Researchers tested whether using a tool's results in a synthetic assignment task affects a language model's answer. They found that when a tool result is used, the model is more likely to select the code from that result, but only if it's not just repeating what was previously written. However, when the same information is presented as plain text, the model behaves similarly. This suggests that while tool results can influence a language model's behavior, they may not carry more authority than plain text in all cases. --- Why it matters: This study matters to AI researchers because it highlights the importance of understanding how different presentation formats affect language models' decisions. The findings have implications for designing synthetic tasks and evaluating the reliability of language models. Source: https://arxiv.org/abs/2608.14992