Question about GODEL-XL experimental in paper

[zh215021]

Table 6 of the paper compares the effect of full fine-tuning on four tasks between GPT-3 and GODEL-XL. Why does this result look worse than GODEL-L in table 3-5 under a larger scale (GODEL-L is set with few-shot).