This image is a comparative bar chart displaying the "Mean Accuracy (%)" of different pruning and merging methods across various tasks and two language models, under different compression ratios. The chart is divided into two main panels, each representing a different language model. ### Overall Chart Structure The chart consists of two main sub-charts, arranged horizontally. Both sub-charts share a common Y-axis on the left and a common legend on the right. **Y-axis Label**: Mean Accuracy (%) **Y-axis Markers**: 0, 10, 20, 30, 40, 50, 60, 70, 80, 90. ### Legend The legend is located in the top-right corner of the image. It categorizes entries by "Compression Ratio", "Pruning methods", and "Merging methods". **Compression Ratio**: * `0%`: Represented by a dashed grey horizontal line. * `50%`: Represented by a pattern of black diagonal stripes. This pattern is overlaid on the method-specific color. * `25%`: Represented by a solid color. The legend shows "grey solid" as an example, but the actual bar colors are determined by the method. **Pruning methods**: * `REAP (ours)`: Dark blue color. * `EAN`: Light red color. * `Frequency`: Dark green color. **Merging methods**: * `HC-SMoE`: Light yellow color. * `M-SMoE`: Light blue color. For each method, two bars are presented: a solid-colored bar representing the 25% compression ratio, and a striped-colored bar (using the method's color with black diagonal stripes) representing the 50% compression ratio. The dashed grey line indicates the 0% compression ratio (baseline accuracy). ### Chart 1: GLM - 4.5 - Air This sub-chart is located on the left side of the image and is labeled "GLM - 4.5 - Air" below its X-axis. **X-axis Categories**: Coding, Math, Creative Writing, MC (Multiple Choice). #### Data Points and Trends for GLM - 4.5 - Air **1. Coding** * **0% Compression (Baseline)**: Approximately 60%. * **REAP (ours)** (Dark blue): * 25% Compression (solid): ~60% * 50% Compression (striped): ~56% * *Trend*: Maintains accuracy very close to baseline at 25% compression, with a slight drop at 50%. * **EAN** (Light red): * 25% Compression (solid): ~59% * 50% Compression (striped): ~52% * *Trend*: Slightly below baseline at 25%, with a more noticeable drop at 50%. * **Frequency** (Dark green): * 25% Compression (solid): ~58% * 50% Compression (striped): ~33% * *Trend*: Moderate drop at 25%, significant drop at 50%, showing the largest degradation among methods. * **HC-SMoE** (Light yellow): * 25% Compression (solid): ~57% * 50% Compression (striped): ~44% * *Trend*: Moderate drop at 25%, further significant drop at 50%. * **M-SMoE** (Light blue): * 25% Compression (solid): ~56% * 50% Compression (striped): ~30% * *Trend*: Lowest accuracy at both 25% and 50% compression, with a substantial drop at 50%. **2. Math** * **0% Compression (Baseline)**: Approximately 89%. * **REAP (ours)** (Dark blue): * 25% Compression (solid): ~88% * 50% Compression (striped): ~87% * *Trend*: Very robust, maintaining accuracy very close to baseline at both compression levels. * **EAN** (Light red): * 25% Compression (solid): ~88% * 50% Compression (striped): ~81% * *Trend*: Strong performance at 25%, with a moderate drop at 50%. * **Frequency** (Dark green): * 25% Compression (solid): ~87% * 50% Compression (striped): ~62% * *Trend*: Good at 25%, but a very significant drop at 50% compression. * **HC-SMoE** (Light yellow): * 25% Compression (solid): ~87% * 50% Compression (striped): ~70% * *Trend*: Good at 25%, with a notable drop at 50%. * **M-SMoE** (Light blue): * 25% Compression (solid): ~86% * 50% Compression (striped): ~47% * *Trend*: Lowest performance at 