List your company

Scientific Reasoning on GPQA (Acc, Out Tok, Delta CoT)

61.19AccuracyRoT
18.96629.92840.8951.852Sep 26, 2025Nov 5, 2025Dec 15, 2025Jan 24, 2026Mar 5, 2026Apr 14, 2026May 25, 2026

Evaluation results

MethodLinks
RoTModel=Qwen3-4B2025.0961.194,773.4312.3
RoTModel=Qwen3-14B2025.0959.623,336.2516.6
CoTModel=Qwen3-4B2025.0958.215,443.16—
CoTModel=Qwen3-14B2025.0957.693,999.04—
RoTModel=Qwen3-8B2025.0957.144,165.1919.4
CoTModel=Qwen3-8B2025.0953.975,169.84—
RoT+TIModel=Qwen3-4B2025.0953.733,021.0344.5
RoT+TIModel=Qwen3-8B2025.0947.621,646.2768.2
GACmode=+ Token-φ2026.0543.5——
GACmode=w/o φ (plain SFT)2026.0542.4——
RoT+TIModel=Qwen3-14B2025.0942.311,954.9451.1
HPT2026.0540.4——
LUFFY2026.0540.1——
KL-ctrl2026.0540——
CHORDmode=Token-φ, fixed sched.2026.0539.6——
GradNorm-ctrl2026.0539.5——
SRFT2026.0539.2——
Nash-MTL2026.0539.1——
CAGrad2026.0538.8——
MGDA2026.0538.4——
SFT-best + RL2026.0537.2——
SFT-best2026.0534.8——
Qwen2.5-7B-Inst.2026.0531.5——
CoTModel=Qwen3-1.7B2025.0930.235,615.28—
RoT+TIModel=Qwen3-1.7B2025.0930.231,448.2674.2
RoTModel=Qwen3-0.6B2025.0929.414,158.042.9
CoTModel=Qwen3-0.6B2025.0927.944,039.28—
RoTModel=Qwen3-1.7B2025.0925.585,496.932.1
RoT+TIModel=Qwen3-0.6B2025.0920.59797.3480.3