Does every pair need direct comparison?

Not necessarily, but comparisons must provide enough connected information to relate the items; disconnected groups cannot be ranked against each other from the data alone.

Can a system with a higher score lose?

Yes. Predictions are probabilistic.

Does the model require a neural network?

No. Neural scoring models are one possible application of its pairwise likelihood.

AI Glossary / Statistics and probability / Advanced

Bradley-Terry Model

Definition

The Bradley-Terry model is a probabilistic model for pairwise comparisons. It assigns each item a positive strength and models the probability that one item defeats, or is preferred to, another as its strength divided by the sum of their strengths. An equivalent formulation applies a logistic sigmoid to the difference between two real-valued scores.

What it means

People can often compare two answers more easily than assign each answer an absolute quality score. The Bradley-Terry model turns collections of such comparisons into relative scores and predicted preference probabilities. A higher score means a higher modeled chance of preference under the evaluation conditions. It is not an intrinsic measurement of truth, safety, or universal usefulness. If judges favor verbosity, familiar phrasing, or presentation order, the collected preferences can reflect those influences as well as the intended quality criterion.

How it works

Let positive strengths be wA and wB. The model gives P(A preferred to B) = wA/(wA+wB). Writing each strength as the exponential of a score gives P(A preferred to B) = sigmoid(sA-sB).
Fit scores by maximizing the likelihood of observed comparisons, or use a justified regularized or Bayesian approach. In the basic independent-comparison formulation, each preference contributes a binary likelihood term. Extensions can include comparison-specific effects, such as presentation order, or variation associated with judges.
Only relative strengths are identified. Multiplying all strengths by the same positive number leaves probabilities unchanged; equivalently, adding a constant to every score changes nothing. A reference score or another normalization is therefore needed. Sparse or separated comparison results can also cause estimation difficulties.

Three practical examples

An evaluation team collects blinded comparisons between two assistants' answers to the same prompts. A preference model summarizes the comparisons, while the team checks how the selected prompts and judges affect interpretation.
A reward-modeling project assigns scores to candidate responses and fits their differences to preference labels. These response scores can depend on the prompt rather than being one permanent score for each assistant.
A product team compares alternative descriptions in paired judgments. The fitted model helps summarize relative preference, but comprehension and factual accuracy are measured separately when they matter to the decision.

Think of it this way

Imagine a competition table that describes chances of winning rather than declaring a guaranteed winner. The numbers acquire meaning through comparisons with other competitors, not through the arbitrary location of zero on the score scale.

Worked example

Consider three invented assistants with strengths wA = 4, wB = 2, and wC = 1. The model predicts P(A beats B) = 4/(4+2) = 2/3, P(B beats C) = 2/(2+1) = 2/3, and P(A beats C) = 4/(4+1) = 4/5.
For twelve comparable A-versus-B judgments, the expected number favoring A is 12 times 2/3 = 8. For nine B-versus-C judgments, the expected number favoring B is 6. For fifteen A-versus-C judgments, the expected number favoring A is 12. These are expectations, not guaranteed counts.
Suppose the observed counts happen to be exactly 8 of 12, 6 of 9, and 12 of 15. These pairwise proportions are jointly compatible with the strengths above. Each binomial likelihood is maximized at its observed proportion, and this single set of relative strengths achieves all three maxima simultaneously. Under the stated basic model, it therefore supplies a maximum-likelihood fit, up to the common scale convention.
The corresponding scores can be sA = log(4), sB = log(2), and sC = 0. The A-versus-B score difference is log(2), approximately 0.6931, representing odds of 2 to 1 rather than probability 0.6931. The logistic conversion gives probability 2/3.
Multiplying the strengths by ten gives 40, 20, and 10, with identical predictions. Adding five to all three log-strength scores likewise changes no pairwise difference.
Now consider a different pattern: A beats B with probability 0.75, B beats C with probability 0.75, and C beats A with probability 0.75. The required strength ratios would be wA/wB = 3, wB/wC = 3, and wC/wA = 3. Multiplying them gives 27, but the strengths must cancel to give 1. A basic single-score Bradley-Terry model cannot represent all three probabilities exactly. Such a pattern can indicate context dependence or a poor model fit, rather than an arithmetic mistake to conceal.

Illustrative case study

Fictional company Maple Answers compares three assistants using a common set of business prompts. Judges receive paired answers with system identities hidden and display positions randomized. The team records ties instead of silently forcing all uncertain judgments into wins or losses.
Analysts fit an appropriate preference model, inspect comparisons that the model explains poorly, and report uncertainty around relative performance. They also examine factual accuracy and results by task type. When one assistant leads overall but performs poorly on a particular task, the report preserves that distinction. This scenario is fictional.

Common misconceptions and mistakes

A probability of 0.8 means an 80-percent modeled preference chance for the specified comparison, not that an answer is 80-percent correct.
Scores have an arbitrary shared origin. Comparing raw scores from unrelated fits can be misleading without a common reference and compatible data.
The basic two-outcome model does not automatically handle ties. Tie treatment or an extended model must be specified.
Observed wins can form cycles even when probabilities follow the model. The limitation concerns representing systematic cyclic preference probabilities with one scalar strength per item.
Repeated judgments from the same people or prompts need appropriate treatment of dependence; counting every label as independent can overstate precision.

Business use

Use the model to summarize a well-designed pairwise evaluation, not to replace evaluation design. Define the preference question, sample relevant tasks and judges, record uncertainty, and inspect model fit. Keep separate requirements such as cost, latency, factual quality, and task-specific suitability visible when selecting a system.

Check your understanding

With strengths 4 and 1, what is the first item's preference probability?

Answer: 0.8.

Does scaling both strengths change it?

Answer: No.

Can the basic model exactly represent the three 0.75 cyclic preferences above?

Answer: No; their required odds ratios contradict one another.

Related terms

Preference learning
Ranking
Pairwise comparison
reward model
logistic sigmoid
maximum likelihood
relative strength

Learn next

Study uncertainty in rankings, comparison-graph connectivity, separation, and extensions for ties or judge effects. Compare preference scores with independently measured task outcomes.

Sources

Turner and Firth, Bradley-Terry Models in R: https://stat.ethz.ch/CRAN/web/packages/BradleyTerry2/vignettes/BradleyTerry.html . Jurafsky and Martin, Speech and Language Processing, preference-modeling account: https://web.stanford.edu/~jurafsky/slp3/old_aug25/9.pdf . All numerical examples and the company scenario are original illustrations.

Previous: Bracketed notation Next: Branching factor