In the Arena: How LMSys changed LLM Benchmarking Forever

Apologies for lower audio quality; we lost recordings and had to use backup tracks.

Our guests today are Anastasios Angelopoulos and Wei-Lin Chiang, leads of Chatbot Arena, fka LMSYS, the crowdsourced AI evaluation platform developed by the LMSys student club at Berkeley, which became the de facto standard for comparing language models. Arena Elo is often more cited than MMLU scores to many folks, and they have attracted >1,000,000 people to cast votes since its launch, leading top model trainers to cite them over their own formal academic benchmarks:

The Limits of Static Benchmarks

We’ve done two benchmarks episodes: Benchmarks 101 and Benchmarks 201. One issue we’ve always brought up with static benchmarks is that 1) many are getting saturated, with models scoring almost perfectly on them 2) they often don’t reflect production use cases, making it hard for developers and users to use them as guidance.

The fundamental challenge in AI evaluation isn't technical - it's philosophical. How do you measure something that increasingly resembles human intelligence? Rather than trying to define intelligence upfront, Arena let users interact naturally with models and collect comparative feedback. It's messy and subjective, but that's precisely the point - it captures the full spectrum of what people actually care about when using AI.

The Pareto Frontier of Cost vs Intelligence

Because the Elo scores are remarkably stable over time, we can put all the chat models on a map against their respective cost to gain a view of at least 3 orders of magnitude of model sizes/costs and observe the remarkable shift in intelligence per dollar over the past year:

This frontier stood remarkably firm through the recent releases of o1-preview and price cuts of Gemini 1.5:

The Statistics of Subjectivity

In our Benchmarks 201 episode, Clémentine Fourrier from HuggingFace thought this design choice was one of shortcomings of arenas: they aren’t reproducible. You don’t know who ranked what and what exactly the outcome was at the time of ranking. That same person might rank the same pair of outputs differently on a different day, or might ask harder questions to better models compared to smaller ones, making it imbalanced.

Another argument that people have brought up is confirmation bias. We know humans prefer longer responses and are swayed by formatting - Rob Mulla from Dreadnode had found some interesting data on this in May:

The approach LMArena is taking is to use logistic regression to decompose human preferences into constituent factors. As Anastasios explains: "We can say what components of style contribute to human preference and how they contribute." By adding these style components as parameters, they can mathematically "suck out" their influence and isolate the core model capabilities.

This extends beyond just style - they can control for any measurable factor: "What if I want to look at the cost adjusted performance? Parameter count? We can ex post facto measure that."

This is one of the most interesting things about Arena: You have a data generation engine which you can clean and turn into leaderboards later. If you wanted to create a leaderboard for poetry writing, you could get existing data from Arena, normalize it by identifying these style components. Whether or not it’s possible to really understand WHAT bias the voters have, that’s a different question.

Private Evals

One of the most delicate challenges LMSYS faces is maintaining trust