Figure 3(b) visibly labels its two reward panels as MATH and Countdown and shows random-mask SPG suffering severe late-training degradation while SPG+IM remains stable; however, the surrounding Section 4.2 prose describes the affected tasks as GSM8K and MATH. The fixed source is internally inconsistent about whether the non-MATH collapse panel is Countdown or GSM8K, so task identity for that panel remains unresolved rather than being silently normalized. · CiteArk
Figure 3(b) visibly labels its two reward panels as MATH and Countdown and shows random-mask SPG suffering severe late-training degradation while SPG+IM remains stable; however, the surrounding Section 4.2 prose describes the affected tasks as GSM8K and MATH. The fixed source is internally inconsistent about whether the non-MATH collapse panel is Countdown or GSM8K, so task identity for that panel remains unresolved rather than being silently normalized.
来源：paper-fixed:PDF pp. 7 and 9, Section 4.2 Q4, Figure 3(b), and visible panel labels
