For the 10% of similar documents without correct answers, question-answering accuracy dropped from 68% to 45.5%.

NVIDIA and Texas A&M University researchers found that mixing hard distractors—documents semantically similar to a question but containing no answer—into long inputs at just a 10% share cut Qwen2.5-7B-Instruct's question-answering accuracy from 68% to 45.5%, a failure mode they term the "first ink drop effect." The paper, revised August 20, 2026, has three authors—one from NVIDIA's AI Technology Center and two from Texas A&M—and distinguishes hard distractors from irrelevant or random documents because they mimic plausible answer candidates and compete with the document that actually holds the answer. The finding matters because retrieval-augmented and long-context systems routinely pad inputs with topically similar passages, and a small share of such near-misses can sharply degrade accuracy. What to watch next is whether the effect holds across other model families and context lengths, and whether retrieval filtering or distractor-aware prompting can recover the lost accuracy.