# Step-level Correctness Discrimination on ProcessBench (GSM8K, MATH, Olympiad Bench, Omni Math) Source: [Step-level Correctness Discrimination on ProcessBench (GSM8K, MATH, Olympiad Bench, Omni Math)]() Collection index: [llms.txt]() Data updated: 2026-02-26T05:45:49.734Z Evaluation of model performance on step-level correctness discrimination across the GSM8K, MATH, Olympiad Bench, and Omni Math components of ProcessBench, measured using error, correctness, and F1 scores for each component. A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. | Scope | Recorded value | | --- | --- | | Task | Step-level Correctness Discrimination | | Dataset (including any stated split) | ProcessBench GSM8K MATH Olympiad Bench Omni Math | | Metric | GSM8K Error Rate | | Unit | PERCENT | | Direction | Lower is better | | Results loaded | 12 of 12 | | Coverage | All results currently returned for this benchmark | Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol. ## Metrics - [GSM8K Error Rate](): Lower is better; unit: PERCENT. - [GSM8K Accuracy](): Higher is better; unit: PERCENT. - [GSM8K F1 Score](): Higher is better; unit: PERCENT. - [MATH Error Rate](): Lower is better; unit: PERCENT. - [MATH Accuracy](): Higher is better; unit: PERCENT. - [MATH F1 Score](): Higher is better; unit: PERCENT. - [Olympiad Bench Error Rate](): Lower is better; unit: PERCENT. - [Olympiad Bench Accuracy](): Higher is better; unit: PERCENT. - [Olympiad Bench F1 Score](): Higher is better; unit: PERCENT. - [Omni Math Error Rate](): Lower is better; unit: PERCENT. - [Omni Math Accuracy](): Higher is better; unit: PERCENT. - [Omni Math F1 Score](): Higher is better; unit: PERCENT. - [Average F1 Score](): Higher is better; unit: PERCENT. ## Recorded results | Method | GSM8K Error Rate | Result date | Report paper | | --- | --- | --- | --- | | RLHFlow-PRM-Deepseek-8B | 0.242 | Not recorded | 69705d92e25ad2fa9d1ed134 | | Math-Shepherd-PRM-7B | 0.324 | Not recorded | 69705d92e25ad2fa9d1ed134 | | Qwen2.5-Math-7B-MC | 0.332 | Not recorded | 69705d92e25ad2fa9d1ed134 | | RLHFlow-PRM-Mistral-8B | 0.338 | Not recorded | 69705d92e25ad2fa9d1ed134 | | EurusPRM-Stage1 | 0.469 | Not recorded | 69705d92e25ad2fa9d1ed134 | | EurusPRM-Stage2 | 0.512 | Not recorded | 69705d92e25ad2fa9d1ed134 | | Qwen2.5-Math-7B-MCRD | 0.556 | Not recorded | 69705d92e25ad2fa9d1ed134 | | Skywork-PRM-Qwen2.5-7B | 0.618 | Not recorded | 69705d92e25ad2fa9d1ed134 | | Qwen2.5-72B-Instruct | 0.628 | Not recorded | 69705d92e25ad2fa9d1ed134 | | Qwen2.5-Math-7B-NAIT | 0.655 | Not recorded | 69705d92e25ad2fa9d1ed134 | | GPT-4o-0806 | 0.7 | Not recorded | 69705d92e25ad2fa9d1ed134 | | Qwen2.5-Math-PRM-7B | 0.72 | Not recorded | 69705d92e25ad2fa9d1ed134 | ### RLHFlow-PRM-Deepseek-8B GSM8K Error Rate: 0.242. Result ID: 69967c0c5698665ec4c3cf61. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"MCE-based PRM","Backbone":"Deepseek-8B"}. This record does not assert independent reproduction. ### Math-Shepherd-PRM-7B GSM8K Error Rate: 0.324. Result ID: 69967c0c5698665ec4c3cf5f. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"MCE-based PRM","Backbone":"7B"}. This record does not assert independent reproduction. ### Qwen2.5-Math-7B-MC GSM8K Error Rate: 0.332. Result ID: 69967c0c5698665ec4c3cf62. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"MCE-based PRM","Labeling":"Monte Carlo Estimation","Backbone":"7B"}. This record does not assert independent reproduction. ### RLHFlow-PRM-Mistral-8B GSM8K Error Rate: 0.338. Result ID: 69967c0c5698665ec4c3cf60. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"MCE-based PRM","Backbone":"Mistral-8B"}. This record does not assert independent reproduction. ### EurusPRM-Stage1 GSM8K Error Rate: 0.469. Result ID: 69967c0c5698665ec4c3cf66. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"Other PRM Baseline","Stage":"Stage 1"}. This record does not assert independent reproduction. ### EurusPRM-Stage2 GSM8K Error Rate: 0.512. Result ID: 69967c0c5698665ec4c3cf67. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"Other PRM Baseline","Stage":"Stage 2"}. This record does not assert independent reproduction. ### Qwen2.5-Math-7B-MCRD GSM8K Error Rate: 0.556. Result ID: 69967c0c5698665ec4c3cf64. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"Filtered-integrated PRM","Labeling":"Reflection-aware label correction","Backbone":"7B"}. This record does not assert independent reproduction. ### Skywork-PRM-Qwen2.5-7B GSM8K Error Rate: 0.618. Result ID: 69967c0c5698665ec4c3cf65. [Reporting paper]() Recorded evaluation configuration: {"Category":"Other PRM Baseline"}. This record does not assert independent reproduction. ### Qwen2.5-72B-Instruct GSM8K Error Rate: 0.628. Result ID: 69967c0c5698665ec4c3cf5e. [Reporting paper]() Recorded evaluation configuration: {"Category":"Critic Model (LLM-as-a-judge)"}. This record does not assert independent reproduction. ### Qwen2.5-Math-7B-NAIT GSM8K Error Rate: 0.655. Result ID: 69967c0c5698665ec4c3cf68. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"Other PRM Baseline","Training":"Noise-Aware Iterative Training","Backbone":"7B"}. This record does not assert independent reproduction. ### GPT-4o-0806 GSM8K Error Rate: 0.7. Result ID: 69967c0c5698665ec4c3cf5d. [Reporting paper]() Recorded evaluation configuration: {"Category":"Critic Model (LLM-as-a-judge)"}. This record does not assert independent reproduction. ### Qwen2.5-Math-PRM-7B GSM8K Error Rate: 0.72. Result ID: 69967c0c5698665ec4c3cf63. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Category":"Filtered-integrated PRM"}. This record does not assert independent reproduction. Page 1 of 1. This is an index page, not the full collection.