IC-213Qwen2-0.5B-Instruct uses English-specific past tense heads and late FFN layers for morphological marking that is absent in Chinese

Ruochen Zhang, Qinan Yu, Matianyu Zang, Carsten Eickhoff, Ellie Pavlick

SourceThe Same but Different: Structural Similarities and Differences in Multilingual Language Modeling

Path patching and ablation on Qwen2-0.5B-Instruct show that the past tense task uses shared copy heads (21.3, 21.4, 21.13) for both English and Chinese, but English-specific past tense heads (19.4, 19.5) and FFN layers (20-24) that are causally relevant only for English. Ablating the past tense heads promotes present tense verbs by 83.21 positions in English while barely affecting Chinese (4.58 positions). Ablating FFN layers 20-24 drops English accuracy from 97.44% to 47.44% and zero-rank from 58.06% to 17.74%, while Chinese zero-rank changes only from 25.08% to 24.19%. The copy heads, by contrast, are essential for both languages.

Evidence
interventional
Key metric
FFN ablation (layers 20-24): EN accuracy 97.44%→47.44%, zero-rank 58.06%→17.74%; ZH zero-rank 25.08%→24.19%. Past tense head ablation: EN present tense promoted 83.21 positions, ZH correct verb +4.58 positions. Copy head ablation: EN past tense rank -141.95, present tense -221.39; ZH verbs -201.82.
Caveat
The authors note that the semantics of the verb is not entirely decoupled from the tense, and that the Chinese task lacks minimal pairs so accuracy cannot be computed; the analysis is limited to one language pair and one morphological phenomenon.
Model
Qwen 2 Qwen2-0.5B-Instruct
Concepts
Depth-dependent structure
Methods
Path Patching [primary]
Related work
Merullo et al. 2023 (FFN for past tense) [builds-on], Ferrando & Voita 2024 (information flow routes) [context]
Related findings
IC-211, IC-212
Extraction
automatic-extraction