Sets the maximum acceptable absolute value of a coefficient in the lower bounding problem. If a coefficient is greater than this number, the solver will use other, less effective methods instead of solving the lower bounding linear problem (LP).
Coefficients are naturally improved through branching. However, if the problem is not numerically well-behaved, increasing this tolerance will force the solver to solve the ill-posed lower bounding problems. This is small-ish by default because the global solution guarantee can be compromised otherwise, but in practice it’s perfectly fine to increase this most of the time.
Your clue that there might be a problem relating to this is that an (MI)LP relaxation was generated and you get a gap that never improves.