GPT-4o is interesting to learn about - but it’d be great to test again with frontier models of May/June 2026 and see if these effects are gone, different, or the same.
Which model you use is a huge wildcard for results like this.
Most of the comments here seem to be from people who haven’t even read the abstract, let alone the paper.
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.
Your example sounds to me a little different: the first one asks the LLM to employ some token budgets in presenting facts in a kind voice, the second asks the LLM to give strong evidence, the difference is not polite versus rude. My one shot guess from your sole example should be that social pressure to prove oneself can affect LLM accuracy.
I have always said please and thank you to LLMs, not to increase accuracy or because I'm stupid. I believe it is more about me than about the LLM, and this is anyway a habit I don't want to lose.
Thomas Aquinas believed cruelty to animals was wrong not because animals have souls (and with that all the standard moral rights), but because it can teach us cruelty to other humans.
Is it worth getting worse results for that reason? From the article:
"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.
Google searches being keyword based, rather than simulated conversations?
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).
I searched for "Hey Google" and got this in response:
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!
I got downvoted for asking a related question recently, but I also don't think people really understood what I was asking - I'm not trying to anthropomorphise LLMs to that extent.
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.
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).
The methods could be better described in the paper, but my understanding is that they did 10 runs for each question for each prompt and took an average of those, so the compared values are not binary. You could do a sign test, but you'd lose power and answer a bit different question.
You can do a generalised mixed effects linear model with binomial outcome (ie a binomial test but with added random effects structure). But unless you want to introduce a richer random effects structure with more variables, it is overkill and overcomplicating things, and the result should be the same as t-tests.
I don't know much about stats, but does "the null is that the success rate is the same" imply that it's a sketchy methodology because they can come up with some findings ("ruder prompts are better/worse!") more often?
You are asking about one-sided vs two-sided tests. Not really "more often" because formal type 1 error rate is still the same. I'd say two-sided tests leave more space for post-hoc theorizing but there are valid situations when there is no clear one-sided hypothesis a priori. Do we really know whether that the hypothesis should have been "ruder prompts are better"?
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.
it sort of makes sense to me,
when asking a question to an expert in the field while you are a student. I would guess the successful interactions on average would be more polite . Like for example if you were asking a question to donald knuth or terrence tao, you'd probably be polite while doing so. Being hostile while asking questions gets you into forum discussion territory.
> 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.
I guess it makes sense since we as humans tend to be far less inclined to help someone who is not polite/is not friendly, so that "bias" is part of the training data, thus influences how LLMs function
> 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.
Which model you use is a huge wildcard for results like this.
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.
"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.
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).
To clarify: sentence search got slightly better at the cost of keyword search. So the result is unusable garbage.
Not feeding them tokens is neglect.
I try to feed them a healthy diet.
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.
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.