Share your thoughts, 1 month free Claude Pro on usSee more

Process Reward Modeling on ProcessBench 1.0 (test)

87.3GSM8K Score

Qwen2.5-Math-PRM-72B

Updated 1mo ago

Evaluation Results

Method	Links
Qwen2.5-Math-PRM-72B 2026.04		87.3	80.6	74.3	71.1	78.3
Qwen2.5-Math-PRM-7B 2026.04		82.4	77.6	67.5	66.3	73.5
Llama3.1-8B-ImplicitPRM-DPO 2026.04		72.1	46	28	26.7	43.2
Skywork-PRM-7B 2026.04		70.8	53.6	22.9	21	42.1
TCRM-Llama3.1-8B (trained on math data) 2026.04		68.9	47.7	34.8	28.3	44.9
Qwen2.5-Math-7B-PRM800K 2026.04		68.2	62.6	50.7	44.3	56.5
Llama3.1-8B-ImplicitPRM-CE 2026.04		67.6	46.2	27.5	29.7	42.8
Skywork-PRM-1.5B 2026.04		59	48	19.3	19.2	36.4
Llama3.1-8B-ImplicitPRM-DPO (ref-free) 2026.04		57.7	42.6	34.1	36.7	42.8
Llama3.1-8B-ImplicitPRM-CE (ref-free) 2026.04		51.2	39.3	32.2	33.6	39.1
RLHFlow-Llama3.1-8B-PRM-Mistral 2026.04		50.4	33.4	13.8	15.8	28.4
RLHFlow-Llama3.1-8B-PRM-Deepseek 2026.04		38.8	33.8	16.9	16.9	26.6
ABC-Llama3.1-8B 2026.04		38.5	29.3	16	15.1	24.7
TCRM-Llama3.1-8B (trained on Skywork data) 2026.04		30.1	24.6	21.3	18	23.7