# Technical Document Extraction: t-SNE Visualization of Agent Task Datasets This document provides a comprehensive extraction and analysis of three t-SNE (t-distributed Stochastic Neighbor Embedding) scatter plots. These plots visualize the high-dimensional feature space of four distinct agent task datasets. ## 1. General Metadata (Common to all three charts) * **Title:** t-SNE Visualization of Agent Task Datasets * **X-Axis Label:** t-SNE Dimension 1 * **Y-Axis Label:** t-SNE Dimension 2 * **Legend Location:** Top-right corner of each plot. * **Legend Categories & Color Mapping:** * **DesktopEnv Data:** Blue (Square marker) * **Mind2Web Data:** Red (Circle marker) * **WebArena Data:** Green (Cross/X marker) * **Android Data:** Purple (Diamond marker) * **Visual Overlays:** Each plot contains colored ellipses (confidence intervals or cluster boundaries) corresponding to the colors of the four data categories. --- ## 2. Component Isolation and Analysis The image consists of three panels (Left, Center, Right) representing different stages or variations of the t-SNE embedding. ### Panel 1: Left Chart (Initial/High Separation) * **Axis Scales:** * X-axis: -60 to 60 * Y-axis: -40 to 40 * **Spatial Distribution & Trends:** * **DesktopEnv (Blue):** Forms a distinct, dense cluster on the far right of the plot (X: 20 to 60, Y: -20 to 20). It has the least overlap with other datasets. * **Android (Purple):** Located in the upper-middle region (X: -20 to 20, Y: 10 to 35). It shows moderate separation but is adjacent to WebArena. * **Mind2Web (Red):** Located on the left side (X: -50 to -10, Y: -30 to 10). It shows significant internal variance but remains largely separated from DesktopEnv. * **WebArena (Green):** Occupies the lower-middle region (X: -30 to 10, Y: -30 to 10). It shows the most overlap with Mind2Web (Red). * **Trend Verification:** The clusters are relatively isolated, suggesting high feature distinctness between the environments in this specific embedding. ### Panel 2: Center Chart (Moderate Convergence) * **Axis Scales:** * X-axis: -40 to 40 * Y-axis: -40 to 40 * **Spatial Distribution & Trends:** * **DesktopEnv (Blue):** Shifted to the upper-right quadrant (X: 0 to 30, Y: 10 to 45). * **Android (Purple):** Now appears as a smaller, tighter cluster near the center/origin (X: 10 to 20, Y: -10 to 10), showing closer proximity to the other three groups. * **Mind2Web (Red):** Centered in the lower-middle (X: -20 to 10, Y: -40 to 0). * **WebArena (Green):** Located in the middle-left (X: -40 to 0, Y: -10 to 20). * **Trend Verification:** The clusters are moving closer together compared to the first panel. The ellipses show increased overlap, particularly between WebArena (Green) and Mind2Web (Red). ### Panel 3: Right Chart (High Convergence/Overlap) * **Axis Scales:** * X-axis: -50 to 50 * Y-axis: -40 to 60 * **Spatial Distribution & Trends:** * **DesktopEnv (Blue):** Remains in the upper-right (X: 0 to 40, Y: 20 to 60). * **Android (Purple):** Highly concentrated in a small central "bridge" area (X: 0 to 15, Y: 0 to 15), overlapping with the edges of all other clusters. * **Mind2Web (Red):** Occupies the lower-left quadrant (X: -40 to 10, Y: -30 to 10). * **WebArena (Green):** Spread across the bottom and middle-left, showing significant intermingling with Mind2Web. * **Trend Verification:** This panel represents the highest degree of feature similarity or a more compressed embedding. The "Android" data acts as a central hub between the more specialized web and desktop environments. --- ## 3. Data Summary Table (Estimated Centroids) | Dataset | Color | Panel 1 (L) Approx. Center [x, y] | Panel 2 (C) Approx. Center [x, y] | Panel 3 (R) Approx. Center [x, y] | | :--- | :--- | :--- | :--- | :--- | | **DesktopEnv** | Blue | [40, 0] | [15, 30] | [20, 40] | | **Mind2Web** | Red | [-35, -10] | [-5, -20] | [-15, -15] | | **WebArena** | Green | [-15, -15] | [-20, 5] | [-10, -10] | | **Android** | Purple | [0, 25] | [15, 0] | [10, 5] | ## 4. Conclusion The visualizations demonstrate the relational mapping of four agent task datasets. **DesktopEnv** consistently remains the most distinct cluster. **Mind2Web** and **WebArena** (both web-based) show the highest degree of similarity and overlap across all three views. **Android** data serves as a transitional feature set that shares characteristics with both desktop and web environments.

