Evaluating Large Language Models for Generalization and Robustness via Data Compression

Existing methods for evaluating large language models face challenges such as data contamination, sensitivity to prompts, and the high cost of benchmark creation. To address this, we propose a lossless data compression based evaluation approach that tests how models' predictive abilities generalize after their training cutoff. Specifically, we collect comprehensive test data spanning 83 months from 2017 to 2023 and split the data into training and testing periods according to models' training data cutoff. We measure: 1) the compression performance on the testing period as a measure of generalization on unseen data; and 2) the performance gap between the training and testing period as a measure of robustness. Our experiments test 14 representative large language models with various sizes on sources including Wikipedia, news articles, code, arXiv papers, and multi-modal data. We find that the compression rate of many models reduces significantly after their cutoff date, but models such as Mistral and Llama-2 demonstrate a good balance between performance and robustness. Results also suggest that models struggle to generalize on news and code data, but work especially well on arXiv papers. We also find the context size and tokenization implementation have a big impact of on the overall compression performance.

Finding Challenging Metaphors that Confuse Pretrained Language Models

Metaphors are considered to pose challenges for a wide spectrum of NLP tasks. This gives rise to the area of computational metaphor processing. However, it remains unclear what types of metaphors challenge current state-of-the-art models. In this paper, we test various NLP models on the VUA metaphor dataset and quantify to what extent metaphors affect models' performance on various downstream tasks. Analysis reveals that VUA includes a large number of metaphors that pose little difficulty to downstream tasks. We would like to shift the attention of researchers away from these metaphors to instead focus on challenging metaphors. To identify hard metaphors, we propose an automatic pipeline that identifies metaphors that challenge a particular model. Our analysis demonstrates that our detected hard metaphors contrast significantly with VUA and reduce the accuracy of machine translation by 16\%, QA performance by 4\%, NLI by 7\%, and metaphor identification recall by over 14\% for various popular NLP systems.

SciMMIR: Benchmarking Scientific Multi-modal Information Retrieval

Multi-modal information retrieval (MMIR) is a rapidly evolving field, where significant progress, particularly in image-text pairing, has been made through advanced representation learning and cross-modality alignment research. However, current benchmarks for evaluating MMIR performance in image-text pairing within the scientific domain show a notable gap, where chart and table images described in scholarly language usually do not play a significant role. To bridge this gap, we develop a specialised scientific MMIR (SciMMIR) benchmark by leveraging open-access paper collections to extract data relevant to the scientific domain. This benchmark comprises 530K meticulously curated image-text pairs, extracted from figures and tables with detailed captions in scientific documents. We further annotate the image-text pairs with two-level subset-subcategory hierarchy annotations to facilitate a more comprehensive evaluation of the baselines. We conducted zero-shot and fine-tuning evaluations on prominent multi-modal image-captioning and visual language models, such as CLIP and BLIP. Our analysis offers critical insights for MMIR in the scientific domain, including the impact of pre-training and fine-tuning settings and the influence of the visual and textual encoders. All our data and checkpoints are publicly available at https://github.com/Wusiwei0410/SciMMIR.

CMMMU: A Chinese Massive Multi-discipline Multimodal Understanding Benchmark

As the capabilities of large multimodal models (LMMs) continue to advance, evaluating the performance of LMMs emerges as an increasing need. Additionally, there is an even larger gap in evaluating the advanced knowledge and reasoning abilities of LMMs in non-English contexts such as Chinese. We introduce CMMMU, a new Chinese Massive Multi-discipline Multimodal Understanding benchmark designed to evaluate LMMs on tasks demanding college-level subject knowledge and deliberate reasoning in a Chinese context. CMMMU is inspired by and strictly follows the annotation and analysis pattern of MMMU. CMMMU includes 12k manually collected multimodal questions from college exams, quizzes, and textbooks, covering six core disciplines: Art & Design, Business, Science, Health & Medicine, Humanities & Social Science, and Tech & Engineering, like its companion, MMMU. These questions span 30 subjects and comprise 39 highly heterogeneous image types, such as charts, diagrams, maps, tables, music sheets, and chemical structures. CMMMU focuses on complex perception and reasoning with domain-specific knowledge in the Chinese context. We evaluate 11 open-source LLMs and one proprietary GPT-4V(ision). Even GPT-4V only achieves accuracies of 42%, indicating a large space for improvement. CMMMU will boost the community to build the next-generation LMMs towards expert artificial intelligence and promote the democratization of LMMs by providing diverse language contexts.

Kun: Answer Polishment for Chinese Self-Alignment with Instruction Back-Translation

In this paper, we introduce Kun, a novel approach for creating high-quality instruction-tuning datasets for large language models (LLMs) without relying on manual annotations. Adapting a self-training algorithm based on instruction back-translation and answer polishment, Kun leverages unlabelled data from diverse sources such as Wudao, Wanjuan, and SkyPile to generate a substantial dataset of over a million Chinese instructional data points. This approach significantly deviates from traditional methods by using a self-curation process to refine and select the most effective instruction-output pairs. Our experiments with the 6B-parameter Yi model across various benchmarks demonstrate Kun's robustness and scalability. Our method's core contributions lie in its algorithmic advancement, which enhances data retention and clarity, and its innovative data generation approach that substantially reduces the reliance on costly and time-consuming manual annotations. This methodology presents a scalable and efficient solution for improving the instruction-following capabilities of LLMs, with significant implications for their application across diverse fields. The code and dataset can be found at https://github.com/Zheng0428/COIG-Kun

Language Model as an Annotator: Unsupervised Context-aware Quality Phrase Generation

