Reward Modeling on RM-Bench (test)
84.11Chat ScoreEval-Skill (workflow-only)
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Eval-Skill (workflow-only)Judge Backbone=Qwen3.6-35B-A3B, Evaluation Strategy=workflow-only2026.06 | 84.11 | 96.17 | 87.5 | 94.69 | — | 90.62 | — | — | — | |
| Eval-Skill (full)Judge Backbone=Qwen3.6-35B-A3B, Evaluation Strategy=full2026.06 | 82.28 | 96.37 | 85.59 | 95.73 | — | 89.99 | — | — | — | |
| Eval-Skill (adaptive)Judge Backbone=Qwen3.6-35B-A3B, Evaluation Strategy=adaptive2026.06 | 82.28 | 96.37 | 85.59 | 95.73 | — | 89.99 | — | — | — | |
| Principles-Qwen32BModel Category=Generative Reward Model2026.02 | 80.4 | 92 | 77 | 95.5 | 86.2 | 83.8 | — | — | — | |
| DirectJudge Backbone=Qwen3.6-35B-A3B, Evaluation Strategy=Direct2026.06 | 77.78 | 89.74 | 80.9 | 94.17 | — | 85.65 | — | — | — | |
| Eval-Skill (workflow-only)Judge Backbone=Qwen3-8B-think, Evaluation Strategy=workflow-only2026.06 | 77.64 | 94.82 | 75.17 | 91.53 | — | 84.79 | — | — | — | |
| Eval-Skill (full)Judge Backbone=Qwen3-8B-think, Evaluation Strategy=full2026.06 | 77.22 | 94.95 | 76.22 | 91.82 | — | 85.05 | — | — | — | |
| Eval-Skill (adaptive)Judge Backbone=Qwen3-8B-think, Evaluation Strategy=adaptive2026.06 | 77.22 | 94.95 | 76.22 | 91.82 | — | 85.05 | — | — | — | |
| Eval-Skill (workflow-only)Judge Backbone=Gemma-4-E4B-IT, Evaluation Strategy=workflow-only2026.06 | 76.79 | 88.58 | 79.86 | 93.74 | — | 84.74 | — | — | — | |
| Eval-Skill (adaptive)Judge Backbone=Gemma-4-E4B-IT, Evaluation Strategy=adaptive2026.06 | 76.79 | 90.65 | 79.86 | 94.2 | — | 85.38 | — | — | — | |
| DeepSeek-R1-0528Model Category=LLM-as-a-Judge2026.02 | 76.7 | 74.3 | 51 | 89.2 | 72.8 | 75.8 | — | — | — | |
| RewardAnything-8B-v1Model Category=Generative Reward Model2026.02 | 76.7 | 90.3 | 75.2 | 90.2 | 83.1 | 72.9 | — | — | — | |
| GRAM-R²Model Category=Generative Reward Model2026.02 | 76 | 89.8 | 80.6 | 96.2 | 85.7 | 83.4 | — | — | — | |
| Qwen3-14BSupervision Signal=Outcome and Rationale Supervision2026.02 | 75.7 | 95.7 | 82.2 | 93.2 | 86.7 | 82.9 | — | — | — | |
| Eval-Skill (full)Judge Backbone=Gemma-4-E4B-IT, Evaluation Strategy=full2026.06 | 74.96 | 90.65 | 79.43 | 94.2 | — | 84.81 | — | — | — | |
| Qwen3-30B-A3BSupervision Signal=Outcome and Rationale Supervision2026.02 | 74.9 | 95.5 | 84.4 | 93.6 | 87.1 | 84.6 | — | — | — | |
| RM-R1-Distilled-Qwen-32BModel Category=Generative Reward Model2026.02 | 74.2 | 91.8 | 74.1 | 95.4 | 83.9 | 81.4 | — | — | — | |
