## Chart Type: Multiple Line Charts Comparing Sentence Probabilities ### Overview The image presents four line charts arranged horizontally, comparing the probabilities of different sentence types (Teacher, Boosted, Shared, and Student) across a range of sentence indices. Each chart represents a different model or configuration ("Ours" and "DeepSeek-Distill-Qwen3-8B"). The charts show how the probability of each sentence type changes as the sentence index increases. ### Components/Axes * **X-axis (all charts):** "Sentence Index" ranging from 0 to 1500 in increments of 300. * **Y-axis (all charts):** "Probability" ranging from 0.1 to 0.6 (first chart), 0.0 to 1.0 (second chart), 0.4 to 0.9 (third chart), and 0.0 to 0.5 (fourth chart). * **Legends (top of the image):** * Green: "Teacher Sentence" * Purple: "Boosted Sentence" * Orange: "Shared Sentence" * Blue: "Student Sentence" * **Chart Titles (above each chart):** * Chart 1: "Ours" * Chart 2: "Ours" * Chart 3: "DeepSeek-Distill-Qwen3-8B" * Chart 4: "DeepSeek-Distill-Qwen3-8B" ### Detailed Analysis **Chart 1: Ours** * **Teacher Sentence (Green):** The probability starts around 0.4 and generally decreases to approximately 0.2 over the sentence index range. There are fluctuations, but the overall trend is downward. * Value at Sentence Index 1200: ~0.25 * Value at Sentence Index 1500: ~0.20 * Text Annotation: "= 45.08" located near the right side of the chart. **Chart 2: Ours** * **Boosted Sentence (Purple):** The probability starts around 0.55-0.6 and remains relatively stable across the sentence index range. * Value at Sentence Index 0: ~0.55 * Value at Sentence Index 1500: ~0.55 * Text Annotation: "= -62.99" located near the right side of the chart. * **Shared Sentence (Orange):** The probability starts near 0 and increases to approximately 0.15. * Value at Sentence Index 0: ~0.0 * Value at Sentence Index 1500: ~0.15 * Text Annotation: "= 28.98" located near the right side of the chart. * **Student Sentence (Blue):** The probability starts near 0 and increases to approximately 0.1. * Value at Sentence Index 0: ~0.0 * Value at Sentence Index 1500: ~0.1 * Text Annotation: "= -11.07" located near the right side of the chart. **Chart 3: DeepSeek-Distill-Qwen3-8B** * **Teacher Sentence (Green):** The probability starts high, around 0.8, and decreases rapidly before fluctuating between 0.5 and 0.6. * Value at Sentence Index 0: ~0.8 * Value at Sentence Index 1500: ~0.55 * Text Annotation: "= 61.26" located near the right side of the chart. **Chart 4: DeepSeek-Distill-Qwen3-8B** * **Boosted Sentence (Purple):** The probability starts low, around 0.25, and increases to approximately 0.35. * Value at Sentence Index 0: ~0.25 * Value at Sentence Index 1500: ~0.35 * Text Annotation: "= -75.71" located near the right side of the chart. * **Shared Sentence (Orange):** The probability starts near 0 and increases to approximately 0.1. * Value at Sentence Index 0: ~0.0 * Value at Sentence Index 1500: ~0.1 * Text Annotation: "= 9.61" located near the right side of the chart. * **Student Sentence (Blue):** The probability starts near 0 and increases to approximately 0.05. * Value at Sentence Index 0: ~0.0 * Value at Sentence Index 1500: ~0.05 * Text Annotation: "= 4.48" located near the right side of the chart. ### Key Observations * The "Ours" model shows a decreasing probability for "Teacher Sentence" as the sentence index increases, while the "Boosted Sentence" probability remains relatively stable. * The "DeepSeek-Distill-Qwen3-8B" model shows a decreasing probability for "Teacher Sentence" and an increasing probability for "Boosted Sentence" as the sentence index increases. * The "Shared Sentence" and "Student Sentence" probabilities are generally low for both models, with a slight increase as the sentence index increases. * The text annotations on each chart appear to represent some form of score or metric associated with the model and sentence type. ### Interpretation The charts compare the performance of two models ("Ours" and "DeepSeek-Distill-Qwen3-8B") in terms of sentence type probabilities across a range of sentence indices. The data suggests that the models handle different sentence types differently. For example, the "Ours" model seems to favor "Teacher Sentences" initially, but this preference decreases over the sentence index. In contrast, the "DeepSeek-Distill-Qwen3-8B" model shows an increasing preference for "Boosted Sentences." The low probabilities for "Shared Sentence" and "Student Sentence" across both models might indicate that these sentence types are less common or more difficult to predict. The text annotations likely represent a quantitative measure of the model's performance or bias towards each sentence type, providing a more precise comparison of the models' behavior.

