List your company

Arithmetic Reasoning on GSM8K (ACC, PPL)

83.3ACCGRPO
17.010434.220251.4368.6398Aug 13, 2023Dec 29, 2023May 15, 2024Sep 30, 2024Feb 15, 2025Jul 3, 2025Nov 18, 2025

Evaluation results

MethodLinks
GRPOFramework=Agent-R1, Backbone=Qwen3-4B, Type=RL-optimized2025.1183.3—
RLOOFramework=Agent-R1, Backbone=Qwen3-4B, Type=RL-optimized2025.1181.6—
Reinforce++Framework=Agent-R1, Backbone=Qwen3-4B, Type=RL-optimized2025.1178.9—
PPOFramework=Agent-R1, Backbone=Qwen3-4B, Type=RL-optimized2025.1178.1—
ReActFramework=Agent-R1, Backbone=Qwen3-4B, Type=Training-free baseline2025.1153.1—
FP16 baselineBackbone Model=LLaMA-7B2023.0830.251.47
TSLDBackbone Model=LLaMA-7B2023.0826.231.52
Logit KDBackbone Model=LLaMA-7B2023.0825.471.52
TSLDBackbone Model=OPT-6.7B2023.0824.492.14
FP16 baselineBackbone Model=OPT-6.7B2023.0822.521.89
Logit KDBackbone Model=OPT-6.7B2023.0821.081.93
OPT-2.7B FP16Model=OPT-2.7B, Precision=FP162023.0820.392.07
TSLDModel=OPT-2.7B, Quantization=QAT-KD (tensor-wise)2023.0820.242.03
LogitModel=OPT-2.7B, Quantization=QAT-KD (tensor-wise)2023.0820.022.03
GT+LogitModel=OPT-2.7B, Quantization=QAT-KD (tensor-wise)2023.0819.562.12