Skip to content

GCRI · 인공지능과 자율 시스템

모형 평가

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.

02질문

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.

메커니즘

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.

메커니즘

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.

메커니즘

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.

메커니즘

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.

03모니터링

모형 등록부.

팀이 이 주제를 따라가는 데 사용할 화면으로, 그 지표, 상태, 결정을 보여 줍니다. 이름은 예시이며 값은 설명용입니다.

모델 레지스트리

Model evaluation · Model registry · Release gate

모형 하나가 관문에서 차단됨GCRI · 인공지능과 자율 시스템

모델

M-184 triage assistant공공 서비스높음19 of 24 suites차단됨
M-320 flood nowcast과학보통12 / 12공개됨
M-157 document search내부낮음8 of 8공개됨
M-563 agent planner에이전트형높음12 of 16평가 중

모델 카드

의도된 용도
분류 지원, 결정은 사람이
범위 밖
자격 결정
알려진 한계
드문 사례에서 정확도가 낮음
소유자
모델 소유자

사고 기록

  1. D-21INC요약에 환각 인용 포함
  2. D-9MIT검색 근거 추가됨
  3. D-2INC업로드된 파일을 통한 프롬프트 주입
  4. D-1MIT도구 권한 축소됨

Evaluation scores

vs release threshold

Task accuracy87