Lexical tone sandhi in Northern Wu Chinese languages typically adopts a left-dominant mechanism, usually characterized as a tone extension pattern from the leftmost tone rightwards spanning the whole sandhi domain. By default, the underlying tone of the leftmost tone should be the only major factor conditioning the sandhi outputs in a lexical sandhi domain. However, variations still exist in some languages, the causes of which have not been studied extensively. This study examines the categorical variations in disyllabic lexical tone sandhi in Xiangshan Chinese, an under-studied Northern Wu variety. A perceptual analysis of 106 disyllabic lexical compounds produced by 8 native Xiangshan speakers revealed 4 distinctive sandhi patterns. These patterns were cross-checked by independent k-means clustering based on acoustic data. Further analyses of the cluster distributions suggest that the historical tonal categories of the initial tones, and individual speakers, serve as the major sources of the variations. The results provide evidence supporting the reality and relevance of historical tonal categories in tone sandhi, and suggest distinct development paths for sandhi tones and citation tones. Inter-speaker variations largely explain the remaining variation and may reflect different stages of ongoing language change.