## Line Chart: Validation Accuracy Comparison
### Overview
The chart compares the validation accuracy of two machine learning models (GRPO and AIRL-S) across 200 training steps. Two lines represent model performance: blue for GRPO (without PRM) and red for AIRL-S (with PRM). Both models show increasing accuracy over time, with AIRL-S consistently outperforming GRPO after the initial steps.
### Components/Axes
- **X-axis (Step)**: Training progression from 0 to 200 steps, marked at intervals of 50.
- **Y-axis (Accuracy)**: Ranges from 0.38 to 0.44, with increments of 0.02.
- **Legend**: Located at the bottom-right corner, associating:
- **Blue line**: GRPO (w/o PRM)
- **Red line**: AIRL-S (w. PRM)
### Detailed Analysis
1. **GRPO (w/o PRM) [Blue Line]**:
- Starts at 0.38 accuracy at step 0.
- Gradual increase to ~0.415 by step 200.
- Minor fluctuations observed between steps 100–150 (e.g., slight dip at step 150).
- Plateau observed after step 150, stabilizing near 0.418.
2. **AIRL-S (w. PRM) [Red Line]**:
- Starts at 0.38 accuracy at step 0.
- Sharp initial rise to 0.425 by step 100.
- Continued upward trend to ~0.438 by step 200.
- Slight plateau observed after step 150, stabilizing near 0.438.
### Key Observations
- AIRL-S (red) consistently outperforms GRPO (blue) after step 50.
- Both models show diminishing returns after ~150 steps.
- AIRL-S achieves a final accuracy ~0.023 higher than GRPO at step 200.
- GRPO exhibits more volatility in the 50–150 step range compared to AIRL-S.
### Interpretation
The data demonstrates that incorporating PRM (Proximal Regularization Method) in AIRL-S significantly improves validation accuracy compared to GRPO without PRM. The steeper ascent and higher plateau of the red line suggest PRM enhances model stability and convergence. The plateauing behavior indicates that both models reach practical limits of performance beyond 150 steps, with AIRL-S maintaining a clear advantage. This trend highlights the importance of regularization techniques in optimizing model accuracy for validation tasks.
