How an AI says it matters as much as what it says
The founder of AI's most-watched leaderboard says presentation changes comprehension, so his team is building ways to score style separately from correctness.
Lightspeed Venture Partners · Meta Can Read Your Brain, The AI Science Race & AI Ranking With Ion Stoica | LightworkIon Stoica, a Berkeley professor and co-founder of LMArena, was explaining an awkward problem with ranking artificial intelligence. LMArena is the site where anyone can ask a question, get answers from two unnamed chatbots, and vote for a winner. Those votes feed a chess style rating, where beating a strong opponent counts more than beating a weak one.
The difficulty came up when he was asked how subjective votes become an objective ranking. For math or computer code, he said, the answer works or it does not. For a poem, reasonable people can disagree about which version is better.
That is when Stoica made his case for style. “I can tell you the same information but in one case it can be boring, and in other case it can be more energetic, crisp,” he said.
He argued that the difference changes learning. The same facts, presented one way versus another, make what he called a huge difference in what the other person understands and retains. In his framing, model makers and app builders want to learn how to present information so people take it in best.
He compared the issue to user experience in any other field, like a refrigerator or a television, where the underlying capabilities can be the same and how people experience them still matters.
That belief is shaping what his team is trying to build. Stoica said LMArena is working on ways to disentangle factual correctness from stylistic delivery, and that more would be coming from the team on separating the two. His point was that both matter, and both should be counted.
Questions people ask
Do chatbots ever win just by sounding nicer?
Yes. Human voters often favor answers that are clear, confident, and well formatted, even when another answer is more accurate. That is why Arena says it is working to score factual correctness separately from stylistic delivery.
How do you tell an objective task from a subjective one?
For questions with a clear answer, checkers can test the result directly, whether the math is right or the code runs. For open-ended work like writing or advice, there is often no single right answer, so human preference is the closest available measure.