The linear-dynamic-stochastic model of a reaction latency as applied to avoidance experiment is presented. Reactions are classified on the basis of the model into following classes: escape, avoidance, late avoidance, three types of inter-trial responses and "no reaction". The experimental latency distribution is split into latency distributions of the escape, avoidance and late avoidance responses, providing a new insight into latency distribution. The results of fitting the model to latency measurements obtained in the avoidance conditioning experiment are presented. The same processes of the parameter changes as in the escape conditioning are discovered, one causing a latency to decrease and the other causing a latency to increase during learning. The first process affects a latency stronger than the second and, consequently, the latency decreases during learning. The second process is responsible for a decay of inter trial-responses during experiment. The value of the correlation coefficient between the threshold of avoidance reaction and the threshold of escape reaction was also estimated. Negative values of this coefficient were obtained, therefore, on the average, the greater the avoidance reaction threshold the smaller the escape one. In the course of learning the correlation coefficient tends to be equal to - 1, i.e., as a result of training, both thresholds became dependent in a functional (non-random) way. This result may provide an objective index of the "state of learning". The model provides a new tool for analysis of results of latency-based experiments.