## Line Charts: Average Accuracies Across Similarity Thresholds — All Models
### Overview
The image contains three line charts comparing the performance of three models (Owen3-VL-235B-Instruct-FP8, Owen3.5-397B-A17B-FP8, Step3-VL-10B) across similarity thresholds (0.3–0.8). Each chart represents a different accuracy metric: Verified Accuracy, Random-Guess Accuracy, and Overall (Non-Verified) Accuracy. Data points are plotted with distinct colors (blue, orange, green) corresponding to the legend.
---
### Components/Axes
- **X-Axis**: Similarity Threshold (0.3, 0.4, 0.5, 0.6, 0.7, 0.8).
- **Y-Axes**:
- Verified Accuracy (%): 0–20%.
- Random-Guess Accuracy (%): 0–20%.
- Overall (Non-Verified) Accuracy (%): 0–20%.
- **Legend**: Located at the top, mapping colors to models:
- Blue: Owen3-VL-235B-Instruct-FP8.
- Orange: Owen3.5-397B-A17B-FP8.
- Green: Step3-VL-10B.
---
### Detailed Analysis
#### Verified Accuracy
- **Owen3-VL-235B-Instruct-FP8 (Blue)**:
- 0.3: 1.7% → 0.4: 4.7% → 0.5: 5.4% → 0.6: 4.7% → 0.7: 3.6% → 0.8: 2.0%.
- **Trend**: Peaks at 0.5, then declines.
- **Owen3.5-397B-A17B-FP8 (Orange)**:
- 0.3: 7.2% → 0.4: 10.3% → 0.5: 9.4% → 0.6: 8.7% → 0.7: 6.8% → 0.8: 3.6%.
- **Trend**: Peaks at 0.4, gradual decline.
- **Step3-VL-10B (Green)**:
- 0.3: 1.7% → 0.4: 2.9% → 0.5: 4.7% → 0.6: 5.0% → 0.7: 6.8% → 0.8: 7.7%.
- **Trend**: Steady increase.
#### Random-Guess Accuracy
- **Owen3-VL-235B-Instruct-FP8 (Blue)**:
- 0.3: 5.6% → 0.4: 7.7% → 0.5: 10.3% → 0.6: 7.6% → 0.7: 6.0% → 0.8: 5.3%.
- **Trend**: Peaks at 0.5, then declines.
- **Owen3.5-397B-A17B-FP8 (Orange)**:
- 0.3: 0.5% → 0.4: 2.7% → 0.5: 2.0% → 0.6: 4.4% → 0.7: 6.0% → 0.8: 6.6%.
- **Trend**: Gradual increase.
- **Step3-VL-10B (Green)**:
- 0.3: 1.9% → 0.4: 2.7% → 0.5: 3.9% → 0.6: 6.0% → 0.7: 8.3% → 0.8: 8.3%.
- **Trend**: Steady increase.
#### Overall (Non-Verified) Accuracy
- **Owen3-VL-235B-Instruct-FP8 (Blue)**:
- 0.3: 8.0% → 0.4: 12.1% → 0.5: 16.7% → 0.6: 13.1% → 0.7: 11.4% → 0.8: 7.3%.
- **Trend**: Peaks at 0.5, sharp decline.
- **Owen3.5-397B-A17B-FP8 (Orange)**:
- 0.3: 8.0% → 0.4: 12.1% → 0.5: 11.4% → 0.6: 13.1% → 0.7: 11.4% → 0.8: 10.2%.
- **Trend**: Peaks at 0.6, moderate fluctuations.
- **Step3-VL-10B (Green)**:
- 0.3: 3.6% → 0.4: 5.6% → 0.5: 8.6% → 0.6: 11.0% → 0.7: 15.1% → 0.8: 15.1%.
- **Trend**: Steady increase.
---
### Key Observations
1. **Peak Performance**:
- All models achieve highest accuracy at similarity thresholds of 0.5–0.6.
- Step3-VL-10B (green) shows consistent growth across all metrics, peaking at 15.1% in Overall Accuracy.
2. **Decline After Threshold 0.6**:
- Owen3-VL-235B and Owen3.5-397B experience significant drops in Verified and Overall Accuracy beyond 0.6.
3. **Random-Guess Performance**:
- Owen3.5-397B’s Random-Guess Accuracy improves steadily, suggesting potential overfitting or model robustness.
4. **Step3-VL-10B Dominance**:
- Outperforms others in Overall Accuracy, particularly at higher thresholds (0.7–0.8).
---
### Interpretation
The data suggests that similarity thresholds of 0.5–0.6 optimize model performance, with Step3-VL-10B demonstrating superior robustness across thresholds. The decline in Owen3-VL-235B and Owen3.5-397B at higher thresholds may indicate overfitting or sensitivity to input variability. The Step3-VL-10B model’s linear growth in Overall Accuracy implies better generalization, making it the most reliable choice for applications requiring consistent performance across varying similarity levels.