Phrase mining is a fundamental text mining task that aims to identify quality phrases from context. Nevertheless, the scarcity of extensive gold labels datasets, demanding substantial annotation efforts from experts, renders this task exceptionally challenging. Furthermore, the emerging, infrequent, and domain-specific nature of quality phrases presents further challenges in dealing with this task. In this paper, we propose LMPhrase, a novel unsupervised context-aware quality phrase mining framework built upon large pre-trained language models (LMs). Specifically, we first mine quality phrases as silver labels by employing a parameter-free probing technique called Perturbed Masking on the pre-trained language model BERT (coined as Annotator). In contrast to typical statistic-based or distantly-supervised methods, our silver labels, derived from large pre-trained language models, take into account rich contextual information contained in the LMs. As a result, they bring distinct advantages in preserving informativeness, concordance, and completeness of quality phrases. Secondly, training a discriminative span prediction model heavily relies on massive annotated data and is likely to face the risk of overfitting silver labels. Alternatively, we formalize phrase tagging task as the sequence generation problem by directly fine-tuning on the Sequence-to-Sequence pre-trained language model BART with silver labels (coined as Generator). Finally, we merge the quality phrases from both the Annotator and Generator as the final predictions, considering their complementary nature and distinct characteristics. Extensive experiments show that our LMPhrase consistently outperforms all the existing competitors across two different granularity phrase mining tasks, where each task is tested on two different domain datasets.

LatestEval: Addressing Data Contamination in Language Model Evaluation through Dynamic and Time-Sensitive Test Construction

Data contamination in evaluation is getting increasingly prevalent with the emergence of language models pre-trained on super large, automatically crawled corpora. This problem leads to significant challenges in the accurate assessment of model capabilities and generalisations. In this paper, we propose LatestEval, an automatic method that leverages the most recent texts to create uncontaminated reading comprehension evaluations. LatestEval avoids data contamination by only using texts published within a recent time window, ensuring no overlap with the training corpora of pre-trained language models. We develop the LatestEval automated pipeline to 1) gather the latest texts; 2) identify key information, and 3) construct questions targeting the information while removing the existing answers from the context. This encourages models to infer the answers themselves based on the remaining context, rather than just copy-paste. Our experiments demonstrate that language models exhibit negligible memorisation behaviours on LatestEval as opposed to previous benchmarks, suggesting a significantly reduced risk of data contamination and leading to a more robust evaluation. Data and code are publicly available at: https://github.com/liyucheng09/LatestEval.

Avoiding Data Contamination in Language Model Evaluation: Dynamic Test Construction with Latest Materials

Data contamination in evaluation is getting increasingly prevalent with the emerge of language models pre-trained on super large, automatically-crawled corpora. This problem leads to significant challenges in accurate assessment of model capabilities and generalisations. In this paper, we propose LatestEval, an automatic method leverages the most recent texts to create uncontaminated reading comprehension evaluations. LatestEval avoids data contamination by only using texts published within a recent time window, ensuring no overlap with the training corpora of pre-trained language models. We develop LatestEval automated pipeline to 1) gather latest texts; 2) identify key information, and 3) construct questions targeting the information while removing the existing answers from the context. This encourages models to infer the answers themselves based on the remaining context, rather than just copy-paste. Our experiments demonstrate that language models exhibit negligible memorisation behaviours on LatestEval as opposed to previous benchmarks, suggesting a significantly reduced risk of data contamination and leading to a more robust evaluation. Data and code are publicly available at: https://github.com/liyucheng09/LatestEval.

How to Determine the Most Powerful Pre-trained Language Model without Brute Force Fine-tuning? An Empirical Survey

Transferability estimation has been attached to great attention in the computer vision fields. Researchers try to estimate with low computational cost the performance of a model when transferred from a source task to a given target task. Considering the effectiveness of such estimations, the communities of natural language processing also began to study similar problems for the selection of pre-trained language models. However, there is a lack of a comprehensive comparison between these estimation methods yet. Also, the differences between vision and language scenarios make it doubtful whether previous conclusions can be established across fields. In this paper, we first conduct a thorough survey of existing transferability estimation methods being able to find the most suitable model, then we conduct a detailed empirical study for the surveyed methods based on the GLUE benchmark. From qualitative and quantitative analyses, we demonstrate the strengths and weaknesses of existing methods and show that H-Score generally performs well with superiorities in effectiveness and efficiency. We also outline the difficulties of consideration of training details, applicability to text generation, and consistency to certain metrics which shed light on future directions.

A Cross-Attention Augmented Model for Event-Triggered Context-Aware Story Generation

Despite recent advancements, existing story generation systems continue to encounter difficulties in effectively incorporating contextual and event features, which greatly influence the quality of generated narratives. To tackle these challenges, we introduce a novel neural generation model, EtriCA, that enhances the relevance and coherence of generated stories by employing a cross-attention mechanism to map context features onto event sequences through residual mapping. This feature capturing mechanism enables our model to exploit logical relationships between events more effectively during the story generation process. To further enhance our proposed model, we employ a post-training framework for knowledge enhancement (KeEtriCA) on a large-scale book corpus. This allows EtriCA to adapt to a wider range of data samples. This results in approximately 5\% improvement in automatic metrics and over 10\% improvement in human evaluation. We conduct extensive experiments, including comparisons with state-of-the-art (SOTA) baseline models, to evaluate the performance of our framework on story generation. The experimental results, encompassing both automated metrics and human assessments, demonstrate the superiority of our model over existing state-of-the-art baselines. These results underscore the effectiveness of our model in leveraging context and event features to improve the quality of generated narratives.