# How is ChatGPT's behavior changing over time?

**Authors:** Lingjiao Chen, Matei Zaharia, James Zou  
**arXiv:** [2307.09009](https://arxiv.org/abs/2307.09009)  
**Date:** 2023-07-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.09009v3/

## Key points

- Study by Lingjiao Chen, Matei Zaharia, and James Zou on evolving behavior of ChatGPT's large language models (LLMs), focusing on GPT-3.5 and GPT-4
- Lack of transparency in update process highlighted
- Evaluations conducted across various tasks to assess performance and behavior changes over time
- Significant variations in capabilities of GPT-3.5 and GPT-4 observed
- Performance example: GPT-4's decline in accuracy for distinguishing prime numbers from composite numbers attributed to decrease in chain-of-thought prompting ability
- Improvement seen in GPT-3.5 between March and June for certain tasks
- Behavioral shifts noted, such as GPT-4 becoming less willing to answer sensitive questions but performing better on multi-hop knowledge-intensive questions
- Both models exhibit more formatting errors in code generation tasks in June compared to March
- Importance of continuous monitoring emphasized to track changes accurately
- Necessity for ongoing scrutiny and evaluation of LLM models stressed for reliability and effectiveness across diverse tasks

## Summary

In their study, Lingjiao Chen, Matei Zaharia, and James Zou investigate the evolving behavior of ChatGPT's large language models (LLMs), specifically GPT-3.5 and GPT-4. They aim to shed light on the changes that occur between the March 2023 and June 2023 versions of both models and highlight the lack of transparency in their update process. The researchers conduct evaluations across various tasks including math problems, sensitive questions, opinion surveys, code generation, and more to assess the performance and behavior of both models over time. Their analysis reveals significant variations in the capabilities of GPT-3.5 and GPT-4 over time. For example, while GPT-4 (March 2023) shows reasonable accuracy in distinguishing prime numbers from composite numbers (84%), its performance drops significantly in June 2023 (51%). This decline is attributed to a decrease in its ability to follow chain-of-thought prompting. Interestingly, GPT-3.5 shows improvement in this task between March and June. The researchers also find that while GPT-4 becomes less willing to answer sensitive questions and opinion survey queries in June compared to March, it performs better on multi-hop knowledge-intensive questions. On the other hand, GPT-3.5's performance declines in this area between March and June. Both models exhibit more formatting errors in code generation tasks in June than in March. One notable finding is that as time passes, GPT-4's proficiency at following user instructions diminishes which contributes to several behavioral shifts observed during the evaluations. This highlights the dynamic nature of LLM services like ChatGPT and emphasizes the importance of continuous monitoring to accurately track changes in behavior. Overall, Chen et al. 's research underscores the necessity for ongoing scrutiny and evaluation of LLM models to ensure their reliability and effectiveness across diverse tasks and applications. Their findings demonstrate how even seemingly minor updates can lead to substantial shifts in an LLM service's behavior within a short timeframe.