## Bar Chart: Overall Performance Comparison of Policy Models ### Overview The image is a bar chart comparing the overall performance of different policy models. The y-axis represents "Overall Performance," ranging from 25 to 55. The x-axis represents "Policy Model," with several models listed. There are three data series, each represented by a different color bar: "Pass@1" (green), "BoN with InternVL2.5-8B" (red), and "BoN with VisualPRM-8B (ours)" (blue). ### Components/Axes * **Y-axis:** "Overall Performance," with a scale from 25 to 55 in increments of 5. * **X-axis:** "Policy Model," listing the following models: MiniCPM-V2.6, QwenVL2.5-7B, InternVL2.5-8B, InternVL2.5-26B, InternVL2.5-38B, InternVL2.5-78B. * **Legend:** Located at the top-left of the chart, it identifies the three data series: * Green: "Pass@1" * Red: "BoN with InternVL2.5-8B" * Blue: "BoN with VisualPRM-8B (ours)" ### Detailed Analysis Here's a breakdown of the performance for each policy model across the three data series: * **MiniCPM-V2.6:** * Pass@1 (Green): 29.5 * BoN with InternVL2.5-8B (Red): 28.6 * BoN with VisualPRM-8B (Blue): 37.5 * **QwenVL2.5-7B:** * Pass@1 (Green): 41.4 * BoN with InternVL2.5-8B (Red): 41.6 * BoN with VisualPRM-8B (Blue): 45.1 * **InternVL2.5-8B:** * Pass@1 (Green): 32.8 * BoN with InternVL2.5-8B (Red): 33.2 * BoN with VisualPRM-8B (Blue): 41.2 * **InternVL2.5-26B:** * Pass@1 (Green): 36.9 * BoN with InternVL2.5-8B (Red): 39.1 * BoN with VisualPRM-8B (Blue): 45.8 * **InternVL2.5-38B:** * Pass@1 (Green): 44.4 * BoN with InternVL2.5-8B (Red): 44.9 * BoN with VisualPRM-8B (Blue): 50.7 * **InternVL2.5-78B:** * Pass@1 (Green): 46.0 * BoN with InternVL2.5-8B (Red): 46.4 * BoN with VisualPRM-8B (Blue): 51.9 ### Key Observations * The "BoN with VisualPRM-8B (ours)" (blue) consistently outperforms the other two models across all policy models. * The performance of all three data series generally increases as the "InternVL2.5" model number increases (e.g., from 26B to 78B). * The "Pass@1" (green) and "BoN with InternVL2.5-8B" (red) models have very similar performance for each policy model, with the red bar being slightly higher in most cases. ### Interpretation The chart demonstrates that "BoN with VisualPRM-8B (ours)" achieves the highest overall performance compared to "Pass@1" and "BoN with InternVL2.5-8B" across the tested policy models. The trend suggests that increasing the size or complexity of the "InternVL2.5" model generally leads to improved performance for all three configurations. The consistent outperformance of the "VisualPRM-8B" model indicates its effectiveness in the given task or benchmark. The similar performance between "Pass@1" and "InternVL2.5-8B" might suggest a correlation or similar underlying mechanism in their approach.

