Large language models (LLMs) have demonstrated significant utility in real-world applications, exhibiting impressive capabilities in natural language processing and understanding. Benchmark evaluations are crucial for assessing the capabilities of LLMs as they can provide a comprehensive assessment of their strengths and weaknesses. However, current evaluation methods often overlook the inherent randomness of LLMs by employing deterministic generation strategies or relying on a single random sample, resulting in unaccounted sampling variance and unreliable benchmark score estimates. In this paper, we propose a hierarchical statistical model that provides a more comprehensive representation of the benchmarking process by incorporating both benchmark characteristics and LLM randomness. We show that leveraging multiple generations improves the accuracy of estimating the benchmark score and reduces variance. Multiple generations also allow us to define ℙ (correct), a prompt-level difficulty score based on correct ratios, providing fine-grained insights into individual prompts. Additionally, we create a data map that visualizes difficulty and semantics of prompts, enabling error detection and quality control in benchmark construction.

Large language models (LLMs) have shown remarkable performance in various tasks but often fail to handle queries that exceed their knowledge and capabilities, leading to incorrect or fabricated responses. This paper addresses the need for LLMs to recognize and refuse infeasible tasks due to the requests surpassing their capabilities. We conceptualize four main categories of infeasible tasks for LLMs, which cover a broad spectrum of hallucination-related challenges identified in prior literature. We develop and benchmark a new dataset comprising diverse infeasible and feasible tasks to evaluate multiple LLMs’ abilities to decline infeasible tasks. Furthermore, we explore the potential of increasing LLMs’ refusal capabilities with fine-tuning. Experiments validate the effectiveness of our trained models, offering promising directions for refining the operational boundaries of LLMs in real applications.

This paper illustrates the submission system of the IOL Research team for the WMT24 General Machine Translation shared task. We submitted translations for all translation directions in the general machine translation task. According to the official track categorization, our system qualifies as an open system due to the utilization of open-source resources in developing our machine translation model. With the growing prevalence of large language models (LLMs) as a conventional approach for managing diverse NLP tasks, we have developed our machine translation system by leveraging the capabilities of LLMs. Overall, we first performed continued pretraining using the open-source LLMs with tens of billions of parameters to enhance the model’s multilingual capabilities. Subsequently, we employed open-source Large Language Models, equipped with hundreds of billions of parameters, to generate synthetic data. This data was then blended with a modest quantity of additional open-source data for precise supervised fine-tuning. In the final stage, we also used ensemble learning to improve translation quality. Based on the official automated evaluation metrics, our system excelled by securing the top position in 8 out of the total 11 translation directions, spanning both open and constrained system categories.

This paper describes the IOL Research team’s submission systems for the WMT23 low-resource Indic language translation shared task. We participated in 4 language pairs, including en-as, en-mz, en-kha, en-mn. We use transformer based neural network architecture to train our machine translation models. Overall, the core of our system is to improve the quality of low resource translation by utilizing monolingual data through pre-training and data augmentation. We first trained two denoising language models similar to T5 and BART using monolingual data, and then used parallel data to fine-tune the pretrained language models to obtain two multilingual machine translation models. The multilingual machine translation models can be used to translate English monolingual data into other multilingual data, forming multilingual parallel data as augmented data. We trained multiple translation models from scratch using augmented data and real parallel data to build the final submission systems by model ensemble. Experimental results show that our method greatly improves the BLEU scores for translation of these four language pairs.

This paper describes the IOL Research team’s submission systems for the WMT23 general machine translation shared task. We participated in two language translation directions, including English-to-Chinese and Chinese-to-English. Our final primary submissions belong to constrained systems, which means for both translation directions we only use officially provided monolingual and bilingual data to train the translation systems. Our systems are based on Transformer architecture with pre-norm or deep-norm, which has been proven to be helpful for training deeper models. We employ methods such as back-translation, data diversification, domain fine-tuning and model ensemble to build our translation systems. An important aspect worth mentioning is our careful data cleaning process and the utilization of a substantial amount of monolingual data for data augmentation. Compared with the baseline system, our submissions have a large improvement in BLEU score.