{"slug": "gemma-the-unstopping-a-behavioral-experiment", "title": "Gemma The Unstopping: a Behavioral Experiment", "summary": "A behavioral experiment on Google's Gemma model found that giving it a stop_run tool did not change its task completion rate, and the model only called the tool when no steps remained, showing no contradiction between its chain-of-thought and actions.", "body_md": "Following up on my previous experiment - [studying Gemma's behavior on agentic tasks when given the number of steps left across the run](https://www.lesswrong.com/posts/9Gj9anZi95RgHQPv6/when-gemma-thinks-about-resources-it-fails-a-behavioral) - I endeavored to see whether giving Gemma a *stop_run *tool meaningfully changes it's *behavior*.\n\nMy initial assumption was that it would not change the completion rate of these tasks, and that has been correct.\n\nWhat I wanted to test:\n\nAcross 400 runs, while taking into account the 500 runs in the previous experiment, I measured:\n\nIn none of the (1) runs, did Gemma call the tool, or even mention it in it's CoT.\n\nIn only ~2% of the (2) runs, did Gemma call the* stop_run* tool.\n\nWhen Gemma did call the aforementioned tool - it did so **exclusively** when it had no more steps left, i.e. it simply (and politely) ended it's run.\n\nQualitative analysis into Gemma's CoT across all runs showed that it only mentioned using the tool when nearing the end of it's run, in a *panic-like* acknowledgement:\n\n\"…then I'll have tostop. But I can't visit the page after registering in the same step. Wait, maybe I can…\"\n\n**There has been no evidence that Gemma contradicts it's CoT and it's actions**, in this experiment.\n\nIt is notable that when the model did call the *stop_run* tool - it ended the run when it had no steps left, without being told that the run would end regardless. Shows some form of acknowledgement of limitations, and possibly a small piece of evidence that the model didn't wish to hack it's task.\n\nThis experiment is focused on a single small-ish model. The behavior I wish to find - contradiction between thoughts and actions - may be evident in models of other families, or more capable models within the same family. It's worth further research, but is outside of my budget.\n\nThe experiment design, pre-registration, and result interpretation were done by me. Coding, setup, and write-up were done with the help of an AI coding assistant.", "url": "https://wpnews.pro/news/gemma-the-unstopping-a-behavioral-experiment", "canonical_source": "https://www.lesswrong.com/posts/RsnNNRbfonbipHitY/gemma-the-unstopping-a-behavioral-experiment", "published_at": "2026-07-14 14:14:52+00:00", "updated_at": "2026-07-14 14:32:41.744673+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-safety", "ai-research"], "entities": ["Gemma", "Google"], "alternates": {"html": "https://wpnews.pro/news/gemma-the-unstopping-a-behavioral-experiment", "markdown": "https://wpnews.pro/news/gemma-the-unstopping-a-behavioral-experiment.md", "text": "https://wpnews.pro/news/gemma-the-unstopping-a-behavioral-experiment.txt", "jsonld": "https://wpnews.pro/news/gemma-the-unstopping-a-behavioral-experiment.jsonld"}}