## Radar Chart: Model Performance Across Benchmarks ### Overview A radar chart comparing model performance across four benchmarks: AIME25, MATH500, GPQA-Diamond, and AIME24. Five data series are represented with distinct line styles and colors. ### Components/Axes - **Axes**: - Top: AIME25 - Left: MATH500 - Bottom: GPQA-Diamond - Right: AIME24 - **Legend**: - Orange dashed: Human Curated - Dashed green: Skywork-PRM-7B - Dotted yellow: Random - Solid red: Qwen2.5-Math-PRM-7B - Solid purple: ReasonFlux-PRM-7B - **Scale**: 0.0 to 1.0 (outer ring) ### Detailed Analysis - **Human Curated** (orange dashed): Outermost polygon, consistently highest values (~0.8–1.0 across all axes). - **ReasonFlux-PRM-7B** (solid purple): Second-largest polygon, values ~0.6–0.9. - **Qwen2.5-Math-PRM-7B** (solid red): Third-largest, values ~0.5–0.8. - **Skywork-PRM-7B** (dashed green): Values ~0.4–0.7. - **Random** (dotted yellow): Innermost polygon, values ~0.2–0.5. ### Key Observations - Human Curated dominates all benchmarks, suggesting it represents a human performance baseline. - ReasonFlux-PRM-7B consistently outperforms Qwen2.5-Math-PRM-7B and Skywork-PRM-7B. - Random performs significantly worse than all trained models. ### Interpretation The radar chart demonstrates that ReasonFlux-PRM-7B achieves the closest performance to the human-curated benchmark across all tasks, while Qwen2.5-Math-PRM-7B and Skywork-PRM-7B show moderate performance. The Random baseline highlights the effectiveness of the models compared to chance. --- ## Line Graph: Training Reward Over Steps ### Overview A line graph tracking training reward over 180 steps for three models: GRPO, Qwen2.5-Math-PRM-7B, and ReasonFlux-PRM-7B. ### Components/Axes - **X-axis**: Steps (0 to 180) - **Y-axis**: Training Reward (0.1 to 0.4) - **Legend**: - Blue squares: GRPO - Orange triangles: Qwen2.5-Math-PRM-7B - Purple stars: ReasonFlux-PRM-7B ### Detailed Analysis - **ReasonFlux-PRM-7B** (purple stars): - Starts at ~0.28, peaks at ~0.42 by step 180. - Smooth upward trend with minor fluctuations. - **Qwen2.5-Math-PRM-7B** (orange triangles): - Begins at ~0.15, reaches ~0.32 by step 180. - Noisy with oscillations but generally increasing. - **GRPO** (blue squares): - Starts at ~0.05, rises to ~0.30 by step 180. - Steeper initial growth but plateaus earlier. ### Key Observations - ReasonFlux-PRM-7B achieves the highest final reward and maintains stability. - Qwen2.5-Math-PRM-7B shows moderate performance with higher volatility. - GRPO improves rapidly but lags behind the other two models in final reward. ### Interpretation The graph indicates that ReasonFlux-PRM-7B has the most stable and effective training dynamics, while Qwen2.5-Math-PRM-7B and GRPO exhibit trade-offs between growth speed and stability. --- ## Scatter Plot: Accuracy vs. Number of Solutions (GPQA-Diamond) ### Overview A scatter plot showing accuracy (%) against the number of solutions (N = 2¹ to 2⁴) for four models. ### Components/Axes - **X-axis**: Number of Solutions (N) (2¹ to 2⁴) - **Y-axis**: Accuracy (%) (48% to 54%) - **Legend**: - Red triangles: Qwen2.5-Math-PRM-7B - Green dashed: Skywork-PRM-7B - Blue squares: Majority - Purple stars: ReasonFlux-PRM-7B ### Detailed Analysis - **ReasonFlux-PRM-7B** (purple stars): - Accuracy increases from ~48% (N=2¹) to ~54% (N=2⁴). - Steep upward trend. - **Qwen2.5-Math-PRM-7B** (red triangles): - Accuracy rises from ~48% to ~53%. - Slightly less steep than ReasonFlux. - **Skywork-PRM-7B** (green dashed): - Accuracy increases from ~48% to ~51%. - Flatter growth. - **Majority** (blue squares): - Flat line at ~48% across all N values. ### Key Observations - ReasonFlux-PRM-7B shows the strongest improvement with more solutions. - Majority baseline remains constant, indicating no inherent model capability beyond random guessing. - Qwen2.5-Math-PRM-7B and Skywork-PRM-7B show moderate gains. ### Interpretation The scatter plot reveals that ReasonFlux-PRM-7B scales most effectively with increased computational resources (solutions), suggesting superior architectural efficiency. The Majority baseline underscores the importance of model training over brute-force methods. --- ## Cross-Referenced Trends 1. **Consistency Across Metrics**: ReasonFlux-PRM-7B outperforms all models in the radar chart, line graph, and scatter plot. 2. **Training Dynamics**: ReasonFlux-PRM-7B achieves higher rewards faster and more stably than Qwen2.5-Math-PRM-7B and GRPO. 3. **Scalability**: ReasonFlux-PRM-7B benefits most from increased solution counts (N), indicating better generalization. ## Conclusion The data collectively demonstrates that ReasonFlux-PRM-7B is the most performant model across training stability, benchmark accuracy, and scalability. Qwen2.5-Math-PRM-7B and Skywork-PRM-7B show moderate performance, while GRPO and Majority lag behind. Human Curated remains the gold standard, but ReasonFlux-PRM-7B approaches it most closely.