LMArena Elo explained: the blind-battle crowd vote behind 15% of our League Score.

By AI Pro League · Published Oct 30, 2026

LMArena Elo explained starts with its place in our own table: it carries 15% of the League Score and it is the only one of the five components that is not a test at all. Only 10 of the 59 models in our database carry a recorded LMArena Elo, against 31 with a LiveBench score. This explainer uses our own LMArena leaderboard as the worked example throughout, so every number below can be checked against the table itself.

What LMArena Elo measures

Our benchmarks hub describes this component as “crowd preference Elo from blind head-to-head battles (formerly LMSYS Chatbot Arena)”, and our glossary puts it the same way: a crowd platform where humans vote between two anonymous models’ answers, producing preference Elo ratings. There are no set questions, no answer key and no examiner — a voter sees two replies to the same prompt without knowing which model wrote either, and picks the one they prefer. That is why this component anchors our best AI chatbot ranking: it is the closest public proxy in our database for how a model feels to talk to, rather than how it sits an exam. The platform itself is LMArena; the scores we record come from the snapshot named on each row of our leaderboard (the recorded LMArena source), each with a measured date.

How an LMArena Elo score is calculated — and what normalisation does to it

Elo is a relative rating, borrowed from chess: beating a higher-rated opponent moves a rating more than beating a lower-rated one, and the gap between two ratings implies an expected win rate, 1 / (1 + 10^−diff/400). A 100-point gap implies about a 64.0% expected win rate — a real edge, not a rout. Our League Score then normalises the raw Elo across the small field we track — today the recorded range runs from 1197 to 1522, a spread of 325 points — and weights the result at 15%. The worked example, from the top of our table:

Now apply the Elo formula to that top of the table, because it is this component’s central lesson: the gap between first and second in our record is just 8 points, which implies an expected win rate of about 51.2% — a coin flip with a thumb on the scale, not a championship margin. Even the full recorded field, from 1522 down to 1197, implies only about 86.7% for the top model against the bottom one. Elo gaps have to be read in tens and hundreds, never in single points. Full method: how the League Score is built.

What a good LMArena Elo score looks like

In our recorded field the median score is 1479, the floor is 1197 and the ceiling is 1522. The top is compressed — the first five models sit inside a 43-point band — while the tail falls away, which is the normal shape of a preference rating: voters separate the frontier pack from the rest far more clearly than they separate the pack’s own members. The top of the table:

The full table, with a source link on every row, is the LMArena leaderboard, and the standings this component helps shape are the AI model rankings.

Why so few models have an LMArena Elo score

Coverage, not quality, is the main reason: 10 recorded scores out of 59 tracked models means most of the field has no recorded rating in the source we use, and an absent score contributes nothing — it is not treated as a zero. A preference rating also needs a model to have faced the crowd at all, which exam benchmarks do not: a lab can sit a written test on demand, but it cannot manufacture blind votes. That scarcity is why this component carries 15% rather than the 40% given to the LiveBench component, which 31 models carry.

The limits to keep in mind

• Preference is not correctness. A voter can prefer the confident, well-formatted wrong answer; Elo records the preference honestly, which is a different claim from accuracy.
• Small gaps are ties. As the worked example shows, a single-digit Elo gap implies a near-50% expected win rate — treat the top band as a pack, not a podium.
• The field is tiny. Ten models is a leaderboard, not a census — one new rating redraws the normalisation for everyone.
• It is one component. At 15% of the League Score it is the human-preference voice beside LiveBench, SWE-bench Verified, GPQA Diamond and MMLU-Pro — by design, it can break a tie but not crown a winner alone.

For the companion tests, start with our explainers of MMLU-Pro and LiveBench, or the wider guide to AI benchmarks explained. That completes our explainer set: all five League Score components now have one.