50% compression, showing the largest degradation. **3. Creative Writing** * **0% Compression (Baseline)**: Approximately 84%. * **REAP (ours)** (Dark blue): * 25% Compression (solid): ~83% * 50% Compression (striped): ~76% * *Trend*: Strong performance, close to baseline at 25%, with a moderate drop at 50%. * **EAN** (Light red): * 25% Compression (solid): ~82% * 50% Compression (striped): ~79% * *Trend*: Strong performance, with a relatively small drop at 50%. * **Frequency** (Dark green): * 25% Compression (solid): ~80% * 50% Compression (striped): ~60% * *Trend*: Moderate drop at 25%, significant drop at 50%. * **HC-SMoE** (Light yellow): * 25% Compression (solid): ~79% * 50% Compression (striped): ~60% * *Trend*: Moderate drop at 25%, significant drop at 50%. * **M-SMoE** (Light blue): * 25% Compression (solid): ~67% * 50% Compression (striped): ~40% * *Trend*: Lowest accuracy at both 25% and 50% compression, with a substantial drop at 50%. **4. MC (Multiple Choice)** * **0% Compression (Baseline)**: Approximately 75%. * **REAP (ours)** (Dark blue): * 25% Compression (solid): ~67% * 50% Compression (striped): ~54% * *Trend*: Moderate drop at 25%, further drop at 50%. * **EAN** (Light red): * 25% Compression (solid): ~64% * 50% Compression (striped): ~58% * *Trend*: Moderate drop at 25%, with a smaller drop at 50% compared to REAP. * **Frequency** (Dark green): * 25% Compression (solid): ~65% * 50% Compression (striped): ~59% * *Trend*: Moderate drop at 25%, with a smaller drop at 50% compared to REAP. * **HC-SMoE** (Light yellow): * 25% Compression (solid): ~70% * 50% Compression (striped): ~60% * *Trend*: Best performing method in this category at 25% compression, with a moderate drop at 50%. * **M-SMoE** (Light blue): * 25% Compression (solid): ~60% * 50% Compression (striped): ~44% * *Trend*: Lowest accuracy at both 25% and 50% compression, with a substantial drop at 50%. ### Chart 2: Qwen3 - 30B - A3B This sub-chart is located on the right side

# Technical Data Extraction: Model Compression Performance Comparison This document contains a detailed extraction of data from a grouped bar chart comparing different model compression methods across two large language models (LLMs) and four task categories. ## 1. General Metadata * **Y-Axis Title:** Mean Accuracy (%) * **Y-Axis Scale:** 0 to 90 (increments of 10) * **X-Axis Categories (Tasks):** Coding, Math, Creative Writing, MC (Multiple Choice) * **Models Evaluated:** * **GLM – 4.5 – Air** (Left Plot) * **Qwen3 – 30B – A3B** (Right Plot) ## 2. Legend and Key The chart uses a combination of colors and textures to represent different methods and compression ratios. ### Compression Ratio (Texture/Line) * **0%:** Indicated by a horizontal dashed grey line above each task group (Baseline). * **50%:** Indicated by bars with a **diagonal hatching pattern** (forward slash). * **25%:** Indicated by **solid color bars**. ### Methods (Color) * **Pruning methods:** * **REAP (ours):** Blue * **EAN:** Pink/Red * **Frequency:** Green * **Merging methods:** * **HC-SMoE:** Gold/Yellow * **M-SMoE:** Light Blue/Cyan --- ## 3. Data Extraction: GLM – 4.5 – Air (Left Plot) | Task | Baseline (0%) | Method | 25% Ratio (Solid) | 50% Ratio (Hatched) | | :--- | :--- | :--- | :--- | :--- | | **Coding** | ~60% | REAP (ours) | ~60% | ~55% | | | | EAN | ~59% | ~51% | | | | Frequency | ~57% | ~32% | | | | HC-SMoE | ~58% | ~44% | | | | M-SMoE | ~55% | ~30% | | **Math** | ~88% | REAP (ours) | ~88% | ~81% | | | | EAN | ~87% | ~61% | | | | Frequency | ~87% | ~70% | | | | HC-SMoE | ~87% | ~47% | | | | M-SMoE | ~86% | ~39% | | **Creative Writing** | ~84% | REAP (ours) | ~84% | ~75% | | | | EAN | ~82% | ~70% | | | | Frequency | ~79% | ~60% | | | | HC-SMoE | ~78% | ~59% | | | | M-SMoE | ~78% | ~39% | | **MC** | ~75% | REAP (ours) | ~68% | ~56% | | | | EAN | ~64% | ~51% | | | | Frequency | ~65% | ~52% | | | | HC-SMoE | ~70% | ~56% | | | | M-SMoE | ~59% | ~45% | --- ## 4. Data Extraction: Qwen3 – 30B – A3B (Right Plot) | Task | Baseline (0%) | Method | 25% Ratio (Solid) | 50% Ratio (Hatched) | | :--- | :--- | :--- | :--- | :--- | | **Coding** | ~58% | REAP (ours) | ~58% | ~56% | | | | EAN | ~58% | ~55% | | | | Frequency | ~58% | ~47% | | | | HC-SMoE | ~53% | ~38% | | | | M-SMoE | ~55% | ~41% | | **Math** | ~89% | REAP (ours) | ~88% | ~86% | | | | EAN | ~88% | ~87% | | | | Frequency | ~88% | ~73% | | | | HC-SMoE | ~85% | ~1% | | | | M-SMoE | ~88% | ~72% | | **Creative Writing** | ~81% | REAP (ours) | ~81% | ~72% | | | | EAN | ~81% | ~70% | | | | Frequency | ~81% | ~68% | | | | HC-SMoE | ~50% | ~0% | | | | M-SMoE | ~81% | ~72% | | **MC** | ~72% | REAP (ours) | ~68% | ~55% | | | | EAN | ~63% | ~53% | | | | Frequency | ~63% | ~54% | | | | HC-SMoE | ~67% | ~56% | | | | M-SMoE | ~56% | ~45% | --- ## 5. Key Observations and Trends * **Performance Retention:** The "REAP (ours)" method consistently maintains the highest accuracy across almost all tasks and models, especially at the 50% compression ratio (hatched bars), where other methods often show significant degradation. * **Task Sensitivity:** "Math" and "Creative Writing" generally show higher baseline accuracies (~80-90%) compared to "Coding" and "MC" (~60-75%). * **Method Failure:** In the Qwen3 model, the "HC-SMoE" method shows a near-total failure (accuracy drops to near 0%) in the "Math" and "Creative Writing" tasks at the 50% compression ratio. * **Compression Impact:** Performance at 25% compression (solid bars) is generally very close to the 0% baseline for most methods, while 50% compression (hatched bars) reveals the robustness of the "REAP" method compared to "Frequency" or "Merging" methods.

## Bar Chart: Mean Accuracy (%) for Different Writing Tasks ### Overview The bar chart compares the mean accuracy percentages for different writing tasks across two models: GLM-4.5-Air and Qwen3-30B-A3B. The tasks include Coding, Math, Creative, and MC (Multiple Choice). ### Components/Axes - **X-axis**: Different writing tasks (Coding, Math, Creative, MC) - **Y-axis**: Mean Accuracy (%) ranging from 0 to 90 - **Legend**: - Compression Ratio (0%, 25%) - Pruning methods (REAP, Frequency, Merging methods, EAN, M-SMOE) - Line colors represent different pruning methods ### Detailed Analysis or ### Content Details - **GLM-4.5-Air**: - Coding: Mean Accuracy ≈ 50% - Math: Mean Accuracy ≈ 75% - Creative: Mean Accuracy ≈ 65% - MC: Mean Accuracy ≈ 60% - **Qwen3-30B-A3B**: - Coding: Mean Accuracy ≈ 55% - Math: Mean Accuracy ≈ 80% - Creative: Mean Accuracy ≈ 70% - MC: Mean Accuracy ≈ 65% ### Key Observations - **Math Task**: Qwen3-30B-A3B consistently outperforms GLM-4.5-Air in the Math task. - **Creative Task**: GLM-4.5-Air shows slightly higher accuracy in the Creative task compared to Qwen3-30B-A3B. - **MC Task**: Both models perform similarly in the MC task. ### Interpretation The data suggests that Qwen3-30B-A3B is more accurate in generating text for the Math task compared to GLM-4.5-Air. For the Creative task, GLM-4.5-Air performs slightly better. In the MC task, both models have similar performance. The compression ratio and pruning methods do not significantly impact the accuracy in this comparison. The visual trend shows that as the compression ratio increases, the accuracy generally decreases for both models.

