Inter-Annotator Agreement

Definition

Inter-annotator agreement measures how often two or more judges assign the same label to the same item, corrected for the agreement you would expect by chance alone. In Search Evaluation the items are query-document pairs and the labels are relevance grades, so it is the statistic that answers “would a different assessor have graded this the same way?”

It is the label-level statistic in the hierarchy described under Levels of Judge Agreement, and the level people most often quote — and most often over-read.

The Statistics

StatisticWhat it handlesWhere it misleads
Raw agreement / accuracyNothing beyond counting matchesLooks excellent when 90% of items are non-relevant
Cohen’s / Fleiss’ κAgreement beyond chanceMoves with class balance; two collections are not comparable
Krippendorff’s αMultiple annotators, missing data, any measurement scaleSame class-balance sensitivity
Quadratic-weighted κOrdinal grades — a 0-vs-3 error counts more than 0-vs-1An average hides one-sided (systematically lenient) error
MAE on the grade scaleHow far off, on averageSays nothing about direction
Per-class precision / recall / F1Which distinctions the judge can actually makeF1 has no universal chance baseline — it is not on κ’s scale

The Landis-Koch Bands Are Not a Standard

The bands everyone quotes — fair 0.21–0.40, moderate 0.41–0.60, substantial 0.61–0.80 — come from a 1977 biometrics paper by Landis and Koch. They have nothing to do with information retrieval and were offered as a convenience. They are usable as shared vocabulary and should not be used as a gate.

Humans Are the Baseline, Not 1.0

Relevance disagreement long predates LLM judges. Voorhees (2000) showed that different sets of human judgments change many individual labels while system comparisons largely hold; Bailey and colleagues (2008) found that who judges matters — the person who wrote the topic, a task expert, and a random assessor are not interchangeable, and test collections are not fully robust to swapping judges of very different expertise.

The consequence is a rule about setting bars. If your own annotators agree at κ 0.6, demanding κ 0.8 from a model is incoherent. Measure human-human agreement on your own data first; that number, not 1.0, is what to ask of a machine judge. The MT-Bench framing is the same one: GPT-4’s ~80% agreement with human preferences was reported as meaningful precisely because it is “the same level of agreement between humans.”

Reporting Rules

Three rules recur across the LLM as Judge literature, and each exists because a single agreement number concealed something:

Never post a κ without the label distribution and confusion matrix. A judge that is lenient uniformly across the board can post a perfectly normal κ. The leniency is the part that bites — LLM judges are repeatedly found more generous than human assessors.

Report per class, not just overall. Allegro’s RAT is the sharpest case: quadratic-weighted κ of 0.69 overall — genuinely good for a four-level ordinal scale — against per-class F1 of 0.94 on exact matches and 0.83 on complements, but 0.51 and 0.33 on separating “highly substitutable” from “substitutable”. The overall figure was excellent on the calls nobody needed help with and hid poor discrimination on the boundary the business actually argues about. Note that the overall weighted κ and the per-class F1 are different quantities on different scales; F1 0.33 does not mean “coin flip”. For a chance-corrected per-class figure, compute a one-vs-rest binary κ and label it as such.

Don’t quietly collapse graded labels to binary. Squashing 0/1/2/3 into relevant/not manufactures agreement and discards exactly the distinctions NDCG depends on. If you do collapse, report both.

Typical Values in Relevance Judging

Reported figures are consistently modest, and vary with collection, model, prompt, grading scale, the context given to the judge, and who the humans were:

  • Human-human κ can exceed 0.5 on binary assessment; LLM-human κ typically runs 0.3–0.5 depending on the scale.
  • NormasTCU (2026), Brazilian Portuguese legal search: Cohen’s κ 0.32–0.53, MAE 0.46–0.66 on a 0–2 scale — while its nDCG@10 and MRR leaderboards cleared Kendall τ 0.9.
  • Allegro, Polish e-commerce, quadratic-weighted κ 0.69 for a local quantized judge.

Fair-to-moderate label agreement is the normal condition, not a failure state. Better agreement happens when the task is tightly specified — Keller et al. (2026) found judges given only a short query “judge many more documents relevant and have a lower agreement” than judges given a proper topic description, and that generating the description automatically still helps.