The use of large language models (LLMs) is widespread across many domains, including Software Engineering, where they have been used to automate tasks such as program generation and test classification. As LLM-based methods continue to evolve, it is important that we define clear and robust methods that fairly evaluate performance. Benchmarks are a common approach to assess LLMs with respect to their ability to solve problem-specific tasks as well as assess different versions of an LLM to solve tasks over time. For example, the HumanEval benchmark is composed of 164 hand-crafted tasks and has become an important tool in assessing LLM-based program generation. However, a major barrier to a fair evaluation of LLMs using benchmarks like HumanEval is data contamination resulting from data leakage of benchmark tasks and solutions into the training data set. This barrier is compounded by the black-box nature of LLM training data which makes it difficult to even know if data leakage has occurred. To address the data leakage problem, we propose a new benchmark construction method where a benchmark is composed of template tasks that can be instantiated into new concrete tasks using combinatorial test design. Concrete tasks for the same template task must be different enough that data leakage has minimal impact and similar enough that the tasks are interchangeable with respect to performance evaluation. To assess our benchmark construction method, we propose HumanEval_T, an alternative benchmark to HumanEval that was constructed using template tasks and combinatorial test design.

While LLM-Based agents, which use external tools to solve complex problems, have made significant progress, benchmarking their ability is challenging, thereby hindering a clear understanding of their limitations. In this paper, we propose an interactive evaluation framework, named CIBench, to comprehensively assess LLMs' ability to utilize code interpreters for data science tasks. Our evaluation framework includes an evaluation dataset and two evaluation modes. The evaluation dataset is constructed using an LLM-human cooperative approach and simulates an authentic workflow by leveraging consecutive and interactive IPython sessions. The two evaluation modes assess LLMs' ability with and without human assistance. We conduct extensive experiments to analyze the ability of 24 LLMs on CIBench and provide valuable insights for future LLMs in code interpreter utilization.

We investigate the capabilities of transformer large language models (LLMs) on relational reasoning tasks involving abstract symbols. Such tasks have long been studied in the neuroscience literature as fundamental building blocks for more complex abilities in programming, mathematics, and verbal reasoning. For (i) regression tasks, we prove that transformers generalize when trained, but require astonishingly large quantities of training data. For (ii) next-token-prediction tasks with symbolic labels, we show an "inverse scaling law": transformers fail to generalize as their embedding dimension increases. For both settings (i) and (ii), we propose subtle transformer modifications which can reduce the amount of data needed by adding two trainable parameters per head.