PQS 25 (F) - prompt from www.reddit.com

The prompt

this came from researchers at university of michigan and google deepmind. not some random twitter thread. actual peer reviewed stuff

they basically tested every variation of those "you are a world-class financial analyst with 20 years experience at top hedge funds" prompts that everyone copies from linkedin gurus

the expert personas performed worse than just saying nothing at all

like literally leaving the system prompt empty beat the fancy roleplay stuff on financial reasoning tasks

the why is kinda interesting

turns out when you tell the ai its a "wall street expert" it starts acting like what it thinks an expert sounds like. more confident. more assertive. more willing to bullshit you

the hallucination rate nearly doubled with expert personas. 18.7% vs 9.8% with no persona

its basically cosplaying expertise instead of actually reasoning through the problem

they tested across financial qa datasets and math reasoning benchmarks

the workflow was stupidly simple

1. take your query
2. dont add a system prompt or just use "you are a helpful assistant"
3. ask the question directly
4. let it reason without the roleplay baggage

thats it

the thing most people miss is that personas introduce stereotypical thinking patterns. you tell it to be an expert and it starts pattern matching to what experts sound like in its training data instead of actually working through the logic

less identity = cleaner reasoning

im not saying personas are always bad. for creative stuff they help. but for anything where you need actual accuracy? strip them out

the gurus have been teaching us the opposite this whole time