# Technical Document Extraction: Bar Chart Analysis
## Chart Overview
The image is a horizontal bar chart comparing the performance of three "Self-Rewarding" models (M₁, M₂, M₃) against an "SFT Baseline" across three outcome categories: **Self-Rewarding Wins**, **Tie**, and **SFT Baseline Wins**. The chart uses color-coded segments to represent proportions of outcomes.
---
## Key Components
### Legend
- **Green**: Self-Rewarding Wins
- **Blue**: Tie
- **Red**: SFT Baseline Wins
### Axes
- **X-Axis**: Outcome categories (Self-Rewarding Wins, Tie, SFT Baseline Wins)
- **Y-Axis**: Models compared (Self-Rewarding M₃ vs. SFT Baseline, Self-Rewarding M₂ vs. SFT Baseline, Self-Rewarding M₁ vs. SFT Baseline)
---
## Data Points
| Model | Self-Rewarding Wins | Tie | SFT Baseline Wins |
|--------------------------------|---------------------|-------|-------------------|
| Self-Rewarding M₃ vs. SFT | 62.5% | 27.7% | 9.8% |
| Self-Rewarding M₂ vs. SFT | 49.2% | 36.3% | 14.5% |
| Self-Rewarding M₁ vs. SFT | 30.5% | 38.7% | 30.9% |
---
## Observations
1. **Self-Rewarding M₃** dominates with **62.5% Self-Rewarding Wins**, significantly outperforming the SFT Baseline (9.8% wins).
2. **Self-Rewarding M₂** shows moderate performance: **49.2% Self-Rewarding Wins** vs. **14.5% SFT Baseline Wins**.
3. **Self-Rewarding M₁** has the lowest Self-Rewarding Wins (**30.5%**) but nearly equal ties (38.7%) and SFT Baseline Wins (30.9%).
4. All models show **Tie rates** exceeding SFT Baseline Wins, except M₃.
---
## Structural Notes
- Bars are grouped by model, with segments proportional to outcome percentages.
- Percentages sum to 100% for each model, indicating mutually exclusive outcomes.
- Color coding aligns strictly with the legend (green = Self-Rewarding Wins, blue = Tie, red = SFT Baseline Wins).
---
## Conclusion
The chart highlights a clear performance gradient: M₃ > M₂ > M₁ in Self-Rewarding Wins, with diminishing returns as model complexity decreases. Ties and SFT Baseline Wins increase for lower-performing models (M₁).