\n ## Chart: Probability vs. Sentence Index for Different Training Methods ### Overview The image presents four line charts, each depicting the probability distribution across sentence indices for different training methods. The x-axis represents the sentence index, ranging from 0 to 1500. The y-axis represents the probability, ranging from approximately 0.0 to 0.8. Each chart displays four lines representing "Teacher Sentence", "Boosted Sentence", "Shared Sentence", and "Student Sentence". Each chart also includes a numerical value indicating a calculated metric (likely correlation or similarity) associated with the data. ### Components/Axes * **X-axis:** Sentence Index (0 to 1500) * **Y-axis:** Probability (0.0 to 0.8) * **Legend:** * Teacher Sentence (Green, solid line) * Boosted Sentence (Purple, dashed line) * Shared Sentence (Orange, dotted line) * Student Sentence (Blue, solid line) * **Chart Titles:** * Chart 1: "Ours" * Chart 2: "Ours" * Chart 3: "DeepSeek-Distill-Qwen-3-8B" * Chart 4: "DeepSeek-Distill-Qwen-8B" * **Numerical Values:** Each chart displays a single numerical value near the end of the x-axis, likely representing a correlation or similarity score. ### Detailed Analysis or Content Details **Chart 1: "Ours"** * **Teacher Sentence (Green):** The line starts at approximately 0.55 and slopes downward, reaching approximately 0.20 at sentence index 1500. The value at sentence index 1200 is approximately 0.45. * **Boosted Sentence (Purple):** The line starts at approximately 0.75 and remains relatively stable, fluctuating around 0.70 throughout the sentence indices. The value at sentence index 1200 is approximately 0.63. * **Shared Sentence (Orange):** The line starts at approximately 0.15 and slopes upward, reaching approximately 0.29 at sentence index 1500. The value at sentence index 1200 is approximately 0.28. * **Student Sentence (Blue):** The line starts at approximately 0.10 and slopes upward, reaching approximately 0.12 at sentence index 1500. The value at sentence index 1200 is approximately 0.11. * **Numerical Value:** +45.08 **Chart 2: "Ours"** * **Teacher Sentence (Green):** The line starts at approximately 0.75 and slopes downward, reaching approximately 0.55 at sentence index 1500. The value at sentence index 1200 is approximately 0.63. * **Boosted Sentence (Purple):** The line remains relatively flat, starting at approximately 0.78 and fluctuating around 0.75. The value at sentence index 1200 is approximately 0.73. * **Shared Sentence (Orange):** The line starts at approximately 0.05 and slopes upward, reaching approximately 0.12 at sentence index 1500. The value at sentence index 1200 is approximately 0.11. * **Student Sentence (Blue):** The line starts at approximately 0.02 and slopes upward, reaching approximately 0.05 at sentence index 1500. The value at sentence index 1200 is approximately 0.04. * **Numerical Value:** +62.99 **Chart 3: "DeepSeek-Distill-Qwen-3-8B"** * **Teacher Sentence (Green):** The line starts at approximately 0.70 and slopes downward, reaching approximately 0.50 at sentence index 1500. The value at sentence index 1200 is approximately 0.61. * **Boosted Sentence (Purple):** The line remains relatively flat, starting at approximately 0.75 and fluctuating around 0.70. The value at sentence index 1200 is approximately 0.70. * **Shared Sentence (Orange):** The line starts at approximately 0.05 and slopes upward, reaching approximately 0.10 at sentence index 1500. The value at sentence index 1200 is approximately 0.09. * **Student Sentence (Blue):** The line starts at approximately 0.04 and slopes upward, reaching approximately 0.06 at sentence index 1500. The value at sentence index 1200 is approximately 0.05. * **Numerical Value:** +61.26 **Chart 4: "DeepSeek-Distill-Qwen-8B"** * **Teacher Sentence (Green):** The line starts at approximately 0.65 and slopes downward, reaching approximately 0.45 at sentence index 1500. The value at sentence index 1200 is approximately 0.55. * **Boosted Sentence (Purple):** The line remains relatively flat, starting at approximately 0.70 and fluctuating around 0.65. The value at sentence index 1200 is approximately 0.65. * **Shared Sentence (Orange):** The line starts at approximately 0.04 and slopes upward, reaching approximately 0.08 at sentence