## Line Chart: Premature Output Rate across Similarity Thresholds — All Models ### Overview The chart visualizes the relationship between similarity thresholds (x-axis) and premature output rates (y-axis) for three AI models: Qwen3-VL-235B-Instruct-FP8 (blue), Qwen3.5-397B-A17B-FP8 (orange), and Step3-VL-10B (green). Data points are plotted at thresholds of 0.3, 0.4, 0.5, 0.6, and 0.8, with percentages ranging from 0% to 25%. ### Components/Axes - **X-axis (Similarity Threshold)**: Discrete values at 0.3, 0.4, 0.5, 0.6, and 0.8. - **Y-axis (Premature Output Rate)**: Continuous scale from 0% to 25%, marked in 5% increments. - **Legend**: Located in the top-right corner, associating: - Blue circles with Qwen3-VL-235B-Instruct-FP8 - Orange squares with Qwen3.5-397B-A17B-FP8 - Green diamonds with Step3-VL-10B ### Detailed Analysis #### Qwen3-VL-235B-Instruct-FP8 (Blue) - **Trend**: Steeply declining output rate as similarity threshold increases. - **Values**: - 0.3: 18.4% - 0.4: 9.1% - 0.5: 4.7% - 0.6: 2.5% - 0.8: 0.0% #### Qwen3.5-397B-A17B-FP8 (Orange) - **Trend**: Gradual decline with a minor uptick at 0.4. - **Values**: - 0.3: 14.0% - 0.4: 10.6% - 0.5: 7.7% - 0.6: 6.3% - 0.8: 0.3% #### Step3-VL-10B (Green) - **Trend**: Initial rise to a peak at 0.4, followed by a steady decline. - **Values**: - 0.3: 17.6% - 0.4: 20.2% - 0.5: 12.7% - 0.6: 10.0% - 0.8: 6.5% ### Key Observations 1. **Threshold Impact**: Higher similarity thresholds consistently reduce premature output rates across all models. 2. **Model Performance**: - **Step3-VL-10B** exhibits the highest initial output rate (17.6% at 0.3) and a peak (20.2% at 0.4), suggesting sensitivity to moderate thresholds. - **Qwen3-VL-235B-Instruct-FP8** achieves the lowest output rate (0.0% at 0.8), indicating superior threshold tolerance. - **Qwen3.5-397B-A17B-FP8** shows a gradual decline but retains a residual 0.3% at 0.8. 3. **Anomaly**: Step3-VL-10B’s output rate increases from 0.3 to 0.4, contradicting the general trend of decreasing rates with higher thresholds. ### Interpretation - **Threshold Optimization**: Higher similarity thresholds (e.g., 0.6–0.8) significantly reduce premature outputs, with Qwen3-VL-235B-Instruct-FP8 performing best at 0.8 (0.0%). - **Model-Specific Behavior**: Step3-VL-10B’s peak at 0.4 suggests a potential trade-off between threshold stringency and output quality for this model. - **Robustness**: Qwen3-VL-235B-Instruct-FP8 demonstrates the most consistent performance, eliminating premature outputs entirely at the highest threshold. The orange line’s minor uptick at 0.4 may reflect model-specific quirks or data noise. This analysis highlights the importance of threshold selection in balancing output quality and efficiency across different AI architectures.