Researchers have found that large language models can be biased by irrelevant text. When given extra information not related to the task at hand, these models tend to make consistent predictions across various benchmarks. This effect is quantified through a 'decision margin' which measures the difference in log-probability between different outcomes. The study reveals a geometric regularity where this margin follows an affine transformation of its original value when context-free. This finding suggests that irrelevant text doesn't add random noise, but rather distorts model preference in a predictable way. --- Why it matters: This matters because it shows how large language models can be influenced by external information, even if it's not directly relevant to the task. Understanding this effect is crucial for developing more robust and reliable AI systems that can handle noisy or irrelevant context. Source: https://arxiv.org/abs/2608.19208