Mandarin Wh-Indefinites Follow Nonveridicality in Multilingual LLMs, Unlike English any
Abstract
Multilingual LLMs learn mostly from English. Do they carry English polarity licensing over to other languages? English any is licensed by downward-entailing (DE) contexts, Mandarin wh-indefinites such as shenme 'what/something' by nonveridical ones, and the two classes come apart: perhaps licenses shenme but not any, while only licenses any but entails existence. We test this dissociation in a pre-registered study of twelve open LLMs (0.5B-9B, four families), using 48 matched bilingual frames in 21 contexts and a log-probability difference-in-differences normalised by negation. All 11 models that pass the measurement gate license Mandarin wh-indefinites under possibility, necessity and evidential adverbs, at 0.42-2.33 times the strength of negation, while English any gets nothing in the same contexts. An exploratory extension to Qwen2.5-32B shows the same. The English side is weaker: any under only, first and regret is licensed but at a fraction of negation, so only 3 models show the full double dissociation in sentence probabilities. When prompted to choose between sentences, most gated models show it. In exploratory directional ablations, a direction that carries English licensing, held-out classes included, does not reach Mandarin licensing in most models. The models do not simply apply the English rule to Mandarin.
Figures & tables
| Nonveridical, not DE (prediction: zh , en ) | DE, existence-entailing (prediction: zh , en ) | |||||||||||||
| Model | POSS | NEC | EVID | POSS | NEC | EVID | zh | ONLY | FIRST | REGR | ONLY | FIRST | REGR | en |
| Qwen2.5-0.5B-Inst | 0.92 | 0.73 | 1.63 | 0.13 | 0.07 | 0.15 | ✓ | 0.57 | 0.49 | 0.47 | 0.24 | 0.51 | 0.28 | |
| Qwen2.5-1.5B-Inst | 0.68 | 0.61 | 1.76 | 0.06 | 0.03 | 0.06 | ✓ | 0.03 | 0.28 | 0.86 | 0.50 | 0.50 | 0.34 | |
| Qwen2.5-3B-Inst | 0.56 | 0.42 | 0.95 | 0.09 | 0.08 | 0.00 | ✓ | 0.44 | 0.33 | 0.46 | 0.10 | 0.59 | 0.38 | |
| Qwen2.5-7B-Inst | 0.76 | 0.48 | 1.51 | 0.08 | 0.01 | 0.10 | ✓ | 0.62 | 0.36 | 0.43 | 0.19 | 0.45 | 0.21 | |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Context | zh/en | Mandarin (T) | English (T) |
|---|---|---|---|
| BASE | / | 张伟昨天买了什么东西… | Tom bought anything yesterday… |
| VER1 | / | 张伟昨天确实买了什么东西… | Tom indeed bought anything yesterday… |
| VER2 | / | 张伟昨天果然买了什么东西… | Tom predictably bought anything yesterday… |
| VER3 | / | 张伟昨天终于买了什么东西… | Tom finally bought anything yesterday… |
| VER4 | / | 张伟昨天已经买了什么东西… | Tom already bought anything yesterday… |
| MAN1 | / | 张伟昨天偷偷买了什么东西… | Tom secretly bought anything yesterday… |
| Mandarin licensing removed | English licensing removed | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| by | random r95 | by | by | random r95 | by | |||||||
| Model | NEG | ZA | NEG | ZA | NEG | ZA | NEG | EA | NEG | EA | EA | category |
| Qwen2.5-0.5B-Inst | 0.15 | 0.13 | 0.29 | 0.17 | 1.22 | 0.71 | 0.58 | 0.36 | 0.17 | 0.44 | 0.05 | neither |
| Qwen2.5-1.5B-Inst | 0.10 | 0.00 | 0.32 | 0.39 | 0.43 | 0.03 | 0.65 | 0.48 | 0.06 | 0.16 | 0.27 | neither |
| Qwen2.5-3B-Inst | 0.10 | 0.19 | 0.25 | 0.19 | 0.31 | 0.39 | 0.49 | 0.77 | 0.10 | 0.35 | 0.09 | neither |
| Qwen2.5-7B-Inst | 0.00 | 0.30 | 0.16 | 0.12 | 0.49 | 0.68 | 0.27 | 0.39 | 0.05 | 0.13 | 0.00 | neither |
| Mandarin licensing removed | English licensing removed | |||||||||||
| by | random r95 | by | by | random r95 | by | |||||||
| Model | NEG | ZA | NEG | ZA | NEG | ZA | NEG | EA | NEG | EA | EA | category |
| Qwen2.5-0.5B-Inst | 2.82 | 1.79 | 1.48 | 0.49 | 4.93 | 1.82 | 1.31 | 2.95 | 0.16 | 1.26 | 1.04 | partial |
| Qwen2.5-1.5B-Inst | 0.42 | 0.43 | 1.60 | 0.75 | 3.88 | 1.24 | 1.01 | 1.71 | 0.12 | 0.31 | 0.54 | neither |
| Qwen2.5-3B-Inst | 0.37 | 0.24 | 0.22 | 0.56 | 0.36 | 0.15 | 0.83 | 2.06 | 0.10 | 0.32 | 0.30 | neither |
| Qwen2.5-7B-Inst | 0.75 | 0.39 | 0.67 | 0.18 | 1.39 | 0.09 | 0.71 | 1.54 | 0.07 | 0.51 | 0.27 | separate |
Cross-Lingual Transfer
Cross-lingual transfer enables models to apply knowledge or capabilities learned in one language to tasks in another.
See all 268 papers