## Chart: Validation Score vs. Training Step for GRPO and MEL ### Overview The image is a line chart comparing the validation scores of two models, GRPO and MEL, over training steps. The chart shows how the validation score changes as the models are trained. ### Components/Axes * **Title:** Benchmark: OlympiadBench * **X-axis:** Training Step, ranging from 0 to 140 in increments of 20. * **Y-axis:** Validation Score, ranging from 0.450 to 0.625. * **Legend:** Located in the bottom-right corner. * GRPO (Blue) * MEL (Pink) ### Detailed Analysis * **GRPO (Blue):** * Starts at approximately 0.45. * Increases to approximately 0.48 at step 20. * Decreases to approximately 0.47 at step 30. * Increases to approximately 0.51 at step 40. * Increases to approximately 0.515 at step 50. * Increases to approximately 0.57 at step 60. * Decreases to approximately 0.56 at step 70. * Decreases to approximately 0.555 at step 80. * Increases to approximately 0.56 at step 90. * Increases to approximately 0.59 at step 100. * Decreases to approximately 0.58 at step 110. * Decreases to approximately 0.575 at step 120. * Increases to approximately 0.60 at step 130. * Decreases to approximately 0.58 at step 140. * **MEL (Pink):** * Starts at approximately 0.45. * Increases to approximately 0.50 at step 20. * Increases to approximately 0.54 at step 30. * Increases to approximately 0.58 at step 40. * Increases to approximately 0.58 at step 50. * Increases to approximately 0.58 at step 60. * Increases to approximately 0.58 at step 70. * Increases to approximately 0.59 at step 80. * Increases to approximately 0.595 at step 90. * Increases to approximately 0.60 at step 100. * Increases to approximately 0.60 at step 110. * Increases to approximately 0.60 at step 120. * Increases to approximately 0.60 at step 130. * Increases to approximately 0.62 at step 140. ### Key Observations * Both models start with a similar validation score. * MEL generally outperforms GRPO after the initial training steps. * MEL shows a more consistent upward trend, while GRPO fluctuates more. * MEL reaches a higher validation score at the end of the training period. ### Interpretation The chart suggests that the MEL model performs better than the GRPO model on the OlympiadBench benchmark, as indicated by its higher validation scores over the training period. The consistent upward trend of MEL implies a more stable learning process compared to GRPO, which experiences more fluctuations. The data indicates that MEL is a more effective model for this particular benchmark.

\n ## Line Chart: Validation Score vs. Training Step on OlympiadBench ### Overview This image presents a line chart illustrating the validation score of two models, GRPO and MEL, as a function of the training step on the OlympiadBench benchmark. The chart displays the performance improvement of both models during training. ### Components/Axes * **Title:** Benchmark: OlympiadBench (positioned at the top-center) * **X-axis:** Training Step (ranging from approximately 0 to 140, with tick marks at intervals of 20) * **Y-axis:** Validation Score (ranging from approximately 0.45 to 0.625, with tick marks at intervals of 0.025) * **Legend:** Located in the top-right corner. * GRPO (represented by a light blue line with circular markers) * MEL (represented by a light red line with triangular markers) ### Detailed Analysis **GRPO (Light Blue Line):** The GRPO line generally slopes upward, indicating increasing validation score with training steps. * At Training Step 0: Validation Score ≈ 0.46 * At Training Step 20: Validation Score ≈ 0.48 * At Training Step 40: Validation Score ≈ 0.52 * At Training Step 60: Validation Score ≈ 0.56 * At Training Step 80: Validation Score ≈ 0.56 * At Training Step 100: Validation Score ≈ 0.56 * At Training Step 120: Validation Score ≈ 0.58 * At Training Step 140: Validation Score ≈ 0.60 **MEL (Light Red Line):** The MEL line also slopes upward, but with a steeper initial increase and some fluctuations. * At Training Step 0: Validation Score ≈ 0.45 * At Training Step 20: