Contextual licensing of agreement and anaphora: evidence from Korean–Spanish structures and model-based probes

Abstract This study employs Large Language Model (LLM)-based translation as a diagnostic probe for generative predictions regarding Agree and φ-feature valuation in the context of discourse-driven underspecification. The translation direction from Korean – a topic-prominent, radical pro-drop language lacking overt gender marking – to Spanish – a language requiring morpho-syntactically explicit realization of person, number, and gender – compels a process of feature reconstruction rather than direct transfer. Using controlled context-window probes on a 22,800-sentence parallel corpus, the study analyzes how non-biological learners reconstruct missing features. The results show that expanding the discourse context significantly mitigates person mismatches via antecedent accessibility, yet fails to eliminate gender mismatches. This asymmetry reflects the structural scarcity of gender cues in the Korean input, consistent with Minimalist expectations: discourse context facilitates antecedent selection, yet the reliable valuation of specific lexical φ-features necessitates richer formal specification than the source provides. Furthermore, model architecture crucially determines the reconstruction strategy, ranging from conservative non-commitment to the over-generation of spurious feature values. By treating LLMs as learners devoid of Universal Grammar (UG) but shaped by architectural biases – including the attention mechanisms, positional encoding schemes, and fine-tuning objectives that constrain how each model integrates contextual information – this study demonstrates how computational probes can test whether grammatical generalizations depend on structurally specified input or can be recovered from distributional context alone.