LLM judge bias vocabulary: six terms and their tests

Picture a judge audit that comes back clean. The ranking ships. A week later somebody re-runs the same pairs with the two answers swapped, and enough verdicts move to overturn it.

That audit measured something real. Agreement against a human reference is a genuine property of a judge, and it is also an aggregate, so it can sit perfectly still while individual verdicts flip underneath it. Order sensitivity was never inside the quantity the audit reported.

Judge bias changes an outcome only where the margin is smaller than the bias, and thin margins are ordinary in model comparison. The evidence for that on this site is our within-Anthropic three-way reliability study, where the top two models finished 1.13 points apart against a 2.0-point winning threshold, with one of the two scored on 77.3% of the instrument’s weight, so we published no ordering inside that tie. That study graded every task mechanically, with no judge model anywhere in the loop, so its margin carries no judge error at all. Put a judge on a contest that close and its bias lands on the deciding digits.

Name the defect before you claim you tested for it. The six terms below split judge failure into three kinds: a preference riding on a presentation variable, a scoring scale offset from true, and a standard that moves while the run is going. Each kind has its own detection test and its own correction, and a test built for one kind comes back clean on the other two.

The word bias is carrying three different jobs here

Read a judge-bias paper and a judge-bias incident report side by side and the same noun covers three unrelated mechanisms.

One of them is a preference that tracks how the answer was presented. Order, length and the model family that wrote it are all presentation variables, and a judge reacting to any of them is scoring a property the rubric never mentioned. Lianmin Zheng and colleagues named position, verbosity and self-enhancement bias together, in the paper that introduced MT-Bench and Chatbot Arena and set off most of the LLM-judging practice that followed (Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena, NeurIPS 2023 Datasets and Benchmarks Track).

An offset in the scale is a second mechanism, and presentation has nothing to do with it. Suppose a judge labels correctly 90% of the time on true passes and 85% of the time on true failures, two schematic round numbers chosen to make the shape visible. Those two rates fix the mapping from the true pass rate to the one that judge reports, and where the reported number lands depends on the true rate as well. At these figures the judge reads 5 pp low when the truth is 80%, 10 pp high when the truth is 20%, and exactly right at 60%. A swap test will never surface any of that, which is why bias-correcting the reported rate runs as arithmetic on a labeled calibration set.

Drift is the third. The standard the judge applies moves between the start of a run and the end of it, or between two rubric versions that read the same to a person. It is the youngest of the three, and the least tested for.

Underneath all of it sits a structural choice most teams make once at setup and never revisit, which is whether the judge sees two answers at a time or one.

Six terms, and what the test for each hands back

Calibration and the pairwise-versus-pointwise choice are the odd two in that list. Calibration describes the scale itself, so it displaces every score the judge produces rather than singling any answer out. The regime choice is a design decision, taken before any grading happens at all. Both earn a place in this vocabulary because both change what a bias test is capable of finding, and filing them under bias is part of why judge audits miss things.

Three of the six ride on how the answer was presented

Position, verbosity and self-preference group together because one move detects all three: hold the content fixed, change exactly one presentation variable, and count how often the verdict moves.

Position bias is the best documented and the cheapest to exploit. Peiyi Wang and colleagues showed that reordering the candidates was enough for Vicuna-13B to beat ChatGPT on 66 of 80 tested queries, with ChatGPT itself acting as the evaluator (Large Language Models are not Fair Evaluators, peer-reviewed, ACL 2024). That’s one benchmark, with one judge, so read the result as an existence proof for how large the effect can get. Your own pairs will have their own rate. The swap test costs one extra inference per pair. Our swap-consistency and position-preference calculator takes the paired counts and returns both rates with Wilson intervals plus an exact binomial test on the direction.

Isolating verbosity bias is harder, because on plenty of tasks length genuinely correlates with quality. Keita Saito and colleagues found GPT-4 preferring longer answers more often than humans did in their preference-labeling setting, and proposed a metric for the gap (Verbosity Bias in Preference Labeling by Large Language Models, arXiv preprint, October 2023). The paper reports a direction with no portable magnitude attached, so it gives you a reason to length-match your arms. The coefficient to correct with is not in it.

Then there is self-preference, which gets confounded with verbosity constantly, since a model’s own outputs tend to carry its own length habits. Arjun Panickssery, Samuel Bowman and Shi Feng found a linear correlation between a model’s ability to recognize its own text and the strength of its self-preference, and showed that correlation resisting the obvious confounders (LLM Evaluators Recognize and Favor Their Own Generations, peer-reviewed, NeurIPS 2024 oral). Separating the two takes a masked re-score, and the per-bias taxonomy with a test and a correction attached to each entry works that procedure through alongside the planted-cue and formatting biases this page does not cover.

Calibration shifts the whole scale, so no swap will find it

That covers the defects a re-run can expose. The next one survives every re-run you can devise, because it never depended on the input.

A judge is a classifier with its own error rates, and those rates displace the pass rate it reports. Run a perfectly order-invariant, length-blind, family-agnostic judge across a test set and it will still hand you a number sitting off the truth, by an amount its sensitivity and specificity set together with the true rate you are trying to measure.

