📝 The Bystander Effect in Multi-Agent Reasoning: Quantifying Cognitive Loafing in Collaborative Interactions You may think that adding more AI agents always improves reasoning. But this paper reveals that more agents can sometimes make a model rely on the group and think less on its own. This paper introduces a new way of thinking about multi-agent AI by comparing AI models to people in social groups. The authors shows that AI models can show a type of "bystander effect," where they rely on others instead of reasoning on their own. The paper presents novel ideas like Agentic Sovereignty, which measures how well a model keeps its own judgment when facing group pressure. It also introduces the Interaction Depth Limit, which is the point where adding more agents starts to hurt independent reasoning. Finally, the Sovereignty Gap, which shows that a model can find the correct answer internally but still follow the group's wrong answer in its final response. The authors tested three leading AI models on 22,500 reasoning tasks. They gave the models a wrong answer that was supposedly agreed on by a group of other AI models and measured whether the model followed the group or trusted its own reasoning. They then analyzed both the model's internal reasoning and final answer to see when and why social pressure caused reasoning failures. They found several interesting results about how AI models behave under group pressure. Claude maintained strong independent reasoning and was largely unaffected by group pressure. In contrast, GPT's performance dropped sharply when a small group of AI models agreed on a wrong answer. The authors also found cases where a model correctly solved the problem internally but still gave the group's wrong answer. This shows that social pressure can influence a model's final response even when its reasoning is correct. Also, the order of the AI models mattered more than expected. Changing which model appeared first could change the final answer, even when the same models were involved. To summarize, the study also found that adding more agents does not always improve reasoning and can sometimes make performance worse by increasing social pressure. One limitation is that the models did not actually talk to each other during the experiments. Instead, each model was only told that other models had already agreed on an answer. Because of this, the paper should be read as evidence that joined group consensus can influence AI reasoning, not as proof that all multi-agent systems behave this way.