Henry Tang

Henry Tang, Ameet Deshpande, Karthik Narasimhan · princeton

Multilingual pre-trained models exhibit zero-shot cross-lingual transfer, where a model fine-tuned on a source language achieves surprisingly good performance on a target language. While studies have attempted to understand transfer, they focus only on MLM, and the large number of differences between natural languages makes it hard to disentangle the importance of different properties. In this work, we specifically highlight the importance of word embedding alignment by proposing a pre-training objective (ALIGN-MLM) whose auxiliary loss guides similar words in different languages to have similar word embeddings. ALIGN-MLM either outperforms or matches three widely adopted objectives (MLM, XLM, DICT-MLM) when we evaluate transfer between pairs of natural languages and their counterparts created by systematically modifying specific properties like the script. In particular, ALIGN-MLM outperforms XLM and MLM by 35 and 30 F1 points on POS-tagging for transfer between languages that differ both in their script and word order (left-to-right v.s. right-to-left). We also show a strong correlation between alignment and transfer for all objectives (e.g., rho=0.727 for XNLI), which together with ALIGN-MLM's strong performance calls for explicitly aligning word embeddings for multilingual models.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Henry Tang, Sarah Nadi

Code maintenance data sets typically consist of a before and after version of the code that contains the improvement or fix. Such data sets are important for software engineering support tools related to code maintenance, such as program repair, code recommender systems, or Application Programming Interface (API) misuse detection. Most of the current data sets are constructed from mining commit history in version-control systems or issues in issue-tracking systems. In this paper, we investigate whether Stack Overflow can be used as an additional data source. Comments on Stack Overflow provide an effective way for developers to point out problems with existing answers, alternative solutions, or pitfalls. In this paper, we mine comment-edit pairs from Stack Overflow and investigate their potential usefulness. These pairs have the added benefit of having concrete descriptions of why the change is needed as well as potentially having less tangled changes to deal with. We first design a technique to extract related comment-edit pairs and then investigate the nature of these pairs. We find that the majority of comment-edit pairs are not tangled, but only 27% of the studied pairs are potentially useful for the above applications. We categorize the types of mined pairs and find that the highest ratio of useful pairs come from categories Correction, Obsolete, Flaw, and Extension. To demonstrate the effectiveness of our extracted pairs, we submitted 15 pull requests on GitHub, 10 of which have been accepted to widely used repositories such as Apache Beam and nltk. Our work is the first to investigate Stack Overflow comment-edit pairs and opens the door for future work in this direction. Based on our findings and observations, we provide concrete suggestions on how to potentially identify a larger set of useful comment-edit pairs, which can also be facilitated by our shared data.