Prompt Repetition: The Simplest Optimization European AI Teams Are Missing

Google Research published a significant finding in December 2025 that challenges conventional wisdom about prompt engineering: simply duplicating your prompt can unlock dramatic accuracy improvements across all major language models—without adding latency or computational overhead.

Key Developments

The study demonstrated that prompt repetition produces statistically validated gains consistent across OpenAI, Anthropic, Google, and DeepSeek models. Real-world results included:

  • Gemini Flash-Lite: 21-97% accuracy improvement on the NameIndex task
  • GPT-4o-mini: +12% improvement on OpenBookQA with zero latency penalty
  • Cross-model consistency: The effect generalizes across all tested frontier and open-weight models

The mechanism is elegantly simple: causal language models process tokens sequentially during the prefill stage. When you repeat your prompt, tokens in the second instance can attend to a fully populated key-value cache, allowing each subsequent token to observe richer context. Critically, this occurs during prefill—not decoding—so there’s no latency cost.

Why This Matters for European Builders

For Irish and European enterprises already navigating EU AI Act compliance, this finding carries practical weight. With regulatory pressure mounting and budget constraints tightening, discovering a free accuracy win is significant.

European AI teams face a particular challenge: competing with well-funded US labs on model training budgets is unfeasible. But optimizing inference through better prompting? That’s accessible today. This research suggests that many enterprises may be leaving 10-20% accuracy on the table simply through suboptimal prompt structuring.

The consistency across models—OpenAI, Anthropic, Google, and DeepSeek—means European builders aren’t locked into a single vendor. Whether you’re using Claude for safety-critical applications or GPT-4o for production systems, this technique works.

Practical Implications

For developers implementing this:

  1. Zero implementation friction: No model changes, API modifications, or retraining required
  2. Cost-effectiveness: Better accuracy without increased token consumption during decoding
  3. Vendor-agnostic: Works across all tested models, reducing vendor lock-in risk
  4. Quick validation: Easy A/B testing on your specific use cases

However, results vary by task. The 21-97% improvement on NameIndex is substantially larger than the +12% on OpenBookQA, suggesting prompt repetition works better on structured information retrieval tasks than open-ended reasoning.

Open Questions

Several gaps remain:

  • Task specificity: Which problem types benefit most from repetition? How do you identify good candidates?
  • Prompt length scaling: Does the effect diminish with longer prompts or multiple repetitions?
  • Cost-benefit at scale: In production systems handling millions of requests, what’s the practical token-consumption impact even without latency increases?
  • European model testing: How does this generalize to EuroLLM-22B and other EU-developed models?

What’s Next

For Irish and European enterprises, the immediate action is straightforward: test prompt repetition on your highest-value tasks. Given the zero cost and potential double-digit accuracy gains, this belongs in every inference optimization checklist.

The broader implication is subtly powerful: frontier model performance isn’t fixed. Simple post-training optimizations—not requiring model access or retraining—can still unlock significant gains. In a landscape where European competitors face capital constraints against US-dominated labs, that’s meaningful leverage.

Source: Google Research