The First Drop of Ink: Nonlinear Impact of Misleading Information in Long-Context Reasoning

A new arXiv paper (2605.10828) reveals that in large language models, even a small proportion of hard-distractor documents within long contexts causes a sharp performance drop, a phenomenon termed "The First Drop of Ink" effect. The study systematically varied the hard-distractor proportion in fixed-length contexts and found a nonlinear pattern: performance declines steeply with the first small fraction of distractors, while additional distractors yield marginal further decline. Theoretical and empirical analyses grounded in attention mechanics explain this effect, showing that hard distractors cap performance. The research addresses the critical issue of how misleading information affects long-context reasoning in retrieval-augmented generation and agentic systems.