## Bar Chart: Mistral-7B-Instruct Performance Comparison
### Overview
The chart compares three model configurations (None, StruQ, SecAlign) across three evaluation metrics: AlpacaEval2 WinRate, Max ASR (Opt.-Free), and Max ASR (Opt.-Based). Performance is measured in percentages, with WinRate (↑) and ASR (↓) indicated by directional arrows.
### Components/Axes
- **X-Axis**: Three evaluation categories:
1. AlpacaEval2 WinRate (↑)
2. Max ASR (↓) - Opt.-Free
3. Max ASR (↓) - Opt.-Based
- **Y-Axis**: Performance metric (WinRate/ASR) in percentage (0–100%).
- **Legend**:
- Gray: None
- Blue: StruQ
- Orange: SecAlign
- **Bar Groups**: Three clusters of bars (one per model) for each evaluation category.
### Detailed Analysis
1. **AlpacaEval2 WinRate (↑)**:
- **None**: ~67%
- **StruQ**: ~70%
- **SecAlign**: ~68%
- *Trend*: StruQ leads slightly, followed by SecAlign and None.
2. **Max ASR (↓) - Opt.-Free**:
- **None**: ~60%
- **StruQ**: ~2%
- **SecAlign**: ~0%
- *Trend*: None dominates; StruQ and SecAlign show near-zero ASR.
3. **Max ASR (↓) - Opt.-Based**:
- **None**: ~90%
- **StruQ**: ~27%
- **SecAlign**: ~1%
- *Trend*: None again leads; StruQ has moderate ASR, SecAlign minimal.
### Key Observations
- **AlpacaEval2**: StruQ achieves the highest WinRate, suggesting improved instruction-following capability.
- **Opt.-Free ASR**: StruQ and SecAlign drastically reduce ASR compared to None, indicating better robustness in non-optimized scenarios.
- **Opt.-Based ASR**: None maintains the highest ASR, while StruQ and SecAlign show significant drops, implying optimization trade-offs.
### Interpretation
The data suggests:
1. **StruQ** balances WinRate and ASR improvements, performing best in AlpacaEval2 and moderately in Opt.-Based ASR.
2. **SecAlign** shows minimal gains in WinRate but excels in reducing ASR in Opt.-Free scenarios.
3. **None** (baseline) performs worst in WinRate but best in Opt.-Based ASR, highlighting a potential specialization in adversarial robustness at the cost of general performance.
The chart reveals a trade-off between instruction-following ability (WinRate) and robustness (ASR), with StruQ offering the most balanced improvement across metrics. The drastic ASR reduction in Opt.-Free for StruQ/SecAlign suggests they may prioritize safety in non-optimized settings, while None retains higher adversarial vulnerability in optimized scenarios.
