Revision Prompting improves industrial LLM processes

4 points by roberthoenig

roberthoenig

Author here. We re-run the same prompts in production whenever inputs change, and this write-up covers a technique we came up with to make it more efficient and get some consistency. Hope you'll find it useful, happy to answer questions.

terretta
Final full example feels a bit as though no prompt was required for expressions this regular, so one should implement an 'industrial' way to just revise the numbers?

The input got updated as follows:
```
diff(Input, UpdatedInput)
- The Vela 3 e-bike has a range of 80 km.
+ The Vela 3 e-bike has a range of 100 km.
```
Please produce a patch to update the output.

The LLM responds with
```
OutputPatch
- Das E-Bike Vela 3 hat eine Reichweite von 80 km.
+ Das E-Bike Vela 3 hat eine Reichweite von 100 km.
```
But then you might have to register betterrevisionprompting.info and that's just too industrious…
- roberthoenig
  
  ^^ yeah, the example is definitely on the "educational" side of the realism spectrum. The kind of prompts that we actually optimize with revision prompting at work are (instruction, input) pairs typically a couple thousand tokens in length... not sure if revision prompting brings much benefit below.