## Line Graph: Similarity Decay Across Reasoning Steps (GPT5) ### Overview The image depicts a line graph comparing the similarity decay of five AI models over reasoning steps (t_i) measured in GPT5 units. The y-axis represents similarity scores (C_T, t_i) ranging from 0.50 to 0.85, while the x-axis spans 0 to 175 reasoning steps. Five distinct data series are plotted, each corresponding to a different model. ### Components/Axes - **X-axis**: "Reasoning step t_i (GPT5)" (0–175, linear scale) - **Y-axis**: "Similarity (C_T, t_i)" (0.50–0.85, linear scale) - **Legend**: - Blue: DS-R1-Qwen-7B - Orange: Qwen3-8B - Green: Claude-3.7-Sonnet - Purple: GPT-OSS-20B - Brown: Magistral-Small - **Data Series**: Five colored lines with jagged trends, indicating stepwise measurements. ### Detailed Analysis 1. **DS-R1-Qwen-7B (Blue)**: - Starts at ~0.85 similarity at t_i=0. - Declines sharply to ~0.60 by t_i=50. - Stabilizes with minor fluctuations between t_i=75–175. 2. **Qwen3-8B (Orange)**: - Begins at ~0.80 similarity at t_i=0. - Gradual decline to ~0.55 by t_i=100. - Slight recovery to ~0.58 by t_i=150. 3. **Claude-3.7-Sonnet (Green)**: - Initial similarity ~0.75 at t_i=0. - Sharp drop to ~0.60 by t_i=50. - Fluctuates between ~0.55–0.65 until t_i=175. 4. **GPT-OSS-20B (Purple)**: - Lowest starting point (~0.60 at t_i=0). - Steep decline to ~0.50 by t_i=50. - Erratic fluctuations between ~0.45–0.55 until t_i=175. 5. **Magistral-Small (Brown)**: - Mid-range start (~0.70 at t_i=0). - Gradual decline to ~0.58 by t_i=100. - Stabilizes with minor oscillations until t_i=175. ### Key Observations - **Initial Decline**: All models show rapid similarity decay in the first 50 steps. - **Stability Variance**: DS-R1-Qwen-7B and Magistral-Small stabilize faster than others. - **Lowest Performance**: GPT-OSS-20B consistently exhibits the lowest similarity. - **Notable Dip**: Claude-3.7-Sonnet shows a pronounced drop at t_i=50, followed by volatility. ### Interpretation The graph illustrates how similarity to a target metric (C_T) degrades as reasoning steps increase. Models with higher initial similarity (DS-R1-Qwen-7B, Qwen3-8B) degrade more rapidly, suggesting potential overfitting or inefficiency in maintaining coherence over extended reasoning. The persistent low performance of GPT-OSS-20B may indicate architectural limitations or training data gaps. The green line’s volatility (Claude-3.7-Sonnet) could reflect sensitivity to specific reasoning tasks. Notably, no model maintains high similarity beyond ~100 steps, highlighting a universal challenge in long-context reasoning for current AI systems.