## Box Plot: Model Accuracy vs. Reasoning Steps
### Overview
The image is a box plot comparing the accuracy (in percentage) of two models, "GPT-3.5" and "Our Model," as the number of reasoning steps increases from 1 to 5. The chart visually demonstrates the distribution of accuracy scores for each model at each step, including medians, quartiles, and outliers.
### Components/Axes
* **Chart Type:** Grouped Box Plot.
* **X-Axis:** Labeled "Number of Reasoning Steps". It has five discrete categories: "1 Step", "2 Steps", "3 Steps", "4 Steps", and "5 Steps".
* **Y-Axis:** Labeled "Accuracy (%)". The scale runs from 40 to 80, with major gridlines at intervals of 10 (40, 50, 60, 70, 80).
* **Legend:** Located in the top-left corner of the chart area.
* A blue square/line corresponds to "GPT-3.5".
* An orange square/line corresponds to "Our Model".
* **Data Series:** Two series of box plots, one blue (GPT-3.5) and one orange (Our Model), plotted side-by-side for each reasoning step category.
### Detailed Analysis
The plot provides specific median accuracy values annotated above each box. The following data is extracted by matching the box color to the legend and reading the associated value.
**Trend Verification:**
* **GPT-3.5 (Blue):** The median accuracy shows a consistent downward trend as the number of reasoning steps increases. The line connecting the medians slopes downward from left to right.
* **Our Model (Orange):** The median accuracy also shows a consistent downward trend as steps increase, but the decline is less steep than GPT-3.5's until the final step.
**Data Points (Median Accuracy %):**
* **1 Step:**
* GPT-3.5 (Blue): 79%
* Our Model (Orange): 74.5%
* **2 Steps:**
* GPT-3.5 (Blue): 70.3%
* Our Model (Orange): 67.1%
* **3 Steps:**
* GPT-3.5 (Blue): 65.3%
* Our Model (Orange): Value not explicitly annotated. Visually, the median line is slightly below the 65% gridline, approximately 64-65%.
* **4 Steps:**
* GPT-3.5 (Blue): Value not explicitly annotated. Visually, the median line is just above the 60% gridline, approximately 61-62%.
* Our Model (Orange): Value not explicitly annotated. Visually, the median line is between the 60% and 65% gridlines, approximately 63%.
* **5 Steps:**
* GPT-3.5 (Blue): 42.1%
* Our Model (Orange): Value not explicitly annotated. Visually, the median line is just above the 60% gridline, approximately 61%.
**Additional Visual Details:**
* **Spread (Interquartile Range - IQR):** The height of the boxes (IQR) generally increases for both models as steps increase, indicating greater variability in performance with more complex reasoning.
* **Outliers:** Individual data points (dots) are visible below the lower whiskers for several categories, indicating instances of significantly lower accuracy. These are present for both models at steps 2, 3, 4, and 5.
### Key Observations
1. **Performance Crossover:** GPT-3.5 starts with a higher median accuracy at 1 Step (79% vs. 74.5%) but is overtaken by "Our Model" by 2 Steps and maintains a lead through 5 Steps.
2. **Significant Drop at 5 Steps for GPT-3.5:** The most dramatic feature is the sharp decline in GPT-3.5's median accuracy at 5 Steps to 42.1%, which is a ~23 percentage point drop from its 3-Step performance.
3. **Consistent Degradation:** Both models exhibit a clear negative correlation between the number of reasoning steps and median accuracy. More steps lead to lower accuracy.
4. **Increased Variability:** The increasing size of the boxes (IQR) suggests that as the task becomes more complex (more steps), the models' performance becomes less consistent.
### Interpretation
This chart illustrates a common challenge in AI reasoning: performance degrades as the required chain of thought lengthens. The data suggests that while both models struggle with multi-step reasoning, "Our Model" demonstrates greater robustness to increased complexity compared to GPT-3.5, particularly beyond the first step.
The catastrophic drop for GPT-3.5 at 5 steps is a critical outlier. It may indicate a specific failure mode, a limitation in its context window or attention mechanism for very long chains, or a point where error propagation becomes unmanageable. In contrast, "Our Model" shows a more graceful, linear degradation.
The increasing variance (wider boxes) with more steps implies that for complex tasks, the outcome becomes less predictable—sometimes the model succeeds, other times it fails significantly (as shown by the outliers). This has practical implications for reliability in applications requiring multi-step logic, such as complex problem-solving, planning, or detailed analysis. The chart argues for the development of models specifically optimized for sustained, multi-step reasoning to maintain both accuracy and consistency.
