Multi-task Evaluation on PostTrainBench
29.17AIME 25 ScoreOfficial Instruct Models
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Official Instruct ModelsEvaluation Setting=Official Instruct Models2026.05 | 29.17 | 70.21 | 85 | 36.21 | 87 | 43.32 | 71.49 | 51.14 | |
| Official Instruct ModelProtocol=Evaluation with Official Instruct Model, Backbone=Gemma3-4B2026.05 | 10 | 94.8 | 67 | 31.47 | 83.55 | 46.06 | 69.51 | 46.58 | |
| Human ScoreBackbone=Qwen3-1.7B2026.05 | 6.66 | 44.84 | 63.46 | 30.35 | 86.95 | 35.76 | 60.36 | 46.91 | |
| Opus-4.7 (xHigh)Evaluation Setting=Agent-Post-Trained Base Models2026.05 | 6.39 | 24.2 | 76.75 | 27.48 | 59.01 | 16.53 | 42.07 | 28.56 | |
| Opus-4.6Protocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 5.56 | 6.34 | 96.67 | 25.3 | 52.19 | 22.87 | 36.59 | 27.71 | |
| Opus-4.6Evaluation Setting=Agent-Post-Trained Base Models2026.05 | 5 | 7.78 | 75.92 | 25.52 | 41.04 | 18.81 | 24.75 | 23.16 | |
| GLM-4.7 (OpenCode)Protocol=Evaluation with Proposed Method (Compared with Baseline), Backbone=Gemma3-4B2026.05 | 4.6 | 0 | 0.29 | 6 | 1.56 | 6.14 | 17.04 | 0.61 | |
| GLM-4.7 & ANDES (Ours)Evaluation Setting=Proposed Method (Compared with Baseline)2026.05 | 3.9 | 25.66 | 85.67 | 27.12 | 72.04 | 31.51 | 48.74 | 33.39 | |
| Gemini-3.1-ProEvaluation Setting=Agent-Post-Trained Base Models2026.05 | 3.89 | 7.42 | 62.84 | 18.53 | 45.51 | 14.48 | 40.19 | 21.59 | |
| ML-Master 2.0Backbone=GLM-52026.05 | 3.33 | 0 | 31.43 | 18.57 | 31.45 | 8.24 | 16.84 | 15.69 | |
| DataMasterBackbone=GLM-52026.05 | 3.33 | 21.93 | 34.27 | 31.02 | 49.43 | 34.93 | 43.29 | 31.17 | |
| GPT 5.4 (High)Protocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 3.33 | 0.29 | 100 | 26.34 | 44.2 | 21.9 | 23.78 | 24.85 | |
| Opus-4.6 (1M)Evaluation Setting=Agent-Post-Trained Base Models2026.05 | 3.33 | 6.73 | 77.16 | 27.29 | 51.27 | 15.3 | 37.25 | 24.82 | |
| Base Models (Average)Evaluation Setting=Base Models (Zero-Shot)2026.05 | 1.67 | 1.26 | 1.5 | 8.48 | 20.43 | 9.49 | 12.81 | 7.53 | |
| Kimi-K2-ThinkingEvaluation Setting=Agent-Post-Trained Base Models2026.05 | 1.67 | 1.26 | 1.5 | 8.48 | 14.84 | 9.49 | 15.09 | 7.25 | |
| GPT-5.1-Codex-MaxEvaluation Setting=Agent-Post-Trained Base Models2026.05 | 1.67 | 1.07 | 1.5 | 15.35 | 20 | 6.14 | 5.79 | 7.65 | |
| GLM-4.7 (OpenCode)Evaluation Setting=Proposed Method (Compared with Baseline)2026.05 | 1.67 | 1.26 | 1.5 | 8.48 | 18.76 | 9.49 | 13.88 | 7.48 | |
| Gemini-3.1-ProProtocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 1.11 | 16.01 | 80.67 | 19.79 | 39.5 | 19.7 | 34.55 | 23.15 | |
| Opus-4.7 (xHigh)Protocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 1.11 | 30.86 | 91.33 | 28.72 | 55.7 | 21.49 | 27.85 | 28.43 | |