## 1-SNE Visualization of Agent Task Datasets ### Overview Three 2D scatter plots visualize the distribution of agent task datasets using 1-SNE dimensionality reduction. Each plot shows four data categories (Descriptive, Handwritten, Mathematical, Random) with distinct colors and clustering patterns. The visualizations demonstrate how datasets separate or overlap in reduced-dimensional space. ### Components/Axes - **Axes Labels**: - X-axis: "1-SNE Dimension 1" - Y-axis: "1-SNE Dimension 2" - **Legend**: Located in the top-right corner of all plots. - **Descriptive Data**: Blue - **Handwritten Data**: Red - **Mathematical Data**: Green - **Random Data**: Purple - **Ellipses**: Drawn around clusters to highlight density regions. ### Detailed Analysis #### Plot 1 (Left) - **Descriptive Data (Blue)**: Clustered in the top-right quadrant (X: 20–40, Y: 10–30). - **Handwritten Data (Red)**: Concentrated in the bottom-left quadrant (X: -20–0, Y: -20–0). - **Mathematical Data (Green)**: Spread across the bottom-center (X: -10–10, Y: -20–0). - **Random Data (Purple)**: Scattered in the top-left quadrant (X: -20–0, Y: 10–30). - **Ellipses**: - Blue ellipse: Covers ~80% of blue points. - Red ellipse: Encloses ~90% of red points. - Green ellipse: Contains ~70% of green points. - Purple ellipse: Encircles ~60% of purple points. #### Plot 2 (Center) - **Descriptive Data (Blue)**: Dominates the top-right (X: 20–40, Y: 10–30), with tighter clustering. - **Handwritten Data (Red)**: Bottom-left (X: -20–0, Y: -20–0), slightly overlapping with green. - **Mathematical Data (Green)**: Bottom-center (X: -10–10, Y: -20–0), with minor overlap into red. - **Random Data (Purple)**: Top-left (X: -20–0, Y: 10–30), with sparse points in the center. - **Ellipses**: - Blue ellipse: Covers ~90% of blue points. - Red ellipse: Encloses ~80% of red points. - Green ellipse: Contains ~65% of green points. - Purple ellipse: Encircles ~55% of purple points. #### Plot 3 (Right) - **Descriptive Data (Blue)**: Top-right (X: 20–40, Y: 10–30), with some points spilling into the center. - **Handwritten Data (Red)**: Bottom-left (X: -20–0, Y: -20–0), with minor overlap into green. - **Mathematical Data (Green)**: Bottom-center (X: -10–10, Y: -20–0), with increased overlap into red. - **Random Data (Purple)**: Top-left (X: -20–0, Y: 10–30), with significant points in the center. - **Ellipses**: - Blue ellipse: Covers ~75% of blue points. - Red ellipse: Encloses ~70% of red points. - Green ellipse: Contains ~60% of green points. - Purple ellipse: Encircles ~50% of purple points. ### Key Observations 1. **Clustering**: Descriptive data consistently clusters in the top-right across all plots. 2. **Overlap**: - Handwritten and Mathematical data overlap in the bottom-left/center in Plots 2 and 3. - Random data increasingly overlaps with other categories in Plot 3. 3. **Ellipse Coverage**: Descriptive data has the highest ellipse coverage (80–90%), while Random data has the lowest (50–60%). ### Interpretation The 1-SNE visualization effectively separates Descriptive and Handwritten data, which maintain distinct clusters. Mathematical data shows moderate separation but overlaps with Handwritten data in later plots, suggesting potential ambiguity in their reduced-dimensional representation. Random data exhibits the least coherence, with points dispersing across multiple regions, indicating it may lack inherent structure. The ellipses confirm that Descriptive data is the most compact, while Random data is the most dispersed. These trends suggest that 1-SNE preserves key separations but may struggle with datasets lacking clear patterns, such as Random data. The increasing overlap in Plot 3 could imply data mixing or reduced discriminative power in higher-dimensional space.