Korean stem-final conjugations illustrate the interaction between lexical exceptionality and heterogeneous phonological processes. When /p/-, /t/-, and /s/-final stems occur before vowel-initial suffixes, the irregular classes in these paradigms undergo intervocalic lenition, each exhibiting a distinct alternation pattern. Learners must therefore not only identify which roots trigger lenition, but also determine the corresponding repair strategy. This study investigates how lexically-specific phonological patterns are acquired when multiple repair strategies are available. We employ a lexically scaled MaxEnt model (Linzen et al., 2013; Hughto et al., 2019) to learn these paradigm-specific alternations and run simulations under two learning scenarios: (1) when repair strategies occur at equal frequencies and (2) when one strategy significantly outnumbers the others. Results show that the model favors a least-cost solution by treating statistically dominant morpheme classes as the general pattern. We conclude by discussing the model’s sensitivity to lexical statistics, predictions for empirical testing, and implications for language acquisition.