## Line Graph: Similarity Trends Across AI Models Over Reasoning Steps ### Overview The graph displays the similarity metric (C_T, t_i) of five AI models across 30 reasoning steps (t_i, GPT5). Each line represents a model's performance decay/growth pattern, with values ranging from 30 to 60 on the y-axis. ### Components/Axes - **X-axis**: Reasoning step t_i (GPT5) [0-30] - **Y-axis**: Similarity (C_T, t_i) [30-60] - **Legend**: - Blue circles: DS-R1-Qwen-7B - Orange diamonds: Qwen3-8B - Green squares: Claude-3.7-Sonnet - Purple stars: GPT-OSS-20B - Brown triangles: Magistral-Small - **Grid**: Minor gridlines at every 5-unit interval ### Detailed Analysis 1. **DS-R1-Qwen-7B (Blue)** - Starts at ~60 similarity at t_i=0 - Sharp decline to ~35 by t_i=10 - Stabilizes between 34-36 from t_i=15-30 2. **Qwen3-8B (Orange)** - Begins at ~60 similarity - Gradual decline to ~30 by t_i=15 - Fluctuates between 30-34 after t_i=15 3. **Claude-3.7-Sonnet (Green)** - Initial drop from 40 to 30 (t_i=0-5) - Rises to 38 by t_i=20 - Maintains 35-38 range until t_i=30 4. **GPT-OSS-20B (Purple)** - Starts at 40 similarity - Plummets to 25 by t_i=15 - Recovers to 35 by t_i=30 5. **Magistral-Small (Brown)** - Begins at 45 similarity - Drops to 30 by t_i=10 - Oscillates between 30-35 throughout ### Key Observations - **Initial Decline**: All models show steep similarity drops in early steps (t_i=0-10) - **Recovery Patterns**: Claude-3.7-Sonnet and GPT-OSS-20B demonstrate partial recovery after mid-step troughs - **Stability**: DS-R1-Qwen-7B and Qwen3-8B stabilize at lower similarity values post-t_i=15 - **Volatility**: Magistral-Small exhibits the most fluctuation (±5 similarity units) ### Interpretation The graph reveals that AI model performance (measured by similarity) generally degrades with increased reasoning steps, though some models show resilience or partial recovery. The sharpest initial declines (DS-R1-Qwen-7B and Qwen3-8B) suggest these models may be overfitting to early reasoning patterns. The recovery observed in Claude-3.7-Sonnet and GPT-OSS-20B implies adaptive mechanisms that maintain partial performance despite step increases. Magistral-Small's volatility indicates inconsistent reasoning capabilities across steps. These trends highlight the challenge of maintaining consistent performance in complex reasoning tasks across different AI architectures.