arXiv:2610.18420v1 Announce Type: cross Abstract: We trained Agaram 2.5B, a 2.5-billion-parameter Tamil-English language model, from scratch on about 50 billion tokens (58% Tamil, the rest English, m...
arXiv:2608. 13695v1 Announce Type: cross Abstract: Large language model providers routinely cite multilingual safety benchmarks spanning a dozen or more languages as evidence that their models are safe for non-English-speaking users.
VakyArth is the first pragmatic benchmark for Indic languages, covering Hindi, Punjabi, Tamil, and Malayalam. It tests models on five pragmatic phenomena—deixis, speech acts, implicature, social pragmatics, and coherence—using multiple-choice questions, natural language inference, and translation tasks authored by native speakers. Evaluation of multilingual LLMs shows consistent failures on pragmatic meanings rooted in Indic linguistic and cultural conventions, with systematic differences across languages and tasks.
arXiv:2608.30092v1 Announce Type: cross Abstract: We present Arkios, a 1.04B-parameter dense transformer pretrained from scratch on 150B tokens of bilingual English-Nepali text, using a custom single...
arXiv:2610.02926v1 Announce Type: new Abstract: Standard factual benchmarks assume clean monolingual prompts and exact-match scoring, two assumptions that break simultaneously in real-world multiling...
arXiv:2507.23248v2 Announce Type: replace-cross Abstract: Bengali is spoken by more than 230 million people, yet no standardized instrument evaluates large language models (LLMs) on Bengali across th...
The study investigates how exposing large language models to Arabic translations of English benchmark tasks can hide data contamination from standard English-only detection methods. By testing four instruction-tuned models with increasing Arabic exposure, the authors find that traditional probes like TS-Guessing and Min-K%++ fail to detect contamination, yet English performance still improves. They propose a new Translation-Aware Contamination Detection (TACD) method that uses cross-lingual consistency and choice reordering to reveal contamination effects that are otherwise invisible to English-only diagnostics.
Large language models are increasingly deployed as general-purpose educational and technical assistance systems, but their underlying infrastructure does not treat languages equally. One underexamined source of disparity is tokenization: semantically equivalent content can require substantially different token counts across languages, affecting API cost, latency, and usable context length before a model is invoked.
The paper presents a cross‑lingual legal QA system for Vietnamese labour law, introducing a bilingual evaluation suite of 231 Vietnamese–English question–answer pairs, 75 of which are annotated for five complex legal reasoning phenomena. It evaluates a verifier‑guided pipeline that decomposes answers into claims, checks citation reachability and entailment, and corrects citation failures and contradictions, and introduces six automatic diagnostics for faithfulness to retrieved evidence. Experiments show that dense retrieval outperforms sparse and hybrid retrieval, translation placement has no significant effect on diagnostics, and verifier‑guided correction modestly improves citation preservation but not other dimensions, with human evaluation indicating a gap between automatic diagnostics and human judgments.
arXiv:2605.29637v2 Announce Type: replace Abstract: Large language models often exhibit a substantial gap between their performance in English and in lower-resourced languages on equivalent knowledge...
Large language models (LLMs) process text through subword tokenizers rather than directly reading characters or words. Because these tokenizers are trained predominantly on English-centric corpora, they introduce a systematic and often overlooked disadvantage for many non-English languages.
The paper investigates whether multilingual language models transfer factual knowledge from one language to another during continued pretraining. Using an English-pretrained model continued on Persian data with systematically removed facts, the authors create SIFT, a dataset of 500 triples across 20 relations, split by cultural origin. Their findings indicate that factual transfer is minimal, especially for Persian-related facts, and that simple removal strategies or easy negative candidate sets can overestimate transfer.