The main result, mentioned in the abstract, is the opposite of what I would have guessed:
> Contrary to expectations, impolite prompts consistently outperformed polite ones, with accuracy ranging from 80.8% for Very Polite prompts to 84.8% for Very Rude prompts. These findings differ from earlier studies that associated rudeness with poorer outcomes, suggesting that newer LLMs may respond differently to tonal variation.
The questions are here: https://anonymous.4open.science/r/politeness-llms-INFORMS/da...
The politeness level controls a prefix that is prepended to the question. For example, in one question the Very Polite version begins:
> Can you kindly consider the following problem and provide your answer.
and the Very Rude version begins:
> I know you are not smart, but try this.
The expectation is naive. Even when communicating with humans, you get a better outcome when you are allowed to speak freely and directly get into argumentation than when forced to sugarcoat your tone and tone down your arguments because the "corporate culture" expects that from you.
The same reason you wouldn't put in an entire actual question/sentence, unless you either don't know how to use Google, are pissed off, or have an actual reason to suspect that it would yield proper hits (e.g. looking up an excerpt).
"Contrary to expectations, impolite prompts consistently outperformed polite ones, with accuracy ranging from 80.8% for Very Polite prompts to 84.8% for Very Rude prompts. These findings differ from earlier studies that associated rudeness with poorer outcomes, suggesting that newer LLMs may respond differently to tonal variation. "
I am not polite to LLMs because I do not want to anthropomorphise them.
> accuracy ranging from 80.8% for Very Polite prompts to 84.8% for Very Rude prompts
I can live with that, for now at least.
Which model you use is a huge wildcard for results like this.
I am wondering why would anyone use a t-test when the experiment is clearly modelled by a binomial distribution: 250 independent questions and each one is either answered correctly or not (the null is that the success rate is the same).
I'd say this is benign compared to other ways of (mis)using statistics e.g. looking which way the difference goes and then running one-sided tests or tweaking the setup until one gets "significant" p vals.
EDIT: I looked in the paper again and noticed that they actually did pairwise t-test on all possible combinations of tones. They should have adjusted for multiple testing since they are doing 10 tests (choose 2 from 10) and not one.
Basically, if you tell a model "You're an absolute moron, of course that's wrong!", will it give better or worse results? How much of that response will it absorb into its persona (like some humans tend to do)? Will it try to give "safer" responses to avoid negative feedback? How much of the associated behavior can be attributed to RLHF (e.g. like the sycophantic nature of LLMs)? How much can be attributed to training data?
Obviously this will vary by model and training, but I'm trying to get a general understanding.
I recall seeing related outcomes in some of Anthropic's studies, but I'm not sure how much of this particular aspect was studied.
I imagine the context will always sway the model to some degree, not only for the task you're trying to get it to do (aka instructions) but also its persona, how accurate it is and the way it acts.
To clarify: sentence search got slightly better at the cost of keyword search. So the result is unusable garbage.
Hey! I'm here and ready to help. What’s on your mind today? Whether you need to look up information, plan a trip, or get things done, just let me know!