HOTFIXR: Hardness Optimized Training data For Improving X-Lingual Reasoning is a data generation framework that uses models to probe and learn a student model's multilingual weaknesses, and generates data to mitigate them that can improve multilingual performance.
Abstract
Language-specific competency (LSC) is the phenomenon of a language model performing better or worse depending on the language of the prompt. In other words, a language model outputs different (and potentially incorrect) responses to the same semantic query when prompted in different languages. Prior work attributes this to an internal misalignment of semantic representation across languages. Currently, there are two main approaches to address LSC in the literature: (1) routing all queries through English, improving performance, but limiting language expressivity to English; or (2) training on language-balanced data, equalizing model performance across languages, but reducing overall performance. In this work, we take a data centric perspective and introduce HOTFIXR: Hardness Optimized Training data For Improving X-Lingual Reasoning. It is a data generation framework that uses models to probe and learn a student model's multilingual weaknesses, and generates data to mitigate them. HOTFIXR can generate multilingual synthetic training data that can improve multilingual performance. We evaluate on three in-distribution tasks, three out-of-distribution tasks, and four out-of-distribution languages. On average, HOTFIXR (1) improves in-distribution performance by 6.2%, (2) reduces catastrophic forgetting (induced by fine-tuning) on OOD tasks by 3.7%, and (3) on OOD languages by 7.1%. Overall, as many real-world applications requires multilingual LLMs, our work contributes to the efforts of making LLMs multilingually proficient. We will release code upon acceptance.
This paper tests whether prompting the same (frozen) SLM in three typologically diverse languages and aggregating the outputs can improve classification without retraining or translation, and suggests that cross-lingual diversity rather than surface-level input variation drives the gain.
It is found that, in this benchmark, supervised fine-tuning (SFT) provides a strong baseline, substantially improving argument language consistency and end-to-end function call accuracy and, under consistent model selection, SFT achieves performance comparable to, and sometimes exceeding more complex reinforcement learning (RL) approaches.
Siddharth Chauhan, Thomas Butler, Abhishek Singhania et al.· 0 citations
Centroid Intervention Fusion is proposed, a projection fusion framework that consolidates multiple multilingual intervention projections into a single language-shared operator and outperforms the strongest prior pairwise intervention baseline by up to +3.3% across four model backbones.
Large Language Models (LLMs) have achieved remarkable performance largely due to the availability of high-quality and large-scale training data. However, this reliance creates a significant performance disparity between high-resource languages such as English and low-resource languages like Arabic. The imbalance in data exposure during the pre-training phase persists into the bilingual supervised fine-tuning (SFT) stage, limiting the models’ semantic depth and reasoning capabilities in low-resource languages.To address the aforementioned imbalance, we depart from traditional approaches by computing the optimal learning rates for each language in the dataset separately and transforming these learning rates into language-specific weighting coefficients used in the loss function. Experimental results obtained on the XQuAD test set after training on the MURI dataset demonstrate that the proposed method preserves performance in English while significantly improving Arabic performance, achieving 56.36 F1 and 40.27 Exact Match (EM) scores.
D. Özcan, Çağrı Yeşil, Hasan Saribas et al.· Signal Processing and Commun...· 0 citations
We present L3Cube-IndicQuest v2, a large-scale gold-standard multilingual question-answering benchmark for evaluating the India-specific factual knowledge of Large Language Models (LLMs). The benchmark comprises 3,471 curriculum-grounded English question--answer pairs spanning nine domains, curated from educational curricula, competitive examination materials, and domain-specific reference books. We introduce a practical hybrid construction strategy that combines context-grounded LLM-based question generation and validation with semantic deduplication and human verification, enabling scalable creation of benchmark data while preserving annotation quality. The benchmark is translated into 19 Indic languages, yielding a publicly released multilingual dataset of 69,420 question--answer pairs across 20 languages. We evaluate six LLMs under three protocols: LLM-as-a-judge and two deterministic lexical criteria, exact-substring and word-overlap matching. All three produce almost the same model ranking, showing that the results do not depend on the choice of judge. The frontier commercial model leads by a wide margin, and among open-weight models Gemma4 31B outperforms the Indic-specialised Sarvam 30B in every evaluated Indic language.
Rinit Jain, Tirthraj Mahajan, Advait Joshi et al.· 0 citations