## Bar Chart: Pass@1 Performance Comparison of Policy Models ### Overview This bar chart compares the "Pass@1" performance of different policy models, using three different approaches: "Pass@1" (green bars), "BoN with InternVL2.5-8B" (salmon/reddish bars), and "BoN with VisualPRM-8B (ours)" (blue bars). The chart displays overall performance on the y-axis against various policy models on the x-axis. ### Components/Axes * **X-axis:** Policy Model. Categories include: MiniCPM-V2.6, QwenVL2.5-7B, InternVL2.5-8B, InternVL2.5-26B, InternVL2.5-38B, InternVL2.5-78B. * **Y-axis:** Overall Performance (labeled as "Overall Performance"). Scale ranges from 25 to 55. * **Legend:** Located in the top-left corner. * Green: Pass@1 * Salmon/Reddish: BoN with InternVL2.5-8B * Blue: BoN with VisualPRM-8B (ours) ### Detailed Analysis The chart consists of six groups of three bars, each representing a different policy model. The height of each bar indicates the "Pass@1" performance. * **MiniCPM-V2.6:** * Pass@1: Approximately 29.5 * BoN with InternVL2.5-8B: Approximately 28.6 * BoN with VisualPRM-8B: Approximately 37.5 * **QwenVL2.5-7B:** * Pass@1: Approximately 41.4 * BoN with InternVL2.5-8B: Approximately 41.6 * BoN with VisualPRM-8B: Approximately 45.1 * **InternVL2.5-8B:** * Pass@1: Approximately 41.2 * BoN with InternVL2.5-8B: Approximately 32.8 * BoN with VisualPRM-8B: Approximately 33.2 * **InternVL2.5-26B:** * Pass@1: Approximately 45.8 * BoN with InternVL2.5-8B: Approximately 36.9 * BoN with VisualPRM-8B: Approximately 39.1 * **InternVL2.5-38B:** * Pass@1: Approximately 44.4 * BoN with InternVL2.5-8B: Approximately 44.9 * BoN with VisualPRM-8B: Approximately 50.7 * **InternVL2.5-78B:** * Pass@1: Approximately 46.0 * BoN with InternVL2.5-8B: Approximately 46.4 * BoN with VisualPRM-8B: Approximately 51.9 **Trends:** * **Pass@1:** Generally increases with larger policy models, peaking at 46.0 for InternVL2.5-78B. * **BoN with InternVL2.5-8B:** Shows a more erratic pattern, with a dip at InternVL2.5-8B and a peak at InternVL2.5-78B. * **BoN with VisualPRM-8B:** Demonstrates a consistent upward trend, achieving the highest performance values across most models, especially at InternVL2.5-38B and InternVL2.5-78B. ### Key Observations * "BoN with VisualPRM-8B (ours)" consistently outperforms the other two approaches, particularly with larger policy models. * The performance of "BoN with InternVL2.5-8B" is relatively stable, but generally lower than "Pass@1" and "BoN with VisualPRM-8B". * The largest performance gains are observed when using "BoN with VisualPRM-8B" in conjunction with the InternVL2.5-78B policy model. * InternVL2.5-8B shows a significant drop in performance for the "BoN with InternVL2.5-8B" approach. ### Interpretation The data suggests that the "BoN with VisualPRM-8B" approach significantly enhances the performance of policy models, especially as the model size increases. This indicates that incorporating visual information through the VisualPRM-8B model is beneficial for achieving higher "Pass@1" scores. The consistent upward trend of "BoN with VisualPRM-8B" suggests a strong positive correlation between model capacity and performance when combined with this approach. The outlier at InternVL2.5-8B for "BoN with InternVL2.5-8B" could indicate a compatibility issue or a specific characteristic of that model that hinders the effectiveness of the BoN approach. The chart demonstrates the effectiveness of the proposed "BoN with VisualPRM-8B" method, positioning it as a superior approach compared to the baseline "Pass@1" and "BoN with InternVL2.5-8B" methods.

