Figure 2.

Impact of model scaling on performance across question and case levels. The line chart illustrates the performance of the Qwen2.5 series across different parameter sizes (3B, 7B, 14B, and 32B). The upper line represents the question pass rate (QPR), while the lower line indicates the case pass rate (CPR). The shaded region between the 2 curves highlights the consistency gap, which quantifies the disparity between question-level accuracy and case-level accuracy. As the model scale increases, both metrics improve, and the consistency gap decreases from 26.3% (3B) to 13.5% (32B), indicating performance degradation when moving from question-level accuracy to case-level aggregate accuracy.