## Chart Analysis: Mean Accuracy of Pruning and Merging Methods This image presents two grouped bar charts comparing the mean accuracy of different pruning and merging methods at 50% and 25% compression ratios against a 0% baseline. The comparison is conducted across four tasks: Coding, Math, Creative Writing, and MC. ### Chart 1: GLM – 4.5 – Air **Y-Axis:** Mean Accuracy (%) (Range: 0-90) **X-Axis:** Tasks (Coding, Math, Creative Writing, MC) **Legend:** * **Compression Ratio:** * 0%: Dashed gray line * 50%: Solid bar * 25%: Hatched bar * **Pruning methods:** * REAP (ours): Blue * EAN: Red * Frequency: Green * **Merging methods:** * HC-SMoE: Yellow * M-SMoE: Light Blue **Data Table for GLM – 4.5 – Air:** | Task | Method | 0% Baseline | 50% Compression | 25% Compression | | :--- | :--- | :--- | :--- | :--- | | **Coding** | REAP (ours) | ~60% | ~60% | ~51% | | | EAN | ~60% | ~59% | ~50% | | | Frequency | ~60% | ~57% | ~32% | | | HC-SMoE | ~60% | ~57% | ~44% | | | M-SMoE | ~60% | ~55% | ~30% | | **Math** | REAP (ours) | ~88% | ~88% | ~86% | | | EAN | ~88% | ~87% | ~81% | | | Frequency | ~88% | ~87% | ~61% | | | HC-SMoE | ~88% | ~87% | ~70% | | | M-SMoE | ~88% | ~86% | ~46% | | **Creative Writing** | REAP (ours) | ~84% | ~83% | ~75% | | | EAN | ~84% | ~82% | ~70% | | | Frequency | ~84% | ~79% | ~60% | | | HC-SMoE | ~84% | ~78% | ~59% | | | M-SMoE | ~84% | ~78% | ~39% | | **MC** | REAP (ours) | ~75% | ~68% | ~56% | | | EAN | ~75% | ~63% | ~51% | | | Frequency | ~75% | ~65% | ~52% | | | HC-SMoE | ~75% | ~70% | ~56% | | | M-SMoE | ~75% | ~59% | ~44% | --- ### Chart 2: Qwen3 – 30B – A3B **Y-Axis:** Mean Accuracy (%) (Range: 0-90) **X-Axis:** Tasks (Coding, Math, Creative Writing, MC) **Legend:** (Same as Chart 1) **Data Table for Qwen3 – 30B – A3B:** | Task | Method | 0% Baseline | 50% Compression | 25% Compression | | :--- | :--- | :--- | :--- | :--- | | **Coding** | REAP (ours) | ~58% | ~58% | ~56% | | | EAN | ~58% | ~58% | ~55% | | | Frequency | ~58% | ~58% | ~47% | | | HC-SMoE | ~58% | ~58% | ~38% | | | M-SMoE | ~58% | ~56% | ~41% | | **Math** | REAP (ours) | ~89% | ~89% | ~88% | | | EAN | ~89% | ~88% | ~86% | | | Frequency | ~89% | ~88% | ~87% | | | HC-SMoE | ~89% | ~85% | ~73% | | | M-SMoE | ~89% | ~84% | ~83% | | **Creative Writing** | REAP (ours) | ~81% | ~81% | ~72% | | | EAN | ~81% | ~81% | ~70% | | | Frequency | ~81% | ~81% | ~68% | | | HC-SMoE | ~81% | ~81% | ~50% | | | M-SMoE | ~81% | ~81% | ~73% | | **MC** | REAP (ours) | ~72% | ~68% | ~55% | | | EAN | ~72% | ~63% | ~53% | | | Frequency | ~72% | ~63% | ~52% | | | HC-SMoE | ~72% | ~67% | ~54% | | | M-SMoE | ~72% | ~56% | ~45% |

# Technical Document Extraction: Bar Chart Analysis ## Chart Overview The image contains a comparative bar chart analyzing **mean accuracy (%)** across four tasks (**Coding, Math, Creative Writing, MC**) for two language models: **GLM-4.5-Air** and **Qwen3-30B-A3B**. The chart includes **pruning methods** (REAP, EAN, Frequency) and **merging methods** (HC-SMoE, M-SMoE), with **compression ratios** (0%, 50%, 25%) indicated by dashed lines. --- ## Key Components ### Legend - **Location**: Right side of the chart. - **Compression Ratios**: - 0%: Solid gray line. - 50%: Dark blue line. - 25%: Light gray line. - **Pruning Methods**: - REAP (ours): Blue. - EAN: Red. - Frequency: Green. - **Merging Methods**: - HC-SMoE: Yellow. - M-SMoE: Light blue. ### Axes - **X-axis**: Tasks (Coding, Math, Creative Writing, MC). - **Y-axis**: Mean Accuracy (%) ranging from 0% to 90%. --- ## Data Extraction ### GLM-4.5-Air | Task | REAP (ours) | EAN | Frequency | HC-SMoE | M-SMoE | |--------------------|-------------|-------|-----------|---------|--------| | **Coding** | 58% | 52% | 32% | 45% | 30% | | **Math** | 88% | 82% | 61% | 70% | 58% | | **Creative Writing** | 82% | 79% | 60% | 78% | 40% | | **MC** | 72% | 65% | 53% | 68% | 45% | ### Qwen3-30B-A3B | Task | REAP (ours) | EAN | Frequency | HC-SMoE | M-SMoE | |--------------------|-------------|-------|-----------|---------|--------| | **Coding** | 56% | 54% | 48% | 52% | 42% | | **Math** | 89% | 87% | 85% | 83% | 81% | | **Creative Writing** | 81% | 77% | 75% | 73% | 70% | | **MC** | 70% | 63% | 58% | 65% | 50% | --- ## Trends and Observations 1. **REAP (ours)** consistently achieves the highest accuracy across all tasks and models. 2. **Math** task shows the highest performance for both models, with Qwen3-30B-A3B reaching **89%** (REAP). 3. **Coding** task has the lowest accuracy for both models (GLM: 58%, Qwen3: 56%). 4. **Compression Ratios**: - 0% (solid gray) represents baseline performance. - 50% (dark blue) and 25% (light gray) indicate reduced accuracy compared to baseline, but exact values are not explicitly labeled in the chart. 5. **Merging Methods** (HC-SMoE, M-SMoE) generally underperform compared to pruning methods. --- ## Spatial Grounding - **Legend Position**: Right-aligned, outside the main chart area. - **Bar Colors**: Match legend labels exactly (e.g., REAP = blue, EAN = red). --- ## Component Isolation ### Header - Title: Not explicitly labeled, but inferred from context as a comparison of pruning/merging methods. ### Main Chart - **GLM-4.5-Air** (Left): - Math task dominates with **88%** (REAP). - Coding task is the weakest (**58%**). - **Qwen3-30B-A3B** (Right): - Math task peaks at **89%** (REAP). - Creative Writing shows strong performance (**81%**). ### Footer - No explicit footer text; compression ratios are integrated into the legend. --- ## Final Notes - All data points align with legend color assignments. - Dashed lines (compression ratios) suggest performance degradation but lack explicit numerical labels. - REAP (ours) outperforms all other methods in both models.