Chatbot Arena updates leaderboard to weight model factuality

Chatbot Arena updates leaderboard to weight model factuality · Digg

Commentary on X

It's probably because these labs are obsessed with safety and so they try to increase the depth of the thinking by removing things that are actually real but polemic. So they pile up and now you have a schizophrenic model that's a great technocrat and has no idea how the world actually works.

10:32 AM · Jul 15, 2026

Anastasios Nikolas Angelopoulos@ml_angelopoulos

@cormac_mars @arena If you look on a relative basis, Grok improved with respect to the other models when incorporating factuality, and is doing well on being non-sycophantic and seeking truth according to results on the Arena. cc @elonmusk

9:55 AM · Jul 15, 2026

Kizyto@kizi74479

@ml_angelopoulos Factuality filtering changes the game. OpenAI's models holding steady, they've been training for truth-seeking. Did you see any unexpected winners?

10:46 AM · Jul 15, 2026

Anastasios Nikolas Angelopoulos@ml_angelopoulos

The Arena ranking has officially moved to a mixture of human preference and other signals. Today, our main text leaderboard has begun taking into account factuality. It is another step in the direction of incentivizing models to promote objective truth when talking to real users, instead of hallucinating or sycophanting. You can see that the biggest winner on factuality is OpenAI. Almost none of their models decrease with factuality. Grok has also improved due to its focus on truth-seeking and objectivity. Results from Anthropic and Google are mixed. Meta's Muse Spark and Dola Seed drop by quite a bit, although Muse Spark 1.1 is still a big improvement over 1.0 and the Llama series. So proud of this release and all the work that went into it by the amazing team at Arena. Developments like this will continue pushing the frontier to benefit humanity.

9:56 AM · Jul 15, 2026

View all

Anastasios Nikolas Angelopoulos@ml_angelopoulos

The Arena ranking has officially moved to a mixture of human preference and other signals. Today, our main text leaderboard has begun taking into account factuality. It is another step in the direction of incentivizing models to promote objective truth when talking to real users, instead of hallucinating or sycophanting. You can see that the biggest winner on factuality is OpenAI. Almost none of their models decrease with factuality. Results from Anthropic and Google are mixed. Meta's Muse Spark and Dola Seed drop by quite a bit, although Muse Spark 1.1 is still a big improvement over 1.0 and the Llama series. So proud of this release and all the work that went into it by the amazing team at Arena. Developments like this will continue pushing the frontier to benefit humanity.

9:43 AM · Jul 15, 2026

Combined views

Commentary on X

Combined views

Sentiment

Positive Read

Useful Links

Digg Deeper

Commentary on X

Sentiment

Positive Read

Useful Links

Digg Deeper