ChatLog: Carefully Evaluating the Evolution of ChatGPT Across Time

Shangqing Tu, Chunyang Li, Jifan Yu, Xiaozhi Wang, Lei Hou, Juanzi Li

Abstract

ChatGPT has achieved great success and can be considered to have acquired an infrastructural status. There are abundant works for evaluating ChatGPT on benchmarks. However, existing benchmarks encounter two challenges: (1) Disregard for periodical evaluation and (2) Lack of fine-grained features. In this paper, we construct ChatLog, an ever-updating dataset with large-scale records of diverse long-form ChatGPT responses for 21 NLP benchmarks from March, 2023 to now. We conduct a comprehensive performance evaluation to find that most capabilities of ChatGPT improve over time except for some abilities, and there exists a step-wise evolving pattern of ChatGPT. We further analyze the inherent characteristics of ChatGPT by extracting the knowledge and linguistic features. We find some stable features that stay unchanged and apply them on the detection of ChatGPT-generated texts to improve the robustness of cross-version detection.

Introduction

Evaluating Large Language Models (LLMs) is often treated like a static exam: you give the model a test, record the score, and publish the results. However, researchers from Tsinghua University reveal that ChatGPT (specifically the gpt-3.5-turbo iteration) is a moving target. By tracking the model's performance daily and monthly, the team provides a rare glimpse into the "staircase" of AI evolution.

Most AI benchmarks are snapshots in time. However, OpenAI frequently updates ChatGPT to improve safety, efficiency, or accuracy. These "black-box" updates occur without public changelogs, creating two major issues: benchmark obsolescence and feature drift. Beyond just being "right" or "wrong," the style of the model—its vocabulary, sentence structure, and conciseness—shifts over time, which can break downstream applications or AI-detection tools.

Methods

The researchers employed a two-track monitoring system to address these challenges. ChatLog-Monthly tested ChatGPT on 21 diverse NLP benchmarks every month to observe broad performance trends. ChatLog-Daily sent the same 1,000 questions (sampled from the ELI5 dataset) to the model every single day, running each question three times to account for inherent randomness.

Conceptual framework of ChatLog's monitoring system, illustrating the monthly and daily observation tracks against the model's evolving 'Real Ability'.

Crucially, they didn't just look at accuracy. They used tools like LingFeat and UIE to extract 265 "fine-grained features," such as entity density and lexical diversity. They then used a z-score statistical test to distinguish between minor daily fluctuations and actual version updates from OpenAI.

Key Insights and Ideas

The study revealed several critical patterns in ChatGPT's evolution:

  • The Staircase Pattern: ChatGPT doesn't improve smoothly. It stays stable for weeks and then undergoes a "step change." Significant shifts were detected in late March and late June 2023.
  • The "Conciseness" Trade-off: The model is getting more "efficient." Over time, its answers have become shorter and less redundant. While this improves readability, it may be a result of OpenAI optimizing for lower computational costs (fewer tokens).
  • Regression is Real: Surprisingly, while ChatGPT improved at sentiment analysis (GoEmotions), it significantly worsened at reading comprehension (SQuAD). This suggests that "tuning" a model for one behavior (like following instructions) can inadvertently dull its reasoning edge.
  • Stable Features for Detection: Despite the updates, some linguistic "fingerprints" remain stable. By identifying these unchanging features, the authors built a more robust AI-text detector that doesn't lose accuracy when a new version of ChatGPT is released.

Key Results

The longitudinal data shows significant performance swings across different task categories. For instance, on the GoEmotions dataset, ChatGPT’s F1 score jumped from 24.90% in June to 28.88% in July. Conversely, on SQuAD, its score plummeted from 58.07% in March to 48.20% in April.

Performance metrics for ChatGPT across 21 NLP tasks from March to September 2023, highlighting the variability and 'step-wise' changes in different capabilities.

The researchers also confirmed that the model is producing more concise answers over time, as evidenced by increasing ROUGE-1 precision scores. Furthermore, by combining a standard RoBERTa detector with 10 "stable" linguistic features identified in the study, they improved cross-version detection accuracy by 2.1%, reaching a total accuracy of 98.1%.

Conclusion

The findings highlight that ChatGPT is not a static product but an evolving system. While many capabilities improve, the "black-box" nature of updates can lead to unexpected regressions in reasoning tasks. The authors acknowledge limitations, such as the focus on English-language benchmarks and the specific gpt-3.5-turbo model. The team plans to continue maintaining the ChatLog project, providing a living record of how the world's most famous AI grows—and sometimes stumbles—over time.

Summary updated

Abstract

ChatGPT has achieved great success and can be considered to have acquired an infrastructural status. There are abundant works for evaluating ChatGPT on benchmarks. However, existing benchmarks encounter two challenges: (1) Disregard for periodical evaluation and (2) Lack of fine-grained features. In this paper, we construct ChatLog, an ever-updating dataset with large-scale records of diverse long-form ChatGPT responses for 21 NLP benchmarks from March, 2023 to now. We conduct a comprehensive performance evaluation to find that most capabilities of ChatGPT improve over time except for some abilities, and there exists a step-wise evolving pattern of ChatGPT. We further analyze the inherent characteristics of ChatGPT by extracting the knowledge and linguistic features. We find some stable features that stay unchanged and apply them on the detection of ChatGPT-generated texts to improve the robustness of cross-version detection. We will continuously maintain our project at \url{https://github.com/THU-KEG/ChatLog/}.
BibTeX
@misc{tu2023chatlog,
  author = {Shangqing Tu and Chunyang Li and Jifan Yu and Xiaozhi Wang and Lei Hou and Juanzi Li},
  title = {{ChatLog: Carefully Evaluating the Evolution of ChatGPT Across Time}},
  year = {2023},
  eprint = {2304.14106},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2304.14106},
}
Figure