The input modification for (presumably numerical) "stability" introduces a bias which is then improperly handled by taking the negative beam probability.
Ensure that this change doesn't cause numerically unstable results/investigate alternative workarounds and then merge.
The input modification for (presumably numerical) "stability" introduces a bias which is then improperly handled by taking the negative beam probability.
Ensure that this change doesn't cause numerically unstable results/investigate alternative workarounds and then merge.