An AlpacaEval win rate is a judge's preference, and the longer answer can win it

The number a release leads with

When a new model ships, its announcement often leads with an AlpacaEval win rate. Zephyr-7B, a small open model from 2023, reported a 90.60% win rate on the first version of AlpacaEval.¹ Read that the way most people do and it sounds like a test score, as if the model got about nine answers in ten right. It is not that. An AlpacaEval win rate is the fraction of a fixed set of prompts on which an automatic judge preferred this model's answer to a second model's answer.² Nothing in it measures whether either answer was correct.

The judge is another language model, and the win is its preference. A preference can be swayed by things that have nothing to do with a better answer, and the largest of them is length: a wordier answer tends to win.

How the judge counts a win rate

AlpacaEval runs the same four steps for every model it scores.²

1. Fix the prompts

   The benchmark holds 805 instructions, the same set for every model, from writing tasks to factual questions.²

2. Collect two answers

   For each instruction the model under test writes an answer, and a reference model writes one too. In the first version the reference is text-davinci-003, an older OpenAI model.²

3. Ask the judge

   An automatic judge, a GPT-4 model, reads the two answers to that instruction and says which one it prefers.²

4. Take the fraction

   The win rate is the number of instructions where the judge picked the test model's answer, divided by 805.²

So Zephyr's 90.60% is a plain count you can back out. Multiply 0.9060 by 805 and you get about 730: the GPT-4 judge preferred Zephyr's answer to text-davinci-003's on roughly 730 of the 805 instructions.¹ The number is a preference share, not an accuracy. It never asks whether an answer was right, only which of two answers the judge liked more, and it is measured against one particular reference model rather than against the truth.²

Why a wordier answer scores higher

The judge's preference is not only about quality, and the AlpacaEval team's own follow-up work shows what else moves it. Dubois and colleagues took one model, a GPT-4 turbo, and changed only the instruction telling it how verbose to be, from concise to standard to verbose, while the 805 prompts and the judge stayed fixed.³ On AlpacaEval 2.0 that model's own win rate moved from 22.9% to 64.3%.³ The model was the same, the questions were the same, the judge was the same. The only thing that changed was how much it wrote.

This happens because the judge is a language model, and language models tend to rate a longer, more detailed answer as more helpful, even when the extra words add nothing the shorter answer lacked.³ A model that pads its answers is not answering better. It is giving the judge more of what the judge already rewards.

A shorter cheat beats the length fix

The AlpacaEval authors built a defense against this, the length-controlled win rate, usually written LC. It is a statistical adjustment that estimates what the judge's preference would be if the two answers were the same length, so a model cannot gain by writing more alone.² It helps. On the verbosity test, length control shrank the swing from 22.9%–64.3% down to 41.9%–51.6%, and the amount a model could move its score on wording alone fell from about a quarter to a tenth.³ It also lines up better with people. The Spearman correlation is a measure of how closely two rankings agree, running up to one for a perfect match. Length control raised it, between the AlpacaEval leaderboard and Chatbot Arena, a board built from human head-to-head votes, from 0.94 to 0.98.³ An independent review of the paper reports the same rise.⁴ (The AlpacaEval repository rounds the before-figure to 0.93.²)

Length control narrows the bias without closing it. To show the gap, Zheng and colleagues built a null model: a single canned response that ignores the instruction and returns the same text every time.⁵ On AlpacaEval 2.0 that constant, off-topic answer scored a 76.9% raw win rate and an 86.5% length-controlled win rate.⁵ The debiased number came out higher than the raw one. Because the cheat's answer was short, the length adjustment worked in its favor.⁵ Length control makes winning by verbosity harder, not impossible, and a short empty answer can still beat most real models on the score meant to be harder to game.

The reference model sets the bar

Return to Zephyr's 90.60%. Its paper reported that number,¹ and the model's own card still lists it.⁶ On the current AlpacaEval 2.0 leaderboard, the same zephyr-7b-beta reads 13.20% length-controlled and 10.99% raw.⁷ Nothing about the model changed between those readings. What changed was the setup around it.

The two references sit at very different levels. Version 1.0 scores answers against text-davinci-003, a weak model, so most answers beat it and win rates run high. Version 2.0 scores answers against GPT-4 turbo, a top model, so the same answers rarely win and rates run low.² The 90.60% and the 13.20% answer two different questions. The first asks whether Zephyr beats text-davinci-003. The second asks whether it beats GPT-4 turbo, after the length adjustment. Both can be true at once because the bar moved.

A win rate only means something next to its version, its judge, and its reference, and two win rates are comparable only when all three match.² A leaderboard that holds all three fixed, as the AlpacaEval 2.0 board does, is comparing like with like. A press release that quotes one model's 90.60% beside another model's number from a different version is not.