The study investigates how multilingual large language models (LLMs) handle negation across languages. Using causal interventions on synthetic and FLORES-200 derived datasets, the authors find that negation representations are causally controllable and largely shared across English, French, Vietnamese, and Chinese. They demonstrate that a single sparse autoencoder feature can control polarity across languages and that these representations generalize to additional languages.
arXiv:2609.01356v1 Announce Type: new Abstract: Multilingual large language models (mLLMs) achieve strong performance in machine translation, yet our understanding of the mechanisms by which they tra...
The paper introduces NegCue, a large-scale dataset of 1.8 million samples that includes single-word, multi-word, and affixal negation cues, totaling over 200 unique forms. The authors pre-train encoder-only language models and large language models on this dataset to study how different negation types influence understanding. Experiments on five downstream benchmarks reveal that affixal negations provide the most significant performance gains, whereas single-word negations yield modest improvements, and that additional pre-training benefits both model types.
arXiv:2608. 07629v1 Announce Type: cross Abstract: Multilingual neural machine translation models such as NLLB-200 cover 200 languages but leave thousands unsupported, including most Grassfields Bantu languages of Cameroon.
arXiv:2604. 03532v2 Announce Type: replace-cross Abstract: Large language models (LLMs) show strong multilingual capabilities, yet reliably controlling the language of their outputs remains difficult.
Sparse autoencoders (SAEs) are proposed to extract numerous features from large language model (LLM) representations, yet explaining these features still relies primarily on external observation. This reliance leads to superficial explanations inferred from observed model behavior and computational inefficiency from collecting such behavioral evidence at scale.
The paper compares generative and encoder-based neural models for multilingual Named Entity Recognition (NER) across the eleven languages of the Naamapadam benchmark. Five classic model families, four decoder-only large language models fine‑tuned with LoRA and 4‑bit NF4 quantisation, and nine generative models in zero‑to‑5‑shot inference were evaluated under strict CoNLL span‑level metrics. Encoder-based models (mBERT and XLM‑R) achieved substantially higher F1 scores—up to 0.675 on Hindi—than any generative architecture, with gaps of 7.5–40 percentage points; the best few‑shot result reached only 28% of the encoder baseline. The study identifies three language clusters (encoder‑dominant, partial‑coverage, and failure‑zone) and offers deployment guidelines based on transfer learning and low‑resource NLP principles.
Large language models show strong cross‑lingual abilities, but previous studies have examined them only at isolated training points. This work pretrains a 1.7B multilingual model on nine languages, recording checkpoints at fine granularity to track translation development. Using word‑level translation and a new dataset, the authors find that early learning is dominated by token‑level copying and surface similarities, while later stages involve more general translation mechanisms and refined copying.
The paper proposes the interlingua hypothesis, suggesting that large language models translate by encoding a source sentence into a latent, task‑agnostic feature space and then decoding a target sentence from that space. Three lines of evidence support this: (1) BLEU variance across language pairs is largely explained by language‑specific competences without pair‑specific interactions; (2) many model components influence both monolingual and translation tasks; and (3) fine‑tuning on monolingual data recovers most translation gains seen with aligned documents. These findings converge to support the hypothesis and point toward new ways to understand and improve LLM translation.
arXiv:2609.07876v1 Announce Type: cross Abstract: Recent methods in language model interpretability employ techniques such as sparse autoencoders to decompose residual stream contributions into linea...
The study compares an English-only and a bilingual decoder-only model, each 310 M parameters, trained on eight diverse languages while controlling for English exposure, compute, and document overlap. After aligning on shared English vocabulary, the authors find that token embeddings appear similar, but the deeper hidden states used for prediction diverge across models. This hidden‑state mismatch grows through middle transformer layers and persists despite controls, indicating that contextual processing differs between the models. "whyItMatters":"The findings show that embedding alignment can conceal significant internal representation differences, which is crucial for any downstream work that assumes aligned multilingual models are interchangeable."
arXiv:2609.01322v1 Announce Type: cross Abstract: In many settings, studying causal questions based on text data requires adjusting for confounding information within texts. Yet there is a tradeoff i...
arXiv:2608. 12334v1 Announce Type: cross Abstract: Despite the impressive multilingual capabilities of Large Language Models, the latent dynamics dictating language selection remain poorly understood.