Tag
研究者通过使用 LLM 翻译英文语料构建了日语 SDRT 语料库，并对比了基于提示和微调的 LLM 在日英语篇解析上的表现，发现两者性能相近，表明 SDRT 语篇解析具有跨语言适用性，且模型更依赖上下文语篇关系而非话语标记词。
This paper evaluates cross-lingual fairness of edit-based homoglyph and invisible-character watermarking across 40 language varieties in 24 scripts, finding that removing tokenizers does not eliminate inequity since script, not resource level, drives 96% of capacity variation.
This paper causally analyzes how multilingual LLMs represent negation across English, French, Vietnamese, and Chinese, showing that a shared, causally controllable negation direction—and even a single sparse autoencoder feature per model—can induce or remove negation across languages and generalize to unseen ones.
This paper analyzes semantic synchronization of operator manuals over ISO 11783 CAN buses to enable offline, multilingual LLM assistance on agricultural machinery, benchmarking Markdown, JSON, and XML transfer and evaluating open-weight LLMs with retrieval-augmented generation across multiple languages.
该论文研究英文构建的 LLM 管道能否跨语言复用，通过支票提取、Mintaka 问答和 FinanceBench 三个任务评估准确性、推理努力与行为稳定性，结论是可有条件复用但仍需按语言按领域验证。
PRISM is a framework for evaluating how activation interventions transfer across precision, languages, and model scales in tool-calling language models, showing that quantized NF4 labels and cross-lingual supervision can expose and mitigate unsafe tool-call behaviors even in large models up to 70B parameters.
该论文提出三阶段流程识别多语言大模型中的"Bridge Routing Heads"（BRH），发现五种语言对应的头集合几乎完全互斥，证明多语言多跳推理由语言特异电路实现；消融实验显示因果效应，并通过激活级干预无需训练即可挽救高达51.7%的跨语言推理失败。
BARRAC adapts an English aspect-based sentiment analysis framework to Arabic dialect classification tasks by substituting consumer-review attribute pools with Arabic linguistic devices for sarcasm, dialect, and sentiment, achieving 63.93% mean macro-F1 and outperforming GPT-4o on four of five benchmarks.
This paper introduces Modl, a technique for creating cross-lingual remote-sensing multimodal large language models by composing domain and language LoRAs with mutual orthogonality, achieving superior performance without paired multilingual data.
This paper analyzes reasoning traces in large reasoning models for machine translation, finding that reasoning benefits are conditional and introducing Hierarchical Meta-Summarization to understand trace patterns.
This paper introduces a multilingual dataset of ethical dilemmas to probe LLMs' value preferences and proposes a task vector composition method for modular ethical alignment without retraining.
This paper proposes a phonemically comprehensive, ASCII-only romanization scheme for Thai and Lao, designed for cross-lingual correspondence and machine processing to aid language learning and speech processing.
This research paper investigates why pretraining in large language models fails to transfer knowledge across languages, identifies disjoint token spaces as a fundamental barrier, and proposes mapping languages to a shared token space to improve cross-lingual generalization.
This paper describes six submissions to the WMT26 Model Compression Shared Task, using routing-informed expert pruning and MXFP4 quantization to compress GPT-OSS-20B into smaller translation models with parameters ranging from 4.186B to 7.770B.
This study evaluates cross-lingual clinical annotation projection as constrained text generation, comparing supervised, hybrid, and large language model methods across six languages. Direct LLM projection outperforms previous methods, enabling high-quality multilingual clinical annotation transfer.
The paper proposes SALT, a lightweight post-training method that injects span-level supervision into cross-lingual sentence encoders to improve token representations, achieving top results on multilingual token-level benchmarks and enhancing sentence-level performance.
SWORD introduces a Wikidata-based benchmark to evaluate cross-lingual inconsistencies in LLMs' rejection of factual errors, revealing that models rely on distributional familiarity and show performance degradation in Asian languages.
The paper proposes the interlingua hypothesis, suggesting that large language models perform translation by encoding source text into a latent task-agnostic feature space and decoding from it, supported by empirical evidence on variance, causal influence, and monolingual fine-tuning.
This paper proposes a cross-lingual romanization ecosystem for Sinitic languages, develops specific schemes for Mandarin and Cantonese, and shows improved performance in speech-to-romanization tasks compared to baseline methods.
The paper proposes a latent-space intervention method to improve cross-lingual factual consistency in large language models, achieving better alignment without loss in factual accuracy.