Large language models make consistent moral judgments, but whether these tendencies correspond to the moral makeup of public web text in the same language is not well understood. We test this correspondence for deception dilemmas across Chinese and English in a four-layer empirical study. We use DeepSeek-R1 for ICL probing, Llama-3.1-8B and Qwen2.5-7B for fine-tuning, and GPT-4o for cross-model replication. In both languages, roughly 60% of web text favors utilitarian reasoning (59.8% vs. 62.3%), inconsistent with a purely sociolinguistic explanation. In-context learning shifts model decisions sharply in low-resistance scenarios (Hostage Crisis, V = 0.411, p = 0.001). In a high-resistance scenario (Organ Allocation), one DeepSeek-R1 deployment showed near-zero ICL effects (V = 0.067) at small sample size, but a larger replication on another deployment showed a large effect (V = 0.55 to 0.69), indicating deployment-specific rather than stable resistance. Three findings stand out for multilingual alignment. First, fine-tuning on deontological data reverses direction in English (Llama-EN Deon P(A) rises from ∼0% to 88%), showing that SFT value transfer is not language-neutral. Second, corpus proportions do not carry over to forced choices. The English Llama-3.1-8B base model gives P(A) near zero even though 62.3% of the web text is utilitarian. Third, apparent moral damping in high-resistance scenarios can reflect sample size and deployment variability rather than a stable property of the model or scenario. These results show that corpus moral distributions correspond to model behavior in many conditions, but language, post-training, model architecture, and deployment variability each introduce points where this correspondence breaks.

Large Language Models (LLMs) are increasingly deployed as agents that invoke external tools through structured function calls. While recent work reports strong tool-calling performance under standard English-centric evaluations, the robustness of tool calling under multilingual user interactions remains underexplored. In this work, we introduce MLCL, a diagnostic benchmark, and conduct a systematic evaluation of multilingual tool calling across Chinese, Hindi, and the low-resource language Igbo. Through fine-grained error analysis, we show that many failures occur despite correct intent understanding and tool selection. We identify parameter value language mismatch as a dominant failure mode, where models generate semantically appropriate parameter values in the user’s language, violating language-invariant execution conventions. We further evaluate several inference-time system strategies and find that while these strategies substantially reduce language-induced execution errors, none of them can fully recover English-level performance.