Hello LMSys Team,
First, thank you for providing the Chatbot Arena as an invaluable public research tool.
While conducting long-form narrative stress tests on your platform, I was able to capture a pristine example of a catastrophic repetitive loop failure in the Gemini 1.5 Pro model. The failure was so severe that it generated a 53-page document of nonsensical, looping text.
I believe this log represents a valuable, real-world data point on model degradation under sustained creative load and may be useful for your comparative model analysis.
The full, unabridged chat log is available here: [Link to Google Drive]
As a result of this and other findings, I developed an open-source framework to mitigate these failures, which can be found here: [Link to GitHub repo]
I hope this data is a useful contribution to your excellent research.
Best regards,
“The Dungeon Master Protocol Project"
@dmprotocol.ai - X