Validation Score ≈ 0.47 * At Training Step 40: Validation Score ≈ 0.55 * At Training Step 60: Validation Score ≈ 0.59 * At Training Step 80: Validation Score ≈ 0.59 * At Training Step 100: Validation Score ≈ 0.57 * At Training Step 120: Validation Score ≈ 0.59 * At Training Step 140: Validation Score ≈ 0.61 ### Key Observations * MEL consistently outperforms GRPO in terms of validation score across all training steps. * Both models show diminishing returns in performance improvement as training progresses beyond 80 steps. * GRPO exhibits a more stable and gradual increase in validation score, while MEL shows more volatility. * The difference in validation score between the two models is most pronounced in the initial training stages (0-40 steps). ### Interpretation The chart demonstrates the learning curves of two models (GRPO and MEL) on the OlympiadBench benchmark. MEL appears to be a more effective model, achieving higher validation scores throughout the training process. The initial rapid increase in MEL's performance suggests faster learning or a better initial parameter configuration. The plateauing of both curves towards the end of the training process indicates that further training may not yield significant improvements. The fluctuations in MEL's curve could be due to factors such as the stochastic nature of the training process or the complexity of the OlympiadBench dataset. The consistent upward trend for both models suggests that the training process is generally effective in improving performance on the benchmark. The data suggests that MEL is a better choice for this benchmark, but further investigation might be needed to understand the reasons for its superior performance and the fluctuations in its learning curve.

\n ## Line Chart: Benchmark: OlympiadBench ### Overview The image displays a line chart comparing the validation score performance of two models, labeled "GAPO" and "MEL," over the course of training steps on a benchmark called "OlympiadBench." The chart shows both models improving over time, with the MEL model consistently achieving a higher validation score after the initial training steps. ### Components/Axes * **Chart Title:** "Benchmark: OlympiadBench" (centered at the top). * **Y-Axis:** Labeled "Validation Score." The scale runs from 0.450 to 0.625, with major grid lines and labels at intervals of 0.025 (0.450, 0.475, 0.500, 0.525, 0.550, 0.575, 0.600, 0.625). * **X-Axis:** Labeled "Training_Step." The scale runs from 0 to 140, with major grid lines and labels at intervals of 20 (0, 20, 40, 60, 80, 100, 120, 140). * **Legend:** Located in the bottom-right corner of the chart area. It contains two entries: * A blue line with a circle marker labeled "GAPO". * A red line with a circle marker labeled "MEL". * **Data Series:** Two lines plotted on the chart, corresponding to the legend entries. ### Detailed Analysis **Data Series: GAPO (Blue Line)** * **Trend:** The line shows an overall upward trend with notable volatility. It experiences a dip early in training before recovering and climbing, with several smaller fluctuations along the way. * **Approximate Data Points (Training Step, Validation Score):** * (0, ~0.450) * (10, ~0.475) * (20, ~0.465) - *Local minimum* * (30, ~0.475) * (40, ~0.500) * (50, ~0.510) * (60, ~0.540) * (70, ~0.560) * (80, ~0.555) * (90, ~0.560) * (100, ~0.575) * (110, ~0.565) * (120, ~0.580) * (130, ~0.570) * (140, ~0.575) **Data Series: MEL (Red Line)** * **Trend:** The line shows a strong, consistent upward trend with less volatility than the GAPO line. After an initial dip, it climbs steadily and maintains a clear performance lead over GAPO for the majority of the training process. * **Approximate Data Points (Training Step, Validation Score):** * (0, ~0.475) * (10, ~0.500) * (20, ~0.475) - *Local minimum, similar to GAPO* * (30, ~0.525) * (40, ~0.550) * (50, ~0.565) * (60, ~0.575) * (70, ~0.585) * (80, ~0.590) * (90, ~0.595) * (100, ~0.590) * (110, ~0.600) * (120, ~0.600) * (130, ~0.605) * (140, ~0.625) - *Highest point on the chart* ### Key Observations 1. **Performance Gap:** After training step 20, the MEL model (red) establishes and maintains a clear performance advantage over the GAPO model (blue). The gap is most pronounced between steps 40 and 100. 