In situations where the steepest ascent algorithm fails to yield a significant improvement in the log likelihood function, the user has the option of specifying that the algorithm reverts to taking a Newton-Raphson step. This optimization method is known as the Steepest Ascent with reset.
Cohen, J. (1997). An Overview of AIR’s Direct Estimation Software for Marginal Maximum Likelihood Models. Washington, DC: American Institutes for Research.