# Technical Document Extraction: Normalized Performance Comparison of Expert Routing Strategies ## 1. Image Overview This image is a grouped bar chart comparing the "Normalized Performance" of four different Mixture-of-Experts (MoE) architectural configurations across six distinct NLP evaluation metrics. The chart uses a color-coded legend to differentiate between routing strategies and expert segmentation levels. ## 2. Component Isolation ### A. Header / Legend **Location:** Top-left quadrant of the chart area. **Content:** Four categories with corresponding color swatches. 1. **Blue (#0072B2):** `0 shared expert + 2 out of 16 routed experts (GShard)` 2. **Yellow/Orange (#E69F00):** `1 shared expert + 1 out of 15 routed experts (+ shared expert isolation)` 3. **Green (#009E73):** `1 shared expert + 3 out of 31 routed experts (+ fine-grained expert segmentation)` 4. **Dark Orange (#D55E00):** `1 shared expert + 7 out of 63 routed experts (+ finer expert segmentation)` ### B. Main Chart Area (Axes) * **Y-Axis (Vertical):** * **Label:** `Normalized Performance` * **Range:** 0.5 to 1.2 * **Major Tick Marks:** 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2 * **X-Axis (Horizontal):** * **Label:** `Metrics` * **Categories (Left to Right):** `HellaSwag`, `PIQA`, `ARC-easy`, `ARC-challenge`, `TriviaQA`, `NaturalQuestions` ## 3. Data Extraction (Reconstructed Table) The following table estimates the numerical values based on the visual alignment with the Y-axis grid lines. | Metric | GShard (Blue) | Shared Isolation (Yellow) | Fine-grained (Green) | Finer (Dark Orange) | | :--- | :---: | :---: | :---: | :---: | | **HellaSwag** | ~0.92 | ~0.96 | ~0.98 | 1.00 | | **PIQA** | ~0.98 | ~0.98 | ~0.97 | 1.00 | | **ARC-easy** | ~0.89 | ~0.97 | ~1.00 | 1.00 | | **ARC-challenge** | ~0.92 | ~0.91 | ~0.92 | 1.00 | | **TriviaQA** | ~0.61 | ~0.85 | ~0.93 | 1.00 | | **NaturalQuestions** | ~0.56 | ~0.79 | ~0.88 | 1.00 | ## 4. Trend Verification and Analysis ### General Trend Across all six metrics, the **"1 shared expert + 7 out of 63 routed experts (+ finer expert segmentation)"** (Dark Orange) consistently achieves the highest normalized performance, reaching the 1.0 mark in every category. This indicates it serves as the baseline or the peak performance target for this comparison. ### Specific Series Observations * **GShard (Blue):** Shows the lowest performance across most metrics, particularly on knowledge-heavy tasks like `TriviaQA` and `NaturalQuestions`, where it drops significantly below 0.7. * **Shared Expert Isolation (Yellow):** Provides a substantial performance boost over GShard in almost every category, most notably in `TriviaQA` and `NaturalQuestions`. * **Fine-grained Expert Segmentation (Green):** Generally shows incremental improvements over the "Shared Isolation" strategy. It matches or nearly matches the top performance in `ARC-easy`. * **Task Sensitivity:** The performance gap between the simplest configuration (GShard) and the most complex (Finer segmentation) is narrowest in `PIQA` and widest in `NaturalQuestions` and `TriviaQA`. This suggests that finer expert segmentation is critical for factual retrieval/knowledge tasks. ## 5. Language Declaration The text in this image is entirely in **English**. No other languages are present.

# Technical Document Extraction: Expert Configuration Performance Analysis ## Chart Description This bar chart compares the normalized performance of four expert configurations across six question-answering metrics. The configurations vary in the number of shared experts and routed experts, with increasing granularity in expert segmentation. ### Axis Labels - **X-axis (Metrics):** - HellaSwag - PIQA - ARC-easy - ARC-challenge - TriviaQA - NaturalQuestions - **Y-axis (Normalized Performance):** - Scale: 0.5 to 1.2 - Increment: 0.1 ### Legend Four expert configurations with color-coded bars: 1. **Blue:** 0 shared expert + 2 out of 16 routed experts (GShard) 2. **Orange:** 1 shared expert + 15 routed experts (+ shared expert isolation) 3. **Green:** 1 shared expert + 31 routed experts (+ fine-grained expert segmentation) 4. **Red:** 1 shared expert + 63 routed experts (+ finer expert segmentation) ## Key Data Points & Trends ### Performance by Metric | Metric | GShard (Blue) | Shared +15 (Orange) | Shared +31 (Green) | Shared +63 (Red) | |---------------|---------------|---------------------|--------------------|------------------| | HellaSwag | ~0.92 | ~0.95 | ~0.98 | ~1.00 | | PIQA | ~0.98 | ~0.98 | ~0.97 | ~1.00 | | ARC-easy | ~0.89 | ~0.96 | ~1.00 | ~1.00 | | ARC-challenge | ~0.92 | ~0.91 | ~0.92 | ~1.00 | | TriviaQA | ~0.61 | ~0.85 | ~0.94 | ~1.00 | | NaturalQuestions | ~0.56 | ~0.79 | ~0.88 | ~1.00 | ### Observations - **Performance Scaling:** All configurations achieve near-perfect performance (1.00) on most metrics when using 1 shared expert with 63 routed experts (red bars). - **GShard Limitations:** The baseline GShard configuration (0 shared experts) shows significantly lower performance across all metrics, particularly on TriviaQA and NaturalQuestions. - **Expert Segmentation Impact:** Increasing the number of routed experts (from 15 to 63) improves performance in configurations with 1 shared expert, with the most notable gains in TriviaQA and NaturalQuestions. - **Consistency:** The 1 shared expert + 63 routed experts configuration (red) consistently outperforms other configurations across all metrics. ## Technical Notes - Normalized performance values suggest relative comparison rather than absolute scores. - The chart implies a trade-off between expert sharing and segmentation granularity, with finer segmentation (63 experts) yielding optimal results. - No configuration exceeds the 1.00 performance threshold except in specific cases (e.g., PIQA with 1 shared +15 experts).