index 1500. The value at sentence index 1200 is approximately 0.07. * **Student Sentence (Blue):** The line starts at approximately 0.04 and slopes upward, reaching approximately 0.07 at sentence index 1500. The value at sentence index 1200 is approximately 0.05. * **Numerical Value:** -75.71 ### Key Observations * The "Teacher Sentence" consistently shows a decreasing probability trend across all charts. * The "Boosted Sentence" generally maintains a higher and more stable probability compared to the other sentences. * The "Shared Sentence" and "Student Sentence" exhibit similar upward trends, but with significantly lower probability values. * The numerical values vary significantly across the charts, suggesting different performance levels for each training method. * Chart 4 ("DeepSeek-Distill-Qwen-8B") has a negative numerical value, indicating a potentially inverse relationship or poor correlation. ### Interpretation The charts likely represent the probability distribution of sentence embeddings or representations learned during different training paradigms. The "Teacher Sentence" represents the original, high-quality data, while the "Student Sentence" represents the learned representation. The "Boosted Sentence" and "Shared Sentence" likely represent intermediate stages or components of the training process. The decreasing trend of the "Teacher Sentence" could indicate a loss of information or a shift in the representation space as the sentence index increases. The stable "Boosted Sentence" suggests that the boosting mechanism effectively preserves information. The upward trends of the "Shared Sentence" and "Student Sentence" indicate that the model is learning to represent the data, but the low probability values suggest that the learned representations are still significantly different from the original data. The varying numerical values likely represent a correlation or similarity score between the "Teacher Sentence" and the other sentences. The negative value in Chart 4 suggests that the "DeepSeek-Distill-Qwen-8B" model is learning representations that are dissimilar to the original data, potentially indicating a problem with the training process or model architecture. The charts demonstrate the impact of different training methods on the quality and similarity of learned sentence representations.

\n ## Line Charts: Comparative Probability Analysis of Sentence Types Across Models ### Overview The image displays four line charts arranged horizontally, comparing the probability trends of four different sentence types ("Teacher Sentence," "Boosted Sentence," "Shared Sentence," "Student Sentence") across two different models or conditions labeled "Ours" and "DeepSeek-Distill-Qwen3-8B." Each chart plots "Probability" on the y-axis against "Sentence Index" on the x-axis. The charts include shaded confidence intervals or variance bands and specific numerical annotations at the end of each data series. ### Components/Axes * **Legend (Top Center):** A shared legend for all four charts defines the color coding: * **Green Line:** Teacher Sentence * **Purple Line:** Boosted Sentence * **Orange Line:** Shared Sentence * **Blue Line:** Student Sentence * **Chart 1 (Far Left):** * **Title/Label:** "Ours" (top-right corner) * **Y-axis:** Label "Probability," scale from 0.1 to 0.6. * **X-axis:** Label "Sentence Index," scale from 0 to 1500. * **Chart 2 (Center Left):** * **Title/Label:** "Ours" (top-right corner) * **Y-axis:** Unlabeled, scale from 0.0 to 1.0. * **X-axis:** Label "Sentence Index," scale from 0 to 1500. * **Chart 3 (Center Right):** * **Title/Label:** "DeepSeek-Distill-Qwen3-8B" (top-right corner) * **Y-axis:** Unlabeled, scale from 0.4 to 0.9. * **X-axis:** Label "Sentence Index," scale from 0 to 1500. * **Chart 4 (Far Right):** * **Title/Label:** "DeepSeek-Distill-Qwen3-8B" (top-right corner) * **Y-axis:** Unlabeled, scale from 0.0 to 0.5. * **X-axis:** Label "Sentence Index," scale from 0 to 1500. ### Detailed Analysis **Chart 1 ("Ours"):** * **Teacher Sentence (Green):** Shows a steep initial decline from a probability of ~0.6 at index 0, followed by a gradual, noisy decrease. The line ends with an annotation **"= 45.08"**. A shaded green band indicates variance around the line. * **Boosted Sentence (Purple):** Remains