## Line Chart: Model Similarity Over Heuristic Reasoning Steps ### Overview The image depicts a line chart comparing the similarity metric (c_T, t_i) of five AI models across heuristic reasoning steps (t_i). The y-axis represents similarity scores (0.4–0.9), while the x-axis spans 120 reasoning steps. Five distinct data series are plotted, each corresponding to a different model. ### Components/Axes - **X-axis**: "Reasoning step t_i (Heuristic)" (0–120, linear scale) - **Y-axis**: "Similarity(c_T, t_i)" (0.4–0.9, linear scale) - **Legend**: Located in the top-right corner, mapping colors to models: - Blue: DS-R1-Qwen-7B - Orange: Qwen3-8B - Green: Claude-3.7-Sonnet - Purple: GPT-OSS-20B - Brown: Magistral-Small ### Detailed Analysis 1. **DS-R1-Qwen-7B (Blue)**: - Starts at ~0.85 similarity at t_i=0. - Sharp decline to ~0.55 by t_i=60. - Stabilizes with minor fluctuations (~0.52–0.58) afterward. 2. **Qwen3-8B (Orange)**: - Begins at ~0.88 similarity at t_i=0. - Rapid drop to ~0.5 by t_i=40. - Gradual recovery to ~0.55 by t_i=120. 3. **Claude-3.7-Sonnet (Green)**: - Initial similarity ~0.78 at t_i=0. - Steady decline to ~0.55 by t_i=60. - Slight uptick to ~0.58 by t_i=120. 4. **GPT-OSS-20B (Purple)**: - Starts at ~0.65 similarity at t_i=0. - Oscillates between ~0.45–0.55 throughout. - No clear upward/downward trend. 5. **Magistral-Small (Brown)**: - Begins at ~0.75 similarity at t_i=0. - Sharp drop to ~0.4 by t_i=40. - Recovers to ~0.55 by t_i=120 with volatility. ### Key Observations - **Initial Decline**: All models show a steep similarity drop in the first 40 steps, suggesting diminishing performance with increasing heuristic complexity. - **Stabilization**: Models like DS-R1-Qwen-7B and Qwen3-8B stabilize after t_i=60, while others (e.g., GPT-OSS-20B) remain volatile. - **Recovery Patterns**: Qwen3-8B and Magistral-Small exhibit partial recovery post-t_i=60, whereas Claude-3.7-Sonnet shows minimal improvement. - **Lowest Performance**: GPT-OSS-20B consistently underperforms, maintaining the lowest similarity scores. ### Interpretation The data suggests that heuristic reasoning steps initially correlate with high similarity, but performance degrades as complexity increases. Models with higher initial similarity (e.g., Qwen3-8B) may overfit simpler tasks, leading to sharper declines. Partial recovery in later steps (e.g., Qwen3-8B, Magistral-Small) implies adaptive mechanisms or robustness to complexity. GPT-OSS-20B’s persistent low scores highlight potential limitations in handling heuristic reasoning. The trends underscore trade-offs between model architecture and task complexity, with no single model maintaining high similarity across all steps.