\n ## Grouped Bar Chart: Overall Performance of Policy Models with Different Evaluation Methods ### Overview This image is a grouped bar chart comparing the "Overall Performance" of six different vision-language policy models. Each model is evaluated using three distinct methods, represented by differently colored bars. The chart demonstrates the performance impact of using different reward models within a "Best-of-N" (BoN) sampling strategy compared to a baseline. ### Components/Axes * **Y-Axis:** Labeled "Overall Performance". The scale ranges from 25 to 55, with major gridlines at intervals of 5 (25, 30, 35, 40, 45, 50, 55). * **X-Axis:** Labeled "Policy Model". It lists six specific models: 1. MiniCPM-V2.6 2. QwenVL2.5-7B 3. InternVL2.5-8B 4. InternVL2.5-26B 5. InternVL2.5-38B 6. InternVL2.5-78B * **Legend:** Positioned in the top-left corner of the chart area. It defines three data series: * **Green Square:** `Pass@1` * **Red Square:** `BoN with InternVL2.5-8B` * **Blue Square:** `BoN with VisualPRM-8B (ours)` * **Data Labels:** Each bar has its exact numerical value displayed directly above it. ### Detailed Analysis The chart presents performance data for each policy model across the three evaluation methods. The values are as follows: **1. MiniCPM-V2.6** * Pass@1 (Green): 29.5 * BoN with InternVL2.5-8B (Red): 28.6 * BoN with VisualPRM-8B (Blue): 37.5 **2. QwenVL2.5-7B** * Pass@1 (Green): 41.4 * BoN with InternVL2.5-8B (Red): 41.6 * BoN with VisualPRM-8B (Blue): 45.1 **3. InternVL2.5-8B** * Pass@1 (Green): 32.8 * BoN with InternVL2.5-8B (Red): 33.2 * BoN with VisualPRM-8B (Blue): 41.2 **4. InternVL2.5-26B** * Pass@1 (Green): 36.9 * BoN with InternVL2.5-8B (Red): 39.1 * BoN with VisualPRM-8B (Blue): 45.8 **5. InternVL2.5-38B** * Pass@1 (Green): 44.4 * BoN with InternVL2.5-8B (Red): 44.9 * BoN with VisualPRM-8B (Blue): 50.7 **6. InternVL2.5-78B** * Pass@1 (Green): 46.0 * BoN with InternVL2.5-8B (Red): 46.4 * BoN with VisualPRM-8B (Blue): 51.9 ### Key Observations 1. **Consistent Superiority of VisualPRM-8B:** For every single policy model, the blue bar (`BoN with VisualPRM-8B (ours)`) is the tallest, indicating it yields the highest overall performance. 2. **Performance Hierarchy:** The general trend across all models is: `BoN with VisualPRM-8B` > `BoN with InternVL2.5-8B` ≈ `Pass@1`. The red and green bars are often very close in height. 3. **Scale of Improvement:** The performance gain from using VisualPRM-8B (blue) over the other methods is substantial, ranging from approximately +3.5 points (QwenVL2.5-7B) to +8.4 points (InternVL2.5-8B). 4. **Model Scaling Trend:** Within the InternVL2.5 model family, performance generally increases with model size (from 8B to 78B parameters) for all three evaluation methods. 5. **Outlier - MiniCPM-V2.6:** This model shows a unique pattern where `Pass@1` (29.5) slightly outperforms `BoN with InternVL2.5-8B` (28.6), which is the opposite of the trend seen in all other models. ### Interpretation This chart is likely from a research paper introducing "VisualPRM-8B" as a new process reward model. The data serves as empirical evidence for its effectiveness. * **What the data suggests:** The primary finding is that using VisualPRM-8B as a reward model within a Best-of-N sampling framework (`BoN with VisualPRM-8B`) consistently and significantly boosts the overall performance of various vision-language policy models compared to both a standard Pass@1 evaluation and using a different reward model (InternVL2.5-8B). * **Relationship between elements:** The chart directly compares the proposed method (blue) against two baselines (green and red). The consistent gap between the blue bars and the others across diverse policy models (different architectures and sizes) strengthens the claim that the improvement is attributable to the reward model itself, not a specific policy model. * **Notable implications:** The results imply that the quality of the reward model is a critical factor for improving vision-language model performance via sampling-based methods. The fact that VisualPRM-8B outperforms the larger InternVL2.5-8B when used as a reward model suggests it may have been specifically optimized or trained for this reward modeling task, making it a more effective "judge" for selecting good responses. The scaling trend within InternVL2.5 also hints that larger policy models benefit similarly from better reward models.