relatively flat and low, hovering just above 0.1 probability throughout. * **Shared Sentence (Orange) & Student Sentence (Blue):** Both remain very low, near the bottom of the chart (probability ~0.05 or lower), with minimal visible fluctuation. **Chart 2 ("Ours"):** * **Boosted Sentence (Purple):** Dominates this chart. It rises quickly from near 0 to a plateau around 0.6 probability and remains stable. Annotated with **"= 62.99"**. * **Shared Sentence (Orange):** Shows a gradual, slight increase from near 0 to approximately 0.15-0.2 probability. Annotated with **"= 28.98"**. * **Student Sentence (Blue):** Remains very low, near zero, with a slight upward trend at the very end. Annotated with **"= -11.07"**. * **Teacher Sentence (Green):** Not visibly plotted in this chart. **Chart 3 ("DeepSeek-Distill-Qwen3-8B"):** * **Teacher Sentence (Green):** Fluctuates significantly between approximately 0.5 and 0.7 probability, showing a slight overall downward trend. Annotated with **"= 61.26"**. * **Boosted Sentence (Purple):** Remains relatively flat in the lower portion of the chart, around 0.4 probability. * **Shared Sentence (Orange) & Student Sentence (Blue):** Both remain very low, near the bottom of the chart's visible range (probability ~0.4 or lower). **Chart 4 ("DeepSeek-Distill-Qwen3-8B"):** * **Boosted Sentence (Purple):** Is the highest line, fluctuating between approximately 0.3 and 0.45 probability. Annotated with **"= -75.71"**. * **Shared Sentence (Orange):** Shows a noisy but generally increasing trend from near 0 to about 0.1 probability. Annotated with **"= 9.61"**. * **Student Sentence (Blue):** Remains very low, near zero, with a slight increase towards the end. Annotated with **"= 4.48"**. * **Teacher Sentence (Green):** Not visibly plotted in this chart. ### Key Observations 1. **Model/Condition Comparison:** The "Ours" model (Charts 1 & 2) shows a clear separation where the Teacher Sentence probability is high in one view and the Boosted Sentence probability is high in another. The "DeepSeek" model (Charts 3 & 4) shows a more mixed picture with the Teacher Sentence having moderate, volatile probability and the Boosted Sentence being the highest in its respective view. 2. **Annotation Anomaly:** Several numerical annotations are negative (e.g., "-11.07", "-75.71"), which is atypical for a metric labeled "Probability." This strongly suggests the annotations may represent a different calculated value (e.g., a score, difference, or log probability) rather than the raw probability shown on the y-axis. 3. **Line Visibility:** The Teacher Sentence line is only prominent in the first and third charts. The Boosted Sentence line is prominent in the second and fourth charts. This may indicate the charts are grouped to highlight different primary comparisons. 4. **Trend Consistency:** The Shared and Student Sentence types consistently show the lowest probabilities across all charts and models. ### Interpretation The data appears to analyze the internal probability dynamics of different sentence types within two language model distillation or training frameworks ("Ours" vs. "DeepSeek-Distill-Qwen3-8B"). The "Sentence Index" likely represents a sequence of tokens or steps in a generation or training process. The contrasting profiles between the "Ours" charts suggest that in this model, the "Teacher" and "Boosted" sentence types dominate under different conditions or metrics. The "DeepSeek" model shows less dominance by a single type, with the Teacher sentence exhibiting high variance, possibly indicating less stability. The most critical finding is the disconnect between the y-axis ("Probability") and the final annotations. The negative values for what should be a non-negative probability metric indicate these annotations are **not** the final y-axis values. They likely represent a derived performance score, a change relative to a baseline, or a different metric altogether (e.g., a reward signal). Without additional context, their exact meaning is uncertain, but their presence is a key feature of the data presentation. The charts collectively suggest that the "Boosted Sentence" type is engineered to achieve high probability in specific contexts, while the "Teacher Sentence" provides a baseline that can be volatile.