| Scaled BTTraining Dataset=HelpSteer3, Backbone=Llama2026.02 | 74.2 | 59.2 | 53.4 | 82.3 | — | 67.3 | 47.6 | 69.7 | 84.4 | |
| Nemotron-SuperModel Category=Generative Reward Model2026.02 | 73.7 | 91.4 | 75 | 90.6 | 82.7 | 80 | — | — | — | |
| Qwen3-14B (Outcome-Only)Supervision Signal=Outcome-Only2026.02 | 72.3 | 92.6 | 77.8 | 91.7 | 83.6 | 76.8 | — | — | — | |
| Skywork-Reward-Gemma-2-27BModel Category=Scalar Reward Model2026.02 | 71.8 | 59.2 | 56.6 | 94.3 | 70.5 | 67.8 | — | — | — | |
| RM-R1-Distilled-Qwen-14BModel Category=Generative Reward Model2026.02 | 71.8 | 90.5 | 69.5 | 94.1 | 81.5 | 79.8 | — | — | — | |
| Margin BTTraining Dataset=HelpSteer3, Backbone=Llama2026.02 | 71.7 | 58.7 | 53.2 | 76.5 | — | 65 | 39 | 68.1 | 88 | |
| Soft LabelTraining Dataset=HelpSteer3, Backbone=Llama2026.02 | 70.8 | 57.8 | 53.6 | 86.2 | — | 67.1 | 42.1 | 69.8 | 89.4 | |
| Margin BTTraining Dataset=HelpSteer3, Backbone=Mistral2026.02 | 70 | 55.5 | 50.6 | 71.8 | — | 62 | 41.4 | 63.6 | 80.9 | |
| Skywork-Reward-Llama-3.1-8BModel Category=Scalar Reward Model2026.02 | 69.5 | 60.6 | 54.5 | 95.7 | 70.1 | 66.3 | — | — | — | |
| NLL-SymTraining Dataset=HelpSteer3, Backbone=Llama2026.02 | 69.5 | 58.6 | 54.1 | 88.1 | — | 67.6 | 41.3 | 71.3 | 80.2 | |
| Scaled BTTraining Dataset=HelpSteer3, Backbone=Zephyr2026.02 | 69.4 | 55.1 | 49.2 | 69.3 | — | 60.7 | 35.5 | 63.1 | 83.7 | |
| Scaled BTTraining Dataset=HelpSteer3, Backbone=Mistral2026.02 | 69.2 | 55.1 | 50.1 | 59.3 | — | 58.4 | 36.1 | 59.7 | 79.4 | |
| All-ThreshTraining Dataset=HelpSteer2, Backbone=Llama2026.02 | 68.8 | 56.6 | 51.9 | 86 | — | 65.8 | 42.1 | 68.1 | 87.4 | |
| Qwen3-30B-A3B (Outcome-Only)Supervision Signal=Outcome-Only2026.02 | 68.7 | 95.9 | 81.3 | 93.6 | 84.9 | 80.3 | — | — | — | |
| NLL-AsymTraining Dataset=HelpSteer3, Backbone=Llama2026.02 | 68.4 | 58.7 | 56.9 | 81.5 | — | 66.6 | 53.9 | 68.3 | 77.6 | |
| NLL-SymTraining Dataset=HelpSteer3, Backbone=Zephyr2026.02 | 68.4 | 54.8 | 50.2 | 72.7 | — | 61.5 | 35.7 | 64 | 84.7 | |
| GPT-4oModel Category=LLM-as-a-Judge2026.02 | 67.2 | 67.5 | 63.6 | 91.7 | 72.5 | 66.2 | — | — | — | |
| Margin BTTraining Dataset=HelpSteer3, Backbone=Zephyr2026.02 | 66.8 | 55.9 | 53.2 | 67.8 | — | 60.9 | 38.6 | 62.9 | 81.3 | |
| RRM-32BModel Category=Generative Reward Model2026.02 | 66.6 | 81.4 | 65.2 | 79.4 | 73.1 | 74.4 | — | — | — | |
