As LLMs advance, popular commonsense and NLU benchmarks have become saturated, yet multilingual evaluation still lags. Existing benchmarks cover only a small set of languages, and translation based multilingual benchmarks can introduce bias. We introduce HellaSwagUltra, a multilingual commonsense and NLU benchmark with substantial local cultural knowledge. It is built with an automated, continuously extensible pipeline. Instead of explicitly prompting for reasoning, each example implicitly embeds two commonsense or local knowledge facts in the context, and answer options contain subtle cues about whether those facts are violated, requiring models to pick the most plausible continuation. All test items are fully reviewed and corrected by language experts, with ongoing updates. Even the strongest proprietary model (e.g., Gemini-2.5-Pro) reaches only 62.5% accuracy, while GPT-4o and leading open-source models hover around 40%–50%, demonstrating that multilingual commonsense reasoning remains an open challenge.

Multilingual large language models (LLMs) are advancing rapidly, with new models frequently claiming support for an increasing number of languages. However, existing evaluation datasets are limited and lack cross-lingual alignment, leaving assessments of multilingual capabilities fragmented in both language and skill coverage. To address this, we introduce MuBench, a benchmark covering 61 languages with 3.9M samples and evaluating a broad range of capabilities. We evaluate several state-of-the-art multilingual LLMs and find notable gaps between claimed and actual language coverage, particularly a persistent performance disparity between English and low-resource languages. Leveraging MuBench’s alignment, we propose Multilingual Consistency (MLC) as a complementary metric to accuracy for analyzing performance bottlenecks and guiding model improvement. MuBench provides flexible evaluation formats, including mixed-language testing. Experimental results show that increasing model size does not improve its ability to handle mixed-language contexts. We recruited human experts to evaluate translation quality and cultural sensitivity for 34k samples across 17 languages, and combined these assessments with an LLM-as-a-Judge approach to ensure overall data quality in low resource languages.

Multilingual LLMs increasingly claim broad language coverage, yet their performance still lags far behind English in many languages. Because low resource data is scarce and noisy, improving these languages directly is difficult, making cross lingual transfer from English a natural alternative. Parallel corpora are central to such transfer, but it remains unclear which kinds of parallel data matter most. We propose ParaRater, a corpus selection method that identifies high value English text to translate into target languages, building parallel corpora that most efficiently improve target language performance. ParaRater uses meta learning to tie selection directly to downstream accuracy on native target language data, and applies a two stage filter to find samples that only help when both language versions are used in training, capturing truly impactful parallel data. Experiments across Chinese, German, French, Japanese, Thai, Arabic, Tagalog, and Russian, together with qualitative analyses, support these findings.

The data mixture used in the pre-training of a language model is a cornerstone of its final performance. Static data mixing strategies in Large Language Model (LLM) pre-training are often suboptimal as they fail to adapt to the model’s evolving learning states. Conversely, fully online dynamic updates, while adaptive, incur prohibitive computational costs. To bridge this gap, we propose TiKMiX, an efficient semi-dynamic data mixing framework. Our approach is grounded in a key observation of influence ranking invariance: the relative importance of data domains exhibits strong temporal stability over long training intervals. Leveraging this insight, we propose Group Influence, an efficient approach for quantifying domain impact, and formulate data mixing as a periodic, low-overhead influence maximization problem. Compared with REGMIX, the proposed method reduces computational overhead by 80% and achieves an average performance gain of 2% across nine downstream benchmarks, thereby effectively mitigating data under-digestion.

Mathematical reasoning aims to infer satisfiable solutions based on the given mathematics questions. Previous natural language processing researches have proven the effectiveness of sequence-to-sequence (Seq2Seq) or related variants on mathematics solving. However, few works have been able to explore structural or syntactic information hidden in expressions (e.g., precedence and associativity). This dissertation set out to investigate the usefulness of such untapped information for neural architectures. Firstly, mathematical questions are represented in the format of graphs within syntax analysis. The structured nature of graphs allows them to represent relations of variables or operators while preserving the semantics of the expressions. Having transformed to the new representations, we proposed a graph-to-sequence neural network GraphMR, which can effectively learn the hierarchical information of graphs inputs to solve mathematics and speculate answers. A complete experimental scenario with four classes of mathematical tasks and three Seq2Seq baselines is built to conduct a comprehensive analysis, and results show that GraphMR outperforms others in hidden information learning and mathematics resolving.

Mixed Boolean-Arithmetic (MBA) expressions involve both arithmetic calculation (e.g.,plus, minus, multiply) and bitwise computation (e.g., and, or, negate, xor). MBA expressions have been widely applied in software obfuscation, transforming programs from a simple form to a complex form. MBA expressions are challenging to be simplified, because the interleaving bitwise and arithmetic operations causing mathematical reduction laws to be ineffective. Our goal is to recover the original, simple form from an obfuscated MBA expression. In this paper, we first propose NeuReduce, a string to string method based on neural networks to automatically learn and reduce complex MBA expressions. We develop a comprehensive MBA dataset, including one million diversified MBA expression samples and corresponding simplified forms. After training on the dataset, NeuReduce can reduce MBA rules to homelier but mathematically equivalent forms. By comparing with three state-of-the-art MBA reduction methods, our evaluation result shows that NeuReduce outperforms all other tools in terms of accuracy, solving time, and performance overhead.