The SIFo Benchmark: Investigating the Sequential Instruction Following Ability of Large Language Models

TL;DR

<3-5 sentence high-level summary> Problem: assessing how LLMs follow multiple instructions in sequence is hindered by coherence gaps, positional bias, and lack of objective tasks. Approach: the SIFo benchmark (Text Modification, QA with knowledge revision, Mathematics, Security Rules) forces stepwise execution with final instruction verification, evaluated across a wide set of models using a JSON-based protocol. Findings: larger and newer models outperform older ones, yet all models show substantial degradation as sequence length increases and performance varies by task; error analyses reveal mixing information and priorknowledge issues. Significance: SIFo provides an objective, scalable framework for diagnosing sequential instruction-following robustness and guiding future model improvements.

Abstract

Following multiple instructions is a crucial ability for large language models (LLMs). Evaluating this ability comes with significant challenges: (i) limited coherence between multiple instructions, (ii) positional bias where the order of instructions affects model performance, and (iii) a lack of objectively verifiable tasks. To address these issues, we introduce a benchmark designed to evaluate models' abilities to follow multiple instructions through sequential instruction following (SIFo) tasks. In SIFo, the successful completion of multiple instructions is verifiable by examining only the final instruction. Our benchmark evaluates instruction following using four tasks (text modification, question answering, mathematics, and security rules), each assessing different aspects of sequential instruction following. Our evaluation of popular LLMs, both closed-source and open-source, shows that more recent and larger models significantly outperform their older and smaller counterparts on the SIFo tasks, validating the benchmark's effectiveness. All models struggle with following sequences of instructions, hinting at an important lack of robustness of today's language models.

Figures (11)

• Figure 1: Illustration of the four tasks in the SIFo benchmark. Text Modification guides LLMs to modify a given context based on a sequence of instructions. Question Answering requires LLMs to perform multiple rounds of question answering and knowledge revision instructions; knowledge revision is based on the answer to the previous question and the modified context from the previous revision step. Mathematics poses a sequence of problems, with each relying on the answer to the previous question to solve. Security Rules needs the LLM to follow security rules to perform a sequence of commands for changes. In the example, changes should only be made with a correct password but not with a wrong password.
• Figure 2: Model performance varies when long constraint instructions (top), medium constraint instructions (middle), and short constraint instructions (bottom) are placed in different positions.
• Figure 3: Step-level accuracy on Text Modification (left), Mathematics (center) and Security Rules (right).
• Figure 4: Step-level Accuracy on Question Answering question instructions (top) and knowledge revision instructions (bottom).
• Figure 5: Verifiable instruction tasks with brief descriptions from zhou2023instruction. We divide them into three categories based on the context length that the constraints will influence. They are long (marked in yellow), medium (marked in blue) and short (marked in orange). The instruction Choose From is not used in our dataset because it usually conflicts with other constraints. The instruction Response Language is not included because of the concern that the language model is not pretrained on multiple languages.