Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SysBench: Can Large Language Models Follow System Messages?

Yanzhao Qin, Tao Zhang, Tao Zhang, Yanjun Shen, Wenjing Luo, Haoze Sun, Yan Zhang, Yujing Qiao, Weipeng Chen, Zenan Zhou, Wentao Zhang, Bin Cui

TL;DR

This work introduces SysBench, a benchmark that systematically analyzes system message following ability in terms of three limitations of existing LLMs: constraint violation, instruction misjudgement and multi-turn instability.

Abstract

Large Language Models (LLMs) have become instrumental across various applications, with the customization of these models to specific scenarios becoming increasingly critical. System message, a fundamental component of LLMs, is consist of carefully crafted instructions that guide the behavior of model to meet intended goals. Despite the recognized potential of system messages to optimize AI-driven solutions, there is a notable absence of a comprehensive benchmark for evaluating how well LLMs follow system messages. To fill this gap, we introduce SysBench, a benchmark that systematically analyzes system message following ability in terms of three limitations of existing LLMs: constraint violation, instruction misjudgement and multi-turn instability. Specifically, we manually construct evaluation dataset based on six prevalent types of constraints, including 500 tailor-designed system messages and multi-turn user conversations covering various interaction relationships. Additionally, we develop a comprehensive evaluation protocol to measure model performance. Finally, we conduct extensive evaluation across various existing LLMs, measuring their ability to follow specified constraints given in system messages. The results highlight both the strengths and weaknesses of existing models, offering key insights and directions for future research. The open source library SysBench is available at https://github.com/PKU-Baichuan-MLSystemLab/SysBench.

SysBench: Can Large Language Models Follow System Messages?

TL;DR

This work introduces SysBench, a benchmark that systematically analyzes system message following ability in terms of three limitations of existing LLMs: constraint violation, instruction misjudgement and multi-turn instability.

Abstract

Large Language Models (LLMs) have become instrumental across various applications, with the customization of these models to specific scenarios becoming increasingly critical. System message, a fundamental component of LLMs, is consist of carefully crafted instructions that guide the behavior of model to meet intended goals. Despite the recognized potential of system messages to optimize AI-driven solutions, there is a notable absence of a comprehensive benchmark for evaluating how well LLMs follow system messages. To fill this gap, we introduce SysBench, a benchmark that systematically analyzes system message following ability in terms of three limitations of existing LLMs: constraint violation, instruction misjudgement and multi-turn instability. Specifically, we manually construct evaluation dataset based on six prevalent types of constraints, including 500 tailor-designed system messages and multi-turn user conversations covering various interaction relationships. Additionally, we develop a comprehensive evaluation protocol to measure model performance. Finally, we conduct extensive evaluation across various existing LLMs, measuring their ability to follow specified constraints given in system messages. The results highlight both the strengths and weaknesses of existing models, offering key insights and directions for future research. The open source library SysBench is available at https://github.com/PKU-Baichuan-MLSystemLab/SysBench.
Paper Structure (28 sections, 3 equations, 6 figures, 6 tables)

This paper contains 28 sections, 3 equations, 6 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. System Messages in LLM
  4. Evaluation of Instruction Following
  5. SysBench
  6. Benchmark Design
  7. System Constraint
  8. User Instruction
  9. Multi-turn Conversation
  10. Dataset
  11. Data Construction
  12. Data Statistics
  13. Evaluation Protocol
  14. Experiments
  15. Experimental Setup
  16. ...and 13 more sections

Figures (6)

  • Figure 1: A sample system message, and limitations of LLMs on system message following.
  • Figure 2: Workflow of SysBench. Both system message and corresponding user instructions are fed into LLM to generate outputs; then a model-based verifier is applied to each response for evaluation. All texts are simplified for clearer presentation.
  • Figure 3: Distribution of domains and constraints.
  • Figure 4: The CSR under different types of constraints. Only 8 representative ones are shown.
  • Figure 5: The ISR gain when using ground-truth history; the n-th turn is denoted as $T_n$.
  • ...and 1 more figures