Recent advancements in large language models (LLMs) showcase varied multilingual capabilities across tasks like translation, code generation, and reasoning. Previous assessments often limited their scope to fundamental natural language processing (NLP) or isolated capability-specific tasks. To alleviate this drawback, we aim to present a comprehensive multilingual multitask benchmark. First, we introduce P-MMEval, a large-scale benchmark covering fundamental and capability-specialized datasets. Furthermore, P-MMEval delivers consistent language coverage across various datasets and provides parallel samples. Finally, we conduct extensive experiments on representative multilingual model series to compare performances across models and tasks, explore the relationship between multilingual performances and factors such as tasks, model sizes, languages, and prompts, and examine the effectiveness of knowledge transfer from English to other languages. The resulting insights are intended to offer valuable guidance for future research.

The mechanisms behind multilingual capabilities in Large Language Models (LLMs) have been examined using neuron-based or internal-activation-based methods. However, these methods often face challenges such as superposition and layer-wise activation variance, which limit their reliability. Sparse Autoencoders (SAEs) offer a more nuanced analysis by decomposing the activations of LLMs into a sparse linear combination of SAE features. We introduce a novel metric to assess the monolinguality of features obtained from SAEs, discovering that some features are strongly related to specific languages. Additionally, we show that ablating these SAE features only significantly reduces abilities in one language of LLMs, leaving others almost unaffected. Interestingly, we find some languages have multiple synergistic SAE features, and ablating them together yields greater improvement than ablating individually. Moreover, we leverage these SAE-derived language-specific features to enhance steering vectors, achieving control over the language generated by LLMs. The code is publicly available at https://github.com/Aatrox103/multilingual-llm-features.

Despite their promising performance across various tasks, recent studies reveal that Large language models (LLMs) still exhibit significant deficiencies in handling several word-level and character-level tasks, e.g., word unscrambling and sentence editing, indicating urgent needs for substantial improvements in basic language understanding and manipulation. To address these challenges, it is crucial to develop large-scale benchmarks that can comprehensively assess the performance of LLMs in basic language tasks. In this paper, we introduce a bilingual benchmark, CWUM, to investigate the capabilities and limitations of LLMs in understanding and manipulating natural language at both character and word levels. CWUM consists of 15 simple text editing tasks, e.g., letter counting, word reversing, Chinese character inserting, etc. We conduct extensive experiments on eight advanced LLMs, including base models and instruction-tuned (chat) variants. The experimental results highlight significant failures of existing LLMs on CWUM tasks that humans can solve perfectly with 100% accuracy. On English tasks of CWUM, the average accuracy of GPT-4, LLaMA-3-70B, and Qwen-72B is 66.64%, 39.32%, and 33.16%, respectively, which lags far behind human performance. Instruction-tuning the base model does not lead to a distinct performance improvement, as the average accuracy of LLaMA-3-70B-Instruct on English tasks is only 1.44% higher than that of the base LLaMA-3-70B. Ultimately, we show that supervised fine-tuning (SFT) can enhance model performance on CWUM without compromising its ability to generalize across general tasks.

Recently, large language models (LLMs) have demonstrated breakthrough mathematical problem-solving capabilities in grade school math word problems (MWP). For example, on the MWP benchmark GSM8K, the accuracy of GPT-3.5-Turbo and MetaMath-70B reaches 80.80% and 82.30%, respectively. One question arises, does it mean that LLMs have truly mastered related mathematical problem-solving abilities? In this paper, by presenting two types of benchmarks, where MCGSM8K aims at selecting one correct solution from four solutions, while GSM8K-Judgement judges whether a solution to a given question is true or false, we demonstrate that the ability of most LLMs to evaluate the mathematical reasoning process of MWP is far from sufficient. To compensate for this issue, we propose hybrid supervised fine-tuning data from the training data of GSM8K, MCGSM8K, and GSM8K-Judgement, which significantly improves performance on the proposed reasoning process evaluation benchmarks. For example, fine-tuning improves the performance of LLaMA-2-13B from 33.51% to 70.89% on MCGSM8K. In conclusion, we experimentally demonstrate that most LLMs have limited ability to evaluate the mathematical reasoning process of MWP, which can be enhanced through fine-tuning.