## Bar Chart: Accuracy of LLM and PRM Combinations
### Overview
This bar chart displays the accuracy of different Large Language Models (LLMs) combined with different Program-aided Reasoning Models (PRMs) when evaluated using a Best-of-N approach on an AMC (presumably, a multiple-choice question answering dataset). The accuracy is measured in percentage points. The chart compares four LLM/PRM combinations.
### Components/Axes
* **Title:** "Accuracy of each LLM and PRM combination using Best-of-N on AMC" (Top-center)
* **X-axis:** LLM/PRM combinations: "Qwen2.5-7B-Instruct", "Eurus-2-7B-PRIME", "Phi-4-14B", "Qwen2.5-7B-AIRL-S(Our LLM)" (Bottom-center)
* **Y-axis:** Accuracy (%) - Scale ranges from 30 to 70, with increments of 5. (Left-side)
* **Legend:** Located in the top-left corner, identifying the data series:
* "Accuracy@1" (Pink)
* "Math-Shepherd-Mistral-7B-PRM" (Red)
* "EurusPRM Stage2" (Green)
* "Llama3.1-8B-PRM-Deepseek-Data" (Gray)
* "Qwen2.5-AIRL-S-PRM(Ours PRM)" (Dark Green)
### Detailed Analysis
The chart consists of four groups of bars, one for each LLM/PRM combination on the x-axis. Each group contains five bars, representing the accuracy of each of the five models/combinations listed in the legend.
* **Qwen2.5-7B-Instruct:**
* Accuracy@1: ~33.7%
* Math-Shepherd-Mistral-7B-PRM: ~53.0%
* EurusPRM Stage2: ~54.2%
* Llama3.1-8B-PRM-Deepseek-Data: ~55.6%
* Qwen2.5-AIRL-S-PRM(Ours PRM): ~56.6%
* **Eurus-2-7B-PRIME:**
* Accuracy@1: ~61.4%
* Math-Shepherd-Mistral-7B-PRM: ~63.9%
* EurusPRM Stage2: ~65.1%
* Llama3.1-8B-PRM-Deepseek-Data: ~63.9%
* Qwen2.5-AIRL-S-PRM(Ours PRM): ~56.6%
* **Phi-4-14B:**
* Accuracy@1: ~44.6%
* Math-Shepherd-Mistral-7B-PRM: ~59.0%
* EurusPRM Stage2: ~60.2%
* Llama3.1-8B-PRM-Deepseek-Data: ~61.4%
* Qwen2.5-AIRL-S-PRM(Ours PRM): ~62.6%
* **Qwen2.5-7B-AIRL-S(Our LLM):**
* Accuracy@1: ~59.0%
* Math-Shepherd-Mistral-7B-PRM: ~63.9%
* EurusPRM Stage2: ~65.1%
* Llama3.1-8B-PRM-Deepseek-Data: ~65.1%
* Qwen2.5-AIRL-S-PRM(Ours PRM): ~67.5%
### Key Observations
* The "Accuracy@1" consistently shows the lowest accuracy across all LLM/PRM combinations.
* "Qwen2.5-AIRL-S-PRM(Ours PRM)" generally yields the highest accuracy, especially when combined with "Qwen2.5-7B-AIRL-S(Our LLM)".
* "EurusPRM Stage2" and "Llama3.1-8B-PRM-Deepseek-Data" often perform similarly, with accuracies clustered around the 60-65% range.
* "Math-Shepherd-Mistral-7B-PRM" consistently improves accuracy compared to "Accuracy@1" but is often lower than the other PRMs.
* The largest performance gap for a single LLM/PRM combination is observed for "Qwen2.5-7B-Instruct", where "Accuracy@1" is ~33.7% and "Qwen2.5-AIRL-S-PRM(Ours PRM)" is ~56.6%.
### Interpretation
The data suggests that incorporating PRMs significantly improves the accuracy of LLMs on the AMC dataset. The "Qwen2.5-AIRL-S-PRM(Ours PRM)" consistently demonstrates the best performance, indicating that the researchers' PRM is particularly effective. The relatively low accuracy of "Accuracy@1" suggests that the Best-of-N approach, utilizing multiple reasoning steps, is crucial for achieving higher accuracy. The variations in performance across different LLM/PRM combinations highlight the importance of selecting compatible models for optimal results. The fact that the "Our PRM" consistently performs well with different LLMs suggests it is a robust and generalizable improvement. The data does not provide information on the computational cost or efficiency of these combinations, only their accuracy.
