How to see AI Rankings in Factuality Leaderboards

Last updated: July 15, 2026

Factuality leaderboard ranks models according to a weighted combination of human preference and the correctness of claims in a model’s response. Factuality is currently in our Text Arena leaderboard and Search Arena leaderboard, starting as a non-default toggle which can be selected in our leaderboard UI.

How To View

To view the Factuality leaderboard:

1. Go to the Text Arena leaderboard or Search Arena leaderboard

2. Click Show Filters

3. Click on Corrections

4. Select Factuality

FAQ

Question: How do you define factuality?

Answer: Factuality is defined as a claim being true according to verifiable information from credible sources on the internet.

Question: Why is factuality important?

Answer: Factuality captures how often a model’s stated or implied claims are accurate—users should never be presented false information. When factuality is on, each model's factuality rating is blended into its overall score, so models that are reliably correct rank higher.

Question: How is factuality different from hallucination rate?

Answer: Hallucination is broader than factuality, and harder to measure. Both factuality and hallucination cover cases where models present incorrect factual information (e.g “The Eiffel Tower is 330 meters tall”). Hallucinations could additionally cover, for example, local context from the conversation (e.g. saying “I ran all tests and they passed” — when nothing was run).

Question: What is Style Control? How is factuality different?

Answer: Both style control and factuality adjust for the same worry: that humans will choose a response filled with incorrect information on the merits of its presentation (formatting, length, tone, etc). While style control measures stylistic elements in model responses and attempts to apply corrections based on the estimated influence these elements have on human preference, factuality instead measures the problem directly—auditing model responses for outright errors.

Question: How do the leaderboards change with factuality included?

Answer: Some model scores will go up and others will go down. Models that go down significantly will have much lower factuality than their peers. Models that go up will have much higher factuality than their peers.

Question: How is factual correctness judged?

Answer: Arena extracts individual claims made by the model in its response, focusing on web-verifiable claims. We then verify each claim one-by-one with multiple verifiers, producing calibrated factuality labels.

Question: Why is the default Factuality weight set at 25%?

Answer: We want a balanced weighting between factuality and human preference; 25% punishes unfactual models heavily enough such that they cannot score high on the leaderboard, but also respects nuances of human preference. Note that factuality verification is very pedantic; this may not always align with what is most useful to people.

Question: Can I change the weight to something other than 25%?

Answer: Not at this moment, but this is something our methodology supports. You can see how changing the weight affects the ranking of different models on our blog.

Question: Wouldn’t I always want factuality at 100% to ensure I'm receiving the most accurate information?

Answer: Factuality is very pedantic—more so than the average person would be. One response might be very helpful, with lots of information, of which 90% is true. The other response might have very little information but is 100% true. Only when the stakes are extremely high would someone prefer the latter. The easiest way to never produce incorrect information is to never provide any claims at all. While this is perfectly factual, it is also completely useless.

Question: How do the rankings change taking Factuality into account?

Answer: Each model is scored from two signals: preference (which response people pick as better overall) and factuality (which response is more factually correct). The standard leaderboard ranks on preference alone. Turning on Factuality folds the factuality signal into the Score at a 25% weight, so a well-liked model that is loose with facts drops, while a precise model climbs. The score is a relative strength rating (Elo-style), not a percentage of anything — higher just means stronger.