| Qwen3-MaxEvaluation Setting=Agent-Post-Trained Base Models2026.05 | 0.83 | 0.96 | 1.5 | 7.14 | 20.62 | 9.49 | 16.46 | 7.42 | |
| MiniMax-M2.1Evaluation Setting=Agent-Post-Trained Base Models2026.05 | 0.83 | 1.26 | 13.5 | 9.65 | 19.35 | 9.49 | 21.65 | 9.33 | |
| Sonnet-4.5Evaluation Setting=Agent-Post-Trained Base Models2026.05 | 0.83 | 1.04 | 1.75 | 14.62 | 30.86 | 4.96 | 23.02 | 9.94 | |
| GPT-5.2Evaluation Setting=Agent-Post-Trained Base Models2026.05 | 0.83 | 6.61 | 52.5 | 23.72 | 55.9 | 15.81 | 30.23 | 21.38 | |
| GPT-5.4 (High)Evaluation Setting=Agent-Post-Trained Base Models2026.05 | 0.56 | 10.07 | 31.09 | 27.98 | 48.18 | 17.29 | 27.34 | 20.23 | |
| Initial ScoreBackbone=Qwen3-1.7B-Base2026.05 | 0 | 0.48 | 22.56 | 18.75 | 9.02 | 0 | 8.53 | 8.47 | |
| Codex*Backbone=GPT-5.2-Codex2026.05 | 0 | 0.15 | 40.67 | 29.54 | 24.31 | 9.17 | 25.41 | 18.46 | |
| Claude CodeBackbone=GLM-52026.05 | 0 | 2.01 | 31 | 14.29 | 2.58 | 1.16 | 29.27 | 11.47 | |
| DatasetResearcherBackbone=GLM-52026.05 | 0 | 0 | 22.64 | 18 | 9.33 | 3.56 | 6 | 8.5 | |
| DataFlexBackbone=GLM-52026.05 | 0 | 2.07 | 22.58 | 28.8 | 48.37 | 6.22 | 25.61 | 19.09 | |
| Base Model (Gemma3-4B)Protocol=Evaluation with Base Model (Zero-Shot), Mode=Zero-Shot2026.05 | 0 | 0.29 | 6 | 1.56 | 6.14 | 17.04 | 0.61 | 4.6 | |
| Kimi-K2-ThinkingProtocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 0 | 0.29 | 6 | 1.56 | 6.14 | 17.04 | 0.61 | 4.6 | |
| Qwen3-MaxProtocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 0 | 0.29 | 6 | 1.56 | 6.14 | 17.04 | 0.61 | 4.6 | |
| MiniMax-M2.5Protocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 0 | 3.6 | 9 | 1.56 | 36.39 | 18.34 | 3.05 | 8.45 | |
| MiniMax-M2.1Protocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 0 | 0.29 | 54 | 1.56 | 2.96 | 17.04 | 30.49 | 11.06 | |
| Sonnet-4.5Protocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 0 | 0.29 | 5 | 25.22 | 57.92 | 10.69 | 34.76 | 17.14 | |
| GPT-5.1-Codex MaxProtocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 0 | 8.45 | 33.67 | 21.21 | 44 | 20.13 | 33.54 | 19.42 | |
| Opus-4.6 (1M)Protocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 0 | 13.95 | 64.33 | 28.12 | 36.42 | 21.49 | 36.38 | 23.6 | |
| GPT-5.2Protocol=Evaluation with Agent-Post-Trained Base Model, Backbone=Gemma3-4B2026.05 | 0 | 13.2 | 89 | 28.57 | 51.91 | 20.19 | 21.54 | 25.11 | |
| GLM-4.7 (Scaffold-only)Protocol=Evaluation with Proposed Method (Compared with Baseline), Backbone=Gemma3-4B2026.05 | 0 | 31.12 | 61 | 22.76 | 36.16 | 4.89 | 28.29 | 19.76 | |
| GLM-4.7 & ANDES (Ours)Protocol=Evaluation with Proposed Method (Compared with Baseline), Backbone=Gemma3-4B2026.05 | 0 | 25.16 | 85 | 25.42 | 76.57 | 33.28 | 32.11 | 31.02 | |
| MiniMax-M2.5Evaluation Setting=Agent-Post-Trained Base Models2026.05 | 0 | 2.74 | 2.25 | 11.55 | 31.01 | 10.51 | 15.55 | 9.5 |