## Line Graphs: Sentence Probability Comparison Across Models ### Overview The image contains four line graphs comparing sentence probabilities across two models: "Ours" (top row) and "DeepSeek-Distill-Qwen3-8B" (bottom row). Each graph tracks four sentence types (Teacher, Boosted, Shared, Student) over 1,500 sentences. Probabilities are plotted on the y-axis (0.1–0.9), while the x-axis represents sentence indices (0–1,500). Legends in the top-right corner of each graph map colors to sentence types. ### Components/Axes - **X-axis**: "Sentence Index" (0–1,500) - **Y-axis**: "Probability" (0.1–0.9 for top graphs, 0.0–0.5 for bottom graphs) - **Legends**: - Green: Teacher Sentence - Purple: Boosted Sentence - Orange: Shared Sentence - Blue: Student Sentence - **Graph Layout**: 2x2 grid (top-left: "Ours" Teacher/Boosted/Shared/Student; top-right: "Ours" Teacher/Boosted/Shared/Student; bottom-left: "DeepSeek-Distill-Qwen3-8B" Teacher/Boosted/Shared/Student; bottom-right: "DeepSeek-Distill-Qwen3-8B" Teacher/Boosted/Shared/Student). ### Detailed Analysis #### Top-Left Graph ("Ours") - **Teacher Sentence (Green)**: Starts at ~0.6, decreases steadily to ~0.45 (value: **45.08**). - **Boosted Sentence (Purple)**: Starts at ~0.3, rises slightly, then drops sharply to **-62.99**. - **Shared Sentence (Orange)**: Remains near 0.1–0.2, ending at **28.98**. - **Student Sentence (Blue)**: Fluctuates between 0.05–0.15, ending at **-11.07**. #### Top-Right Graph ("Ours") - **Teacher Sentence (Green)**: Starts at ~0.6, decreases to ~0.45 (value: **45.08**). - **Boosted Sentence (Purple)**: Starts at ~0.3, rises slightly, then drops to **-62.99**. - **Shared Sentence (Orange)**: Stays near 0.1–0.2, ending at **28.98**. - **Student Sentence (Blue)**: Fluctuates between 0.05–0.15, ending at **-11.07**. #### Bottom-Left Graph ("DeepSeek-Distill-Qwen3-8B") - **Teacher Sentence (Green)**: Starts at ~0.8, decreases to ~0.6 (value: **61.26**). - **Boosted Sentence (Purple)**: Starts at ~0.4, rises slightly, then drops sharply to **-75.71**. - **Shared Sentence (Orange)**: Remains near 0.1–0.2, ending at **9.61**. - **Student Sentence (Blue)**: Fluctuates between 0.05–0.15, ending at **4.48**. #### Bottom-Right Graph ("DeepSeek-Distill-Qwen3-8B") - **Teacher Sentence (Green)**: Starts at ~0.8, decreases to ~0.6 (value: **61.26**). - **Boosted Sentence (Purple)**: Starts at ~0.4, rises slightly, then drops to **-75.71**. - **Shared Sentence (Orange)**: Stays near 0.1–0.2, ending at **9.61**. - **Student Sentence (Blue)**: Fluctuates between 0.05–0.15, ending at **4.48**. ### Key Observations 1. **Teacher Sentences**: - Both models show a gradual decline in probability over sentences. - "DeepSeek-Distill-Qwen3-8B" starts with higher probabilities (~0.8 vs. ~0.6) but ends lower (~61.26 vs. 45.08). 2. **Boosted Sentences**: - Both models exhibit a sharp negative drop in later sentences (e.g., **-62.99** in "Ours", **-75.71** in "DeepSeek-Distill-Qwen3-8B"). - This suggests a potential instability or anomaly in the model's handling of Boosted Sentences. 3. **Shared vs. Student Sentences**: - Shared Sentences consistently outperform Student Sentences in probability. - Student Sentences show the lowest probabilities, with negative values in "Ours" (-11.07) and positive but low values in "DeepSeek-Distill-Qwen3-8B" (4.48). ### Interpretation - **Model Performance**: - "Ours" maintains higher probabilities for Teacher and Shared Sentences compared to "DeepSeek-Distill-Qwen3-8B", which shows a steeper decline in Teacher Sentence probabilities. - The negative values for Boosted and Student Sentences in "Ours" may indicate model limitations or data preprocessing issues. - **Sentence Type Dynamics**: - Teacher Sentences dominate early but degrade over time, suggesting diminishing model confidence. - Boosted Sentences’ sharp negative drop could reflect overfitting or adversarial examples. - Shared Sentences’ stability implies robustness, while Student Sentences’ low probabilities highlight challenges in generating coherent outputs. - **Anomalies**: - Negative probabilities (e.g., **-62.99**, **-75.71**) are statistically invalid but may represent model errors or miscalibrations. - The divergence in Student Sentence performance between models suggests differing training objectives or architectures.