Jointly tuning the fixed system prompt and the changing user prompt lets large language models answer more accurately while cutting extra wait time.
Most teams still tweak just one side of the prompt puzzle, so the two parts clash and waste compute.
A system prompt is the short paragraph that sets the model’s role, tone, and guardrails, while a user prompt carries the actual question.
Because the system part stays similar across queries and the user part shifts every time, the paper trains them in two different phases.
Offline, it feeds a big batch of past questions into GPT4o‑mini and iteratively edits both prompts until a built‑in judge model scores each answer above a threshold.
This search produces a polished system prompt and a library of “complement” snippets that show the model how to think aloud in a few lines before answering.
Online, when a fresh query arrives, a tiny helper model or a simple retrieval step picks a matching complement from that library, sticks it under the user prompt, and sends the whole package to the main model along with the optimized system prompt.
That trick removes the need for an extra 7B parameter helper at inference and slashes memory from 18000MiB to 5000MiB, while first‑token latency drops about 25%.
Benchmarks covering general Q&A and tricky math show consistent gains, up to 18% on smaller GPT‑3.5 setups, and even stronger jumps for open models like Qwen2.
The key is that the system prompt now expects the complement’s structure, so the main model follows it instead of ignoring it.
----
Paper – arxiv. org/abs/2507.15675
Paper Title: "P3: Prompts Promote Prompting"
Major AI breakthrough: Diffusion Large Language Models are here!
They're 10x faster and 10x cheaper than traditional LLMs.
Here's everything you need to know:
Its not partisan rhetoric to say that $1 of spending cuts to poor people is a much bigger drag on growth than the possible offsetting support from $1 of tax cuts for rich people even though the budget remains flat. It's just macro mechanics.
Most people have NO IDEA just how VALUABLE Grok 3 DeepSearch is for studying your competitors 🦾💥
Before Grok 3 😔
> Spend big $$$ on social listening tools
> Schedule endless customer calls
> Try to guess what people actually want
After Grok 3 🚀
> Tell grok your idea + competitors
> Analyze realtime discussions on X
> Build exactly what people are asking for
If you’re using Grok 3 in business you have a WILDLY unfair advantage
Steal my prompts and thank me + @jelanifuel later ⬇️