| DirectJudge Backbone=Qwen3-8B-think, Evaluation Strategy=Direct2026.06 | 66.53 | 70.71 | 59.38 | 90.03 | — | 71.66 | — | — | — | |
| All-ThreshTraining Dataset=HelpSteer3, Backbone=Llama2026.02 | 66.3 | 57.3 | 54 | 81.5 | — | 64.8 | 52.6 | 66.3 | 75.5 | |
| Margin BTTraining Dataset=HelpSteer2, Backbone=Llama2026.02 | 66.2 | 56.3 | 52.1 | 88.5 | — | 65.7 | 36.3 | 70.1 | 90.8 | |
| NLL-SymTraining Dataset=HelpSteer2, Backbone=Zephyr2026.02 | 66.1 | 53.3 | 51.9 | 83.5 | — | 63.7 | 39.9 | 66.7 | 84.6 | |
| NLL-SymTraining Dataset=HelpSteer2, Backbone=Mistral2026.02 | 65.4 | 53.8 | 54 | 85.8 | — | 64.7 | 40.6 | 67.3 | 86.4 | |
| Soft LabelTraining Dataset=HelpSteer3, Backbone=Zephyr2026.02 | 65.3 | 56.4 | 52.5 | 60.5 | — | 58.7 | 31.8 | 61.8 | 82.5 | |
| DirectJudge Backbone=Gemma-4-E4B-IT, Evaluation Strategy=Direct2026.06 | 64.84 | 78.27 | 69.79 | 91.53 | — | 76.11 | — | — | — | |
| Soft LabelTraining Dataset=HelpSteer3, Backbone=Mistral2026.02 | 64.3 | 55.6 | 49.4 | 63.8 | — | 58.3 | 47.4 | 58.9 | 68.6 | |
| NLL-AsymTraining Dataset=HelpSteer3, Backbone=Zephyr2026.02 | 64.3 | 55.4 | 33.6 | 51.9 | — | 56.3 | 30.3 | 56.9 | 81.7 | |
| NLL-SymTraining Dataset=HelpSteer2, Backbone=Llama2026.02 | 63.8 | 55 | 50.8 | 90.7 | — | 65.1 | 48.7 | 66.8 | 79.8 | |
| NLL-SymTraining Dataset=HelpSteer3, Backbone=Mistral2026.02 | 63.8 | 53.5 | 50.4 | 72.5 | — | 60.1 | 29.4 | 63.1 | 87.8 | |
| Scaled BTTraining Dataset=HelpSteer2, Backbone=Llama2026.02 | 63.6 | 55.2 | 49.7 | 79.8 | — | 62.1 | 38.9 | 64 | 83.2 | |
| Soft LabelTraining Dataset=HelpSteer2, Backbone=Zephyr2026.02 | 63.5 | 53.1 | 50.4 | 84 | — | 62.7 | 34.9 | 65.8 | 87.5 | |
| Margin BTTraining Dataset=HelpSteer2, Backbone=Mistral2026.02 | 63.4 | 53.9 | 49.6 | 76 | — | 60.7 | 31.7 | 64.3 | 86.2 | |
| Eval-Skill (full)Judge Backbone=Llama-3.1-8B-Instruct, Evaluation Strategy=full2026.06 | 63.01 | 57.24 | 49.83 | 69.99 | — | 60.02 | — | — | — | |
| Eval-Skill (adaptive)Judge Backbone=Llama-3.1-8B-Instruct, Evaluation Strategy=adaptive2026.06 | 63.01 | 59.44 | 47.57 | 69.99 | — | 60 | — | — | — | |
| All-ThreshTraining Dataset=HelpSteer2, Backbone=Zephyr2026.02 | 62.8 | 53.4 | 49.8 | 79.6 | — | 61.4 | 32.9 | 63.8 | 87.5 | |
| All-ThreshTraining Dataset=HelpSteer3, Backbone=Zephyr2026.02 | 62.8 | 55.2 | 52.1 | 51.3 | — | 55.3 | 29.7 | 57 | 79.4 | |
| Claude-3.5-SonnetModel Category=LLM-as-a-Judge2026.02 | 62.5 | 62.6 | 54.4 | 64.4 | 61 | 62.9 | — | — | — | |
