How Is ChatGPT’s Behavior Changing Over Time?

Analysis

Why This Paper Matters

Large language models (LLMs) such as GPT-3.5 and GPT-4 are widely deployed, yet their updates are opaque. This paper is among the first to systematically document that the behavior of the 'same' LLM service can change substantially over a short period. By evaluating the March 2023 and June 2023 versions of GPT-3.5 and GPT-4 across seven diverse tasks, the authors reveal significant performance and behavioral drift. This matters because users and developers often assume that a model's behavior is stable unless the version number changes. The findings challenge that assumption and highlight the need for continuous monitoring.

The paper's significance is amplified by its timing: as LLMs become integrated into critical applications, understanding how they evolve is essential for reliability, safety, and trust. The observed drop in GPT-4's instruction-following ability, particularly chain-of-thought prompting, suggests that updates can inadvertently degrade capabilities. This has implications for prompt engineering and model evaluation practices.

Technical Contributions

• A comprehensive evaluation framework comparing two versions of GPT-3.5 and GPT-4 on seven task categories: math problems, sensitive/dangerous questions, opinion surveys, multi-hop knowledge-intensive questions, code generation, US Medical License tests, and visual reasoning.
• Empirical evidence of performance drift: GPT-4's accuracy on prime vs. composite number questions fell from 84% to 51%, while GPT-3.5 improved on the same task.
• Identification of a common factor behind behavior drifts: GPT-4's decreased ability to follow user instructions, especially chain-of-thought prompting.
• Documentation of changes in model willingness to answer sensitive and opinion survey questions, and increased formatting mistakes in code generation.

Results

The results show substantial variation across tasks and versions. GPT-4's prime vs. composite accuracy dropped from 84% to 51%, partly due to reduced chain-of-thought adherence. GPT-3.5, conversely, improved on this task. GPT-4 became less willing to answer sensitive questions and opinion surveys in June than in March. On multi-hop questions, GPT-4 improved while GPT-3.5 degraded. Both models exhibited more formatting mistakes in code generation in June. These findings demonstrate that behavior drift is not uniform and can affect different capabilities in different directions.

Significance

This paper has broad implications for the AI field. It underscores the need for continuous monitoring and evaluation of deployed LLMs, as silent updates can alter behavior in ways that affect user experience and safety. The findings also suggest that model providers should be more transparent about updates and that developers should implement robust testing pipelines. Furthermore, the observed decrease in instruction-following ability highlights the fragility of prompt-based methods and the importance of evaluating models beyond static benchmarks. Overall, this work calls for a shift toward dynamic, ongoing assessment of LLMs in production.