=Qwen3-1.7B-base2026.0529.10.81
DUETBackbone=Qwen3-1.7B-base, Budget=50%2026.0529.12.51
DUETBackbone=Qwen3-1.7B-base, Budget=Full2026.05291.62
BaselineBackbone=Qwen3-4B-base2026.0528.8—
VIPBackbone=Qwen3-1.7B-base2026.0527.80.95
GRPOBackbone=Qwen3-1.7B-base2026.0527.41
ARRoLBackbone=Qwen3-1.7B-base2026.0526.71.3
GRPOBackbone=Llama-3.2-3B-Instruct2026.0525.41
ARRoLBackbone=Llama-3.2-3B-Instruct2026.0525.31.13
DUETBackbone=Llama-3.2-3B-Instruct, Budget=Full2026.0525.31.26
VIPBackbone=Llama-3.2-3B-Instruct2026.0524.90.95
DAPOBackbone=Llama-3.2-3B-Instruct2026.0524.50.82
DUETBackbone=Llama-3.2-3B-Instruct, Budget=50%2026.0523.22.04
BaselineBackbone=Qwen3-1.7B-base2026.0522.2—
BaselineBackbone=Llama-3.2-3B-Instruct2026.0520.5—