Skip to content

GCRI · 人工知能と自律システム

モデル評価

Model evaluation asks what an AI model can do, how reliably, and where it fails. Benchmarks, red teaming, human studies and field monitoring each measure something different, and none fully predicts behaviour in deployment. This page sets out how evaluations are designed and where they mislead, how results should be reported, and what evidence serves the developers, deployers and regulators who rely on them.

02質問

Model evaluation: how the risk develops.

Validity, calibration and generalisation. The dossier below sets out how the risk develops, what is measured, where the data come from, how it is modelled and who decides.

メカニズム

Benchmarks saturate and leak

Models quickly reach high scores on fixed benchmarks, and test items can leak into training data. High scores then say more about the benchmark than about general capability.

メカニズム

Test conditions differ from use

Evaluations use set prompts, settings and tasks. Real users phrase things differently, combine tools and face new situations, so measured performance may not transfer.

メカニズム

Adversarial testing finds hidden failures

Red teaming and adversarial prompts reveal unsafe or unreliable behaviour that average-case metrics miss. Coverage depends on the testers’ skill and the time allowed.

メカニズム

Performance drifts after deployment

Model updates, changing inputs and new uses alter behaviour over time. Without monitoring, a model evaluated at release can behave differently months later.

03監視

モデルレジストリ。

このテーマを追跡するためにチームが使う画面:その指標、状態、決定。名前は例であり、値はイメージです。

モデル・レジストリ

Model evaluation · Model registry · Release gate

1 つのモデルがゲートで阻止されたGCRI・人工知能と自律システム

モデル

M-184 triage assistant公共サービス高19 of 24 suites阻害中
M-320 flood nowcast科学中12 / 12公開済み
M-157 document search内部低8 of 8公開済み
M-563 agent plannerエージェント型高12 of 16評価中

モデルカード

想定用途
トリアージ支援、判断は人が行う
対象外
適格性の決定
既知の限界
まれな事例での精度が低い
オーナー
モデルの所有者

インシデントログ

  1. D-21INC要約に実在しない引用が含まれる
  2. D-9MIT検索拡張による根拠づけを追加
  3. D-2INCアップロードされたファイルによるプロンプトインジェクション
  4. D-1MITツール権限を縮小

Evaluation scores

vs release threshold

Task accuracy87 of 85 required頑健性80 of 75 requiredFairness gap (lower is better)6 of 5 requiredRefusal on misuse85 of 95 required
Model registry: the screen’s structure, terms and states for Model evaluation.例示の名称と説明用の値です。設計であり、記録でも稼働中のシステムでもありません。
04分野を横断して

Evaluation must reflect the domain where a model is used.

Evaluations are designed for the Labs Evidence series. The International AI Safety Report (2025) describes the limits of current evaluations; NIST’s AI RMF and its generative AI profile (2024) set measurement practice. Models would carry cards under NS-12 of the proposed Nexus Standards.

  • Held-out test sets for a domain.
  • Evaluation protocols with known failure cases.
05手法とエビデンス

How model evaluation is modelled, measured and tested.

モデル、その仮定、データ、不確実性を、科学者やエンジニアが確認する形で示します。構造は公表された手法に従っています。描かれた値は例示であり、知見ではありません。

モデル、方程式、仮定

The model structure used for model evaluation, the assumptions it makes and how each is tested.

要素 i のリスクRi = f(Hi, Ei, Vi, Ci)ハザード、エクスポージャー、脆弱性、対処能力
年間期待損失EAL = ∑k pk · Lk年間確率 pk
結合システムxt+1 = A xt + B ut + εtA carries the couplings, u the shock, ε the error
仮定作られる理由検証の方法
一つの事象の中では、ハザードとエクスポージャーは独立している損失の積分を扱いやすく保つ同時記録に当てはめたコピュラと比較
結合は動作点の近くで線形であるショックを A を通じて伝播させる線形範囲を超えて摂動を与えて比較する
脆弱性曲線は場所を越えて適用できる地域固有の曲線はめったに得られない観測された損失がある場合はそれに照らして検証
入力
来歴付きのバージョン管理されたデータセット
成果物
分布であり、単一の数値ではない
コード
モデルカードと試験とともに公開
再現
同じデータ、コード、環境、シード
標準的なリスクおよび状態空間の定式化です。挙げた前提はこの手法に典型的なもので、審査のために示しており、結果ではありません。
FAIR 原則・GO FAIR と FORCE11W3C PROV・World Wide Web Consortium不確実性の一貫した扱いに関するガイダンス・IPCCCEOS 解析即応データ・地球観測衛星委員会
06指標

何を、どの程度うまく測定しているか。

各指標には、その単位、更新頻度、背後にある出典の種類、値が持つことになるエビデンスのラベルが付いています。値は表示しません:プログラムが自らの権限のもとで値を提供します。

