They introduced,

• Higher clip for more exploration
• Dynamic sampling in case of 0 advantage (all same reward across group)
• Token level loss normalization to fix bias based on response length
• Overlong filtering
• Soft overlong punishment

References