Lissard: Long and Simple Sequential Reasoning Datasets
This addresses a problem for AI researchers and developers by highlighting a key limitation in current language models for tasks involving simple, repetitive rules, though it is incremental as it builds on existing benchmarking efforts.
The paper introduces Lissard, a benchmark with seven tasks to assess language models' ability to handle sequences requiring repetitive procedural execution, showing that state-of-the-art models like GPT-4 and Mistral-7B consistently decline in performance as sequence complexity increases, e.g., failing on lists with 80 items after handling 20 items.
Language models are now capable of solving tasks that require dealing with long sequences consisting of hundreds of thousands of tokens. However, they often fail on tasks that require repetitive use of simple rules, even on sequences that are much shorter than those seen during training. For example, state-of-the-art LLMs can find common items in two lists with up to 20 items but fail when lists have 80 items. In this paper, we introduce Lissard, a benchmark comprising seven tasks whose goal is to assess the ability of models to process and generate wide-range sequence lengths, requiring repetitive procedural execution. Our evaluation of open-source (Mistral-7B and Mixtral-8x7B) and proprietary models (GPT-3.5 and GPT-4) show a consistent decline in performance across all models as the complexity of the sequence increases. The datasets and code are available at https://github.com/unicamp-dl/Lissard