指標単位周期出典の種類ラベル
Benchmark score with uncertaintyscore with confidence intervalper evaluationPublished benchmark results報告値
Contamination checks reportedpercent of evaluations reporting checksper evaluationEvaluation reports申告済み
Red-team findings by severitycount by severity classper exerciseRed-team reports報告値
Performance gap between subgroupsdifference in error ratesper evaluationDisaggregated evaluation観測
Post-deployment incident reportscount by type月次Incident monitoring報告値
較正誤差expected calibration errorper evaluationEvaluation on held-out data観測
07出典

エビデンスの出所。

名前の明示された公開データセットと機関の記録。それぞれにライセンス区分があります。合意に基づく出典は、その条件のもとでのみ使用し、再公開することはありません。

出典提供するものライセンス区分
Stanford HELM (Holistic Evaluation of Language Models)Open, multi-metric evaluations of language modelsOpen source, Apache 2.0 for code; results public
MLCommons benchmarks, including AILuminateStandardised performance and safety benchmarksPublic results, MLCommons terms (to check)
国際 AI 安全性報告書Synthesis of evidence on capabilities, risks and evaluation limitsPublic, terms per publisher
Model cards and system cards from developersDeveloper-reported evaluations and limitationsPublic documents, terms per developer
Epoch AI benchmark and model databasesCompiled benchmark results and model characteristicsオープン、CC BY 4.0
08決定

誰が決め、エビデンスが何に役立つか。

人工知能と自律 システム prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.

誰が決定するか質問このページが提供するエビデンス
開発者Whether a model is ready for release and what limits to stateEvaluation protocols; red-team findings
Deploying organisationsWhether a model is fit for a specific useUse-specific evaluations; subgroup performance
Regulators and AI safety institutesWhat evaluation evidence to requireEvaluation standards and reporting templates

このトピックが生み出す成果物

  • Evaluation protocols with uncertainty and contamination checks
  • Disaggregated evaluation designs for subgroup performance
  • Red-team methods for public sector use cases
  • Evaluation reporting templates
09中核となる問い

What evaluators ask.

  • Does the test measure the capability or property that matters for the intended use (construct validity)?
  • When the model reports high confidence, is it right as often as that confidence implies (calibration)?
  • Does performance hold on data from different places, times or populations (generalisation)?
  • Has test data leaked into training, inflating results (contamination)?
10メカニズム

Why scores mislead.

Models learn statistical regularities in their training data, including spurious ones. A medical imaging model may learn the hospital’s scanner signature rather than the disease; a language model may recognise benchmark questions it saw during training. When deployment data differ from test data, a problem known as 分布シフト, performance can fall sharply and without warning.

Large general-purpose models add new difficulties. Their capabilities are broad and emergent, outputs vary with phrasing, and the line between capability and memorisation is blurred. Aggregate scores can hide failures concentrated in particular groups or tasks.

11手法とデータ

How evaluation is done, and its limits.

  1. 01

    Held-out and out-of-distribution testing

    Testing on data deliberately different from training, analysed with 確率と統計. No test set covers every deployment condition.

  2. 02

    Calibration analysis

    Comparing stated confidence with observed accuracy, and checking how it changes under shift.

  3. 03

    Disaggregated evaluation

    Reporting performance by subgroup and task rather than only on average. Small subgroups give uncertain estimates.

  4. 04

    Human and expert evaluation

    Structured assessment of open-ended outputs, using 社会・行動科学 methods for rater agreement. Human judgement is costly and can be inconsistent.

12手法

どのように研究され、手法が言えないことは何か。

13Where this page’s references

Where this page’s references lead.

このページの読解の基準となる機関と文書。状態と版は変わります。各リンクは発行機関へ移動します。

15質問

質問と回答。

Why are benchmark scores not enough?
Because models can be trained on test items, benchmarks saturate, and test conditions differ from real use. Use-specific, adversarial and post-deployment evaluation are needed as well.
What is red teaming in AI?
Structured adversarial testing in which people or automated tools try to make a model produce harmful, insecure or unreliable outputs, to find failures before deployment.
What is data contamination?
The presence of benchmark test items in a model’s training data, which inflates scores and makes results unreliable as measures of general capability.
Is a high benchmark score evidence a model is safe to deploy?
No. It shows performance on that test under its conditions. Fitness for a use needs evaluation on representative data, with failure analysis and monitoring after deployment.
What is benchmark saturation?
When models reach near-perfect scores, the benchmark no longer discriminates between them, and further gains may reflect overfitting to the test rather than real capability.
Why evaluate after deployment too?
Because data, users and the model itself can change. Monitoring detects drift and new failure modes that pre-release testing could not anticipate.
Will the platform publish evaluation protocols?
Protocols are designed to appear on モデル and in publications once reviewed, each with its scope and limits.
報告書貢献する

このページからエビデンスを提供する。

報告書 is the Consortium’s quarterly publication. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work.

GCRI人工知能と自律システム指数関数的テクノロジーLabs のエビデンス
人工知能と自律システム

手法、データセット、または問いを持ち込んでください。

情報提供と問い合わせは本日から受け付けています。問い合わせには参照番号が付き、次の手順が示されます。申請、支払い、計算資源の割り当て、登録は、担当チーム、規約、審査体制が整ってから開始します。