Lissard: Long and Simple Sequential Reasoning Datasets

TL;DR

Addressing length generalization challenges in large language models, the paper demonstrates that longer sequences demanding repetitive rule application remain hard for even state-of-the-art LLMs. It introduces Lissard, a synthetic benchmark with two task families (input and generation extrapolation) and a key-entity control to systematically vary sequence length. Across GPT-4, GPT-3.5, Mistral, and Mixtral, performance declines as length/complexity increases, revealing a shared fracture point in length extrapolation. The work provides new datasets and scripts to enable future research on long-sequence reasoning and rule-based procedural tasks.

Abstract

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

Figures (2)

• Figure 1: (a) Performance of GPT-4 on the Lissard benchmark (see Table \ref{['tab:key_entity_bins']} for the definition of the bins); (b) Comparative performance of models in the "Object Counting" task.
• Figure 2: Results per task.