The paper notes the opaque view of closed LLMs like ChatGPT, and says improving monitoring and transparency can avoid pitfalls of LLM drift.
Researchers at Stanford University and University of California-Berkeley have published an unreviewed paper on the open access journal arXiv.org, which found that the “performance and behavior” of OpenAI’s ChatGPT large language models (LLMs) have changed between March and June 2023. The researchers concluded that their tests revealed “performance on some tasks have gotten substantially worse over time.”
“The whole motivation for this research: We’ve seen a lot of anecdotal experiences from users of ChatGPT that the models’ behavior is changing over time,” James Zou, a Stanford professor and one of the three authors of the research paper, told VentureBeat. “Some tasks may be getting better or other tasks getting worse. This is why we wanted to do this more systematically to evaluate it across different time points.”Qualifying information
There are some important caveats to the findings and the paper, including that arXiv.org accepts nearly all user-generated papers that comply with its guidelines, and that this particular paper — like many on the site — has not yet been peer-reviewed, nor published in another reputable scientific journal. However, Zou told VentureBeat that the authors do plan to submit it for consideration and review by a journal.
In a tweet in response to the paper and the ensuing discussions, Logan Kilpatrick, OpenAI developer advocate, offered a general thanks to those reporting their experiences with the LLM platform and said they’re actively looking into the issues being shared. Kilpatrick also posted a link to OpenAI’s Evals framework GitHub page which is used to evaluate LLMs and LLM systems with an open-source registry of benchmarks.