## Bar Chart: Overall Performance Comparison Across Policy Models ### Overview The chart compares the overall performance of three policy models (Pass@1, BoN with InternVL2.5-8B, and BoN with VisualPRM-8B) across six different policy architectures (MiniCPM-V2.6, GwenVL2.5-7B, InternVL2.5-8B, InternVL2.5-26B, InternVL2.5-38B, InternVL2.5-78B). Performance is measured on a scale from 25 to 55. ### Components/Axes - **X-axis (Policy Model)**: - MiniCPM-V2.6 - GwenVL2.5-7B - InternVL2.5-8B - InternVL2.5-26B - InternVL2.5-38B - InternVL2.5-78B - **Y-axis (Overall Performance)**: - Scale from 25 to 55 in increments of 5. - **Legend**: - Green: Pass@1 - Red: BoN with InternVL2.5-8B - Blue: BoN with VisualPRM-8B (ours) ### Detailed Analysis 1. **MiniCPM-V2.6**: - Pass@1: 29.5 (green) - BoN with InternVL2.5-8B: 28.6 (red) - BoN with VisualPRM-8B: 37.5 (blue) 2. **GwenVL2.5-7B**: - Pass@1: 41.4 (green) - BoN with InternVL2.5-8B: 41.6 (red) - BoN with VisualPRM-8B: 45.1 (blue) 3. **InternVL2.5-8B**: - Pass@1: 32.8 (green) - BoN with InternVL2.5-8B: 33.2 (red) - BoN with VisualPRM-8B: 41.2 (blue) 4. **InternVL2.5-26B**: - Pass@1: 36.9 (green) - BoN with InternVL2.5-8B: 39.1 (red) - BoN with VisualPRM-8B: 45.8 (blue) 5. **InternVL2.5-38B**: - Pass@1: 44.4 (green) - BoN with InternVL2.5-8B: 44.9 (red) - BoN with VisualPRM-8B: 50.7 (blue) 6. **InternVL2.5-78B**: - Pass@1: 46.0 (green) - BoN with InternVL2.5-8B: 46.4 (red) - BoN with VisualPRM-8B: 51.9 (blue) ### Key Observations - **Trend Verification**: - **Pass@1 (green)**: Gradually increases from 29.5 (MiniCPM-V2.6) to 46.0 (InternVL2.5-78B), showing consistent improvement as policy models scale. - **BoN with InternVL2.5-8B (red)**: Starts at 28.6 (MiniCPM-V2.6) and rises to 46.4 (InternVL2.5-78B), mirroring the Pass@1 trend but with slightly lower values. - **BoN with VisualPRM-8B (blue)**: Outperforms both metrics across all models, starting at 37.5 (MiniCPM-V2.6) and peaking at 51.9 (InternVL2.5-78B). The gap widens significantly for larger models (e.g., InternVL2.5-78B: 51.9 vs. 46.4 for BoN with InternVL2.5-8B). - **Notable Outliers**: - MiniCPM-V2.6 has the lowest performance for all metrics, with BoN with VisualPRM-8B showing the largest relative improvement (37.5 vs. 29.5 for Pass@1). - InternVL2.5-78B achieves the highest performance for BoN with VisualPRM-8B (51.9), surpassing all other models. ### Interpretation The data demonstrates that **BoN with VisualPRM-8B** consistently outperforms the other two policy models across all architectures, with the largest performance gains observed in larger models (InternVL2.5-38B and InternVL2.5-78B). This suggests that the VisualPRM-8B framework enhances the BoN (Bootstrapped Neural) approach more effectively than the InternVL2.5-8B variant, particularly as model complexity increases. The Pass@1 metric shows steady improvement but remains the weakest performer, indicating potential limitations in its design or training data. The results highlight the importance of architectural choices in policy models for achieving high overall performance.