2. **Initial Dip:** Both models experience a performance dip around training step 20, suggesting a common challenge or phase in the early training process on this benchmark. 3. **Volatility vs. Stability:** The GAPO line is more volatile, with sharper peaks and valleys. The MEL line, while not perfectly smooth, demonstrates a more stable and consistent improvement trajectory. 4. **Final Convergence?:** Towards the end of the plotted training (steps 120-140), the GAPO line shows a slight recovery after a dip, while the MEL line continues its strong upward climb, reaching its peak at the final data point. The lines do not appear to converge. ### Interpretation The data suggests that for the "OlympiadBench" benchmark, the MEL training method or model architecture is more effective and robust than GAPO. MEL not only achieves a higher final validation score (~0.625 vs. ~0.575) but also demonstrates more stable learning dynamics after an initial adjustment period. The synchronized dip at step 20 is a critical investigative point. It could indicate a specific difficulty in the benchmark dataset encountered at that stage of training, a learning rate schedule effect, or a characteristic of the optimization landscape. The fact that both models recover but then diverge significantly implies that MEL is better equipped to overcome this hurdle and continue scaling its performance. The chart does not show signs of overfitting (a declining validation score) for either model within the 140 steps, suggesting that performance might continue to improve with further training, particularly for MEL. The primary takeaway is the clear superiority of the MEL approach on this specific task, both in terms of absolute performance and learning stability.

## Line Chart: Benchmark: OlympiadBench ### Overview The chart compares the validation scores of two models, **GRPO** (blue) and **MEL** (pink), across training steps (0–140) on the OlympiadBench benchmark. Both lines exhibit upward trends with fluctuations, but **MEL** consistently outperforms **GRPO** until the final step, where **GRPO** surpasses it slightly. ### Components/Axes - **X-axis**: Training Step (0–140, increments of 20). - **Y-axis**: Validation Score (0.450–0.625, increments of 0.025). - **Legend**: Located in the bottom-right corner. - **GRPO**: Blue line with circular markers. - **MEL**: Pink line with triangular markers. ### Detailed Analysis - **GRPO (Blue)**: - Starts at ~0.45 (step 0). - Rises to ~0.48 (step 20), dips to ~0.47 (step 40), then climbs to ~0.56 (step 60). - Peaks at ~0.61 (step 140). - Notable dip at step 80 (~0.56) and step 100 (~0.58). - **MEL (Pink)**: - Starts at ~0.45 (step 0). - Rises to ~0.53 (step 40), ~0.58 (step 60), ~0.59 (step 80), ~0.60 (step 100), ~0.60 (step 120), and peaks at ~0.625 (step 140). - Steady upward trend with minor fluctuations. ### Key Observations 1. **MEL** maintains a higher validation score than **GRPO** for most steps (e.g., ~0.58 vs. ~0.56 at step 80). 2. **GRPO** surpasses **MEL** only at the final step (140), with scores ~0.61 vs. ~0.625. 3. Both models show volatility in mid-training (steps 40–80), with **GRPO** experiencing sharper dips. 4. Final scores suggest **GRPO** achieves near-parity with **MEL** by step 140. ### Interpretation The data demonstrates that **MEL** initially outperforms **GRPO** in validation scores, likely due to architectural or training advantages. However, **GRPO**’s late-stage improvement (step 140) indicates potential for catching up with extended training. The volatility in mid-training suggests challenges in optimization or overfitting for both models. The final crossover at step 140 raises questions about **GRPO**’s scalability or efficiency in later training phases. This benchmark highlights trade-offs between model design and training duration for OlympiadBench performance.