*** Please feel free to interact with me real-time on Zoom at https://ucsd.zoom.us/j/5803989715. I will be on-line from 12-2PM EDT on Saturday.*** Speakers can sometimes choose between multiple structures to describe the same event (e.g., prepositional vs. double object datives in English). However, typologically different languages have different degrees of flexibility. For example, languages that allow scrambling have more flexible word orders than those that do not. It is, however, unclear whether these patterns reflect internal cognitive biases or are learned through exposure. Learning structural alternations is thus a balancing act of capturing regularities while avoiding overgeneralization from limited data (e.g., using PD and DO with “give” but only PD with “donate”). Here we ask whether learners generalize structural alternations when learning a typologically different grammar, and if so, what guides the generalization. In a series of two miniature novel-language learning experiments, we tested usage-based/statistical accounts against an internal bias account. Usage-based/statistical accounts predict that learners should largely constrain their production and acceptability of alternations to the verb that they learned the structures with and show limited generalization to other verbs. In contrast, an internal bias account predicts that learners who are exposed to scrambling should show a more liberal generalization pattern than those who are not, if learners indeed have tacit knowledge that scrambling signals flexibility. English monolinguals (N = 48) described “give”, “hand”, and “show” dative events in a Korean-English hybrid language (Korean word order, case markers, and verbs, but English nouns). In Experiment 1, all participants were exposed to structural alternations only in “give” sentences. The non-scrambling group was exposed to the canonical order (NOM-DAT-ACC) and a verb-phrase internal alternation (NOM-ACC-DAT). The scrambling-group was exposed to scrambling (DAT-NOM-ACC) in addition to the orders shown to the non-scrambling group (Table 1). Both groups saw only the canonical order for “hand” and “show”. In a subsequent picture description task with novel pictures, scrambling group learners produced the alternated structure (NOM-ACC-DAT) equally often with all verbs, but non-scrambling group learners produced the alternated structure more with “give” than with other verbs (interaction p = .03, Figure 1a). The analogous pattern was found in a subsequent acceptability judgment task (interaction p = .005, Figure 1b). However, we could not rule out the explanation that the scrambling group simply had a harder time tracking verb-specific statistics because they learned one more distinct structure. In Experiment 2, the scrambling group only learned the canonical structure and the scrambled alternation with “give” (Table 1), eliminating this confound. In line with Experiment 1, the non- scrambling group showed patterns of verb-wise conservative generalization in their production (Figure 2a) and acceptability ratings (Figure 2b), whereas the scrambling group generalized across all verbs, though only acceptability results reached statistical significance (interaction p = .03). Post-hoc analyses revealed that Experiment 2 might not have enough power to detect verb differences because 36/48 participants produced alternations in Experiment 1, whereas only 25/48 did in Experiment 2. When combining data from participants who produced alternations from both experiments, the conservativeness in the non-scrambling group was replicated (Figure 3a and 3b). Altogether, learners were verb-wise conservative only when they saw no scrambling (i.e., when they saw no evidence that the structural alternations occur outside verb phrases). This suggests that learners do not merely track statistical patterns in the input but use internal linguistically sophisticated biases to generalize structural alternations.