Copying Your Prompt Twice: I Tested the '+76%' Trick and Found the Truth

The author tested a viral prompt-duplication trick referenced to a Google Research preprint 'Prompt Repetition Improves Non-Reasoning LLMs'. In the original, duplicating the prompt helped non-reasoning models in 47 of 70 comparisons, with the +76% figure coming from a single stress test on Gemini 2.0 Flash-Lite, which has since been discontinued. The author ran their own experiment: 3360 queries, $1.48, on gemini-2.5-flash-lite (reasoning off) and gemini-3.5-flash-lite (thinking minimal and high). On the stress suite (NameIndex and MiddleMatch tasks), duplicating the prompt improved accuracy by +38.8 percentage points on the 2.5 model, replicating the original effect, but had no effect on everyday tasks. A third copy added only +5 p.p., and adding a marker like 'Repeat:' made no difference. On the 3.5 model with thinking minimal, the effect persisted (+36.3 p.p.), but with thinking high, accuracy was already 100% on both prompts. The trick also stabilized an otherwise non-deterministic model. However, the author notes that the trick specifically fixes attention lapses on long lists, not arithmetic or other reasoning deficits, and advises caution: results are limited to specific model IDs and dates, and the dataset is self-made.