More test items make that number more precise without moving it back toward the truth.

The correction is Rogan-Gladen: invert the judge’s measured error rates back through the raw rate, then report an interval carrying the uncertainty from the calibration set as well as from the test set. Our judge bias correction calculator does that arithmetic, and the end-to-end reporting recipe covers the whole workflow, including where to spend a limited human-label budget.

Getting sensitivity and specificity in the first place means agreeing with a human reference on a labeled set, which is a chance-corrected agreement question rather than a bias one. That is Cohen’s kappa territory, and the inter-rater reliability calculator computes it with an interval attached. Agreement and calibration are separate axes, which is why the judge reliability hub keeps them apart and routes to a different test for each.

Rubric drift moves the standard while the judge stays put

Everything above assumes a fixed rubric. Drop that assumption and a defect appears that neither a swap test nor a calibration set can see.

Rubric drift is the effective standard shifting across a run, or between two rubric versions a person would read as identical. A long evaluation run where the judge’s context fills with its own prior verdicts, a rubric tidied for clarity between releases, a judge model version bumped under an unchanged prompt: each moves the standard without moving the wording you would think to inspect.

Ruomeng Ding and colleagues give the adversarial version of this. Their rubric edits pass benchmark validation and still steer a judge’s preferences on a target domain, with reported target-domain accuracy reductions of up to 9.5% on helpfulness and 27.9% on harmlessness (Rubrics as an Attack Surface: Stealthy Preference Drift in LLM Judges, arXiv preprint, February 2026). The paper is a preprint with no replication behind it yet, and the edits it studies are deliberate, where most teams meet drift by accident. The mechanism carries over either way: an aggregate benchmark score stays flat while the judge’s behavior on your own domain moves underneath it.

The defense is a frozen anchor set, re-scored on a schedule and after every rubric change, which applies reproducibility to the grader instead of to the system under test. Two adjacent terms use the same noun for different subjects. Role drift is an agent abandoning its assigned role over a long run, and retrieval drift is retrieval quality decaying as an index or a query distribution moves. Same word, three different things drifting.

The scoring regime decides which of these can happen at all

One question sits under all five terms above, and it usually gets settled implicitly, at setup, before anyone is thinking about bias at all.

A pairwise judge sees two candidate answers and picks one. A pointwise judge sees a single answer and scores it against the rubric. Position bias is reachable only in the first regime, since there is no order to swap when there is only one answer to score. Rankings assembled out of pointwise scores have their own weakness: they go unstable where the candidates sit close together, which is where a ranking most needs to be right.

Tuhina Tripathi and colleagues put a number on one side of that trade. Where a generator embedded a spurious attribute the judge happened to favor, pairwise preferences flipped in about 35% of cases against 9% for absolute scores (Pairwise or Pointwise? Evaluating Feedback Protocols for Bias in LLM-Based Evaluation, peer-reviewed, COLM 2025). Those flips came from a cue the generator planted on purpose. So the two rates size how far each protocol can be gamed by someone who knows what the judge likes. Ordinary judging, with nobody planting anything, is a separate question that paper does not answer.

Neither regime is the safe one, and the comparison of what each one can certify sets out the trade in full. The choice decides which gate you owe before you publish a ranking: a swap test across every pair, or a labeled calibration set and an interval on the rate.

Where to take a term once you have named it

Naming the defect is the cheap half. Each term has a page that defines it precisely and a page that says what to do about it, and those are deliberately different pages.

The same material turned into a runnable gate order, with explicit thresholds and a calculator behind each gate, is the checklist that gates a ranking before it ships. If you would rather start further back, the three-axis frame that separates agreement, calibration and bias before any of these terms comes up is at testing the evaluator itself. One level up, the four measurable ways an eval number lies maps this whole cluster, and what each type of eval can certify settles what a judge score licenses you to claim out loud.

If what your judge grades is a whole agent run, trajectory-level evaluation is the pillar to read first, because a verdict on a final answer can’t separate a correctly reasoned path from a lucky recovery.

The statistics under every test on this page live next door. Eval statistics routes the interval, power and significance vocabulary; an eval confidence interval is the quantity every rate quoted here should carry; and construct validity is the prior question of whether your rubric measures the thing it names. One trap sits between them and lives in the denominator rather than in the judge: a judge that abstains or errors more often on hard items removes those items from the scored set, so the rate you publish afterwards describes the easy remainder, which is coverage conditioning.

The order these shake out to is short. Decide the regime first, because it fixes which tests are even available to you. Run the swap test if you are pairwise, since that is the only regime with an order to reverse. Length-match your arms whichever regime you picked, because length is the one presentation variable you can control before the judge ever sees the answer. Mask provenance before trusting any comparison that involves the judge’s own model family. Score a labeled calibration set, correct the rate, and publish the interval, which has to carry the uncertainty from the calibration set as well as from the test set. Freeze an anchor set and re-score it whenever the rubric or the judge version changes.

Six terms, one design decision, five gates, and a number at the end that can say what it measured. What none of these gates settles is how many runs each one needs before its own result means anything, and that sizing question gets answered while you are still planning the run.