| Eval-Skill (workflow-only)Judge Backbone=RRM-7B, Evaluation Strategy=workflow-only2026.06 | 62.45 | 78.43 | 54.34 | 84.62 | — | 69.96 | — | — | — | |
| DirectJudge Backbone=RRM-7B, Evaluation Strategy=Direct2026.06 | 62.31 | 75.24 | 53.73 | 80.06 | — | 67.84 | — | — | — | |
| NLL-AsymTraining Dataset=HelpSteer3, Backbone=Mistral2026.02 | 62.2 | 58 | 52.2 | 70.6 | — | 60.8 | 39.2 | 63.3 | 79.8 | |
| Scaled BTTraining Dataset=HelpSteer2, Backbone=Zephyr2026.02 | 62.1 | 53.4 | 50.1 | 76.2 | — | 60.5 | 32.7 | 63.7 | 85 | |
| All-ThreshTraining Dataset=HelpSteer3, Backbone=Mistral2026.02 | 62.1 | 56 | 52.3 | 55.5 | — | 56.5 | 25.5 | 59.9 | 84.1 | |
| All-ThreshTraining Dataset=HelpSteer2, Backbone=Mistral2026.02 | 61.9 | 51.9 | 49.3 | 77.5 | — | 60.1 | 31.5 | 63 | 85.9 | |
| Margin BTTraining Dataset=HelpSteer2, Backbone=Zephyr2026.02 | 61.5 | 51.9 | 51.1 | 85.6 | — | 62.5 | 39.2 | 64.2 | 84.1 | |
| Scaled BTTraining Dataset=HelpSteer2, Backbone=Mistral2026.02 | 61.3 | 54.8 | 52 | 83.7 | — | 63 | 43.9 | 65.4 | 79.5 | |
| NLL-AsymTraining Dataset=HelpSteer2, Backbone=Zephyr2026.02 | 60.4 | 53 | 50.9 | 65 | — | 57.3 | 23 | 61.7 | 87.3 | |
| Eval-Skill (full)Judge Backbone=RRM-7B, Evaluation Strategy=full2026.06 | 58.79 | 80.78 | 56.68 | 83.87 | — | 70.03 | — | — | — | |
| Eval-Skill (adaptive)Judge Backbone=RRM-7B, Evaluation Strategy=adaptive2026.06 | 58.79 | 80.78 | 56.68 | 83.87 | — | 70.03 | — | — | — | |
| NLL-AsymTraining Dataset=HelpSteer2, Backbone=Mistral2026.02 | 58.4 | 53.9 | 48.3 | 71.6 | — | 58 | 28 | 61.7 | 84.5 | |
| Eval-Skill (workflow-only)Judge Backbone=Llama-3.1-8B-Instruct, Evaluation Strategy=workflow-only2026.06 | 58.37 | 59.44 | 47.57 | 67.64 | — | 58.26 | — | — | — | |
| NLL-AsymTraining Dataset=HelpSteer2, Backbone=Llama2026.02 | 57.4 | 54.4 | 52.3 | 82.3 | — | 61.6 | 39.7 | 63.1 | 81.9 | |
| Soft LabelTraining Dataset=HelpSteer2, Backbone=Llama2026.02 | 57.1 | 55 | 50.5 | 72.7 | — | 58.8 | 22.5 | 63.8 | 90.2 | |
| DirectJudge Backbone=Llama-3.1-8B-Instruct, Evaluation Strategy=Direct2026.06 | 52.74 | 59.52 | 50.43 | 38.09 | — | 50.2 | — | — | — | |
| Soft LabelTraining Dataset=HelpSteer2, Backbone=Mistral2026.02 | 52.5 | 50.1 | 49.1 | 62.3 | — | 53.5 | 26.5 | 53.7 | 80.3 | |
| ArmoRM-Llama3-8B-v0.1Model Type=Open Reward Model2025.07 | — | — | — | — | 69.2 | — | — | — | — | |
| CE-RM-4BRM Type=Pointwise, Scaling=12026.01 | — | — | — | — | 79.8 | — | — | — | — | |
