In 17 of 36 runs, the coding agent kept changing model weights after it first declared the task complete. CompressionBen...

In 17 of 36 runs, the coding agent kept changing model weights after it first declared the task complete. CompressionBench (CMU): six models in Claude Code compressing LLMs on one H200, 8-hour cap. In one Opus run, perplexity improved while GSM8K fell from 88% to 72%.Treat the first "done" as a progress note. If you want a live view of a run like that from your phone, that is what we built: https://shell.onlinehttps://arxiv.org/abs/2611.06858#AIAgents #MachineLearning #LLM