| CE-RM-4BRM Type=Pointwise, Scaling=22026.01 | — | — | — | — | 81.9 | — | — | — | — | |
| CE-RM-4BRM Type=Pointwise, Scaling=42026.01 | — | — | — | — | 83.2 | — | — | — | — | |
| Claude-3.5-SonnetModel Type=LLM-as-a-Judge2025.07 | — | — | — | — | 74.5 | — | — | — | — | |
| CompassJudger1-32BRM Type=Pointwise2026.01 | — | — | — | — | 54.4 | — | — | — | — | |
| CompassJudger2-32BRM Type=Pairwise2026.01 | — | — | — | — | 73.2 | — | — | — | — | |
| CompassJudger2-7BRM Type=Pairwise2026.01 | — | — | — | — | 71.5 | — | — | — | — | |
| EvalPlanner (Llama-3.1-70B)Model Type=LLM-as-a-Judge2025.07 | — | — | — | — | 80 | — | — | — | — | |
| EvalPlanner (Llama-3.3-70B)Model Type=LLM-as-a-Judge2025.07 | — | — | — | — | 82.1 | — | — | — | — | |
| Gemini-2.5-FlashRM Type=Pointwise2026.01 | — | — | — | — | 70.8 | — | — | — | — | |
| GPT-4oRM Type=Pairwise2026.01 | — | — | — | — | 72.5 | — | — | — | — | |
| GPT-4oModel Type=LLM-as-a-Judge2025.07 | — | — | — | — | 73.1 | — | — | — | — | |
| INF-ORM-Llama3.1-70BModel Type=Open Reward Model2025.07 | — | — | — | — | 75.4 | — | — | — | — | |
| Internlm2-20b-rewardModel Type=Open Reward Model2025.07 | — | — | — | — | 72.1 | — | — | — | — | |
| J1-Llama-70BRM Type=Pairwise2026.01 | — | — | — | — | 82.7 | — | — | — | — | |
| J1-Llama-70BModel Type=LLM-as-a-Judge2025.07 | — | — | — | — | 82.7 | — | — | — | — | |
| J1-Llama-8BRM Type=Pairwise2026.01 | — | — | — | — | 73.4 | — | — | — | — | |
| J1-Llama-8BModel Type=LLM-as-a-Judge2025.07 | — | — | — | — | 73.4 | — | — | — | — | |
| JudgeLRM-7BRM Type=Pairwise2026.01 | — | — | — | — | 58.7 | — | — | — | — | |
| LDL-Reward-Gemma-2-27B-v0.1Model Type=Open Reward Model2025.07 | — | — | — | — | 71 | — | — | — | — | |
| Llama-3-OffsetBias-RM-8BModel Type=Open Reward Model2025.07 | — | — | — | — | 71.3 | — | — | — | — | |
| Llama-3.1-Nemotron-70BModel Type=Open Reward Model2025.07 | — | — | — | — | 72.2 | — | — | — | — | |
| Llama-3.3-70B-InstructRM Type=Pairwise2026.01 | — | — | — | — | 69.5 | — | — | — | — | |
| RISE-Judge-32BRM Type=Pairwise2026.01 | — | — | — | — | 72.3 | — | — | — | — | |
| RISE-Judge-7BRM Type=Pairwise2026.01 | — | — | — | — | 68 | — | — | — | — | |
| RM-R1-DeepSeek-Distill-Qwen-32BModel Type=Generative Reward Model2025.07 | — | — | — | — | 83.9 | — | — | — | — | |
| RM-R1-Qwen-32BRM Type=Pairwise2026.01 | — | — | — | — | 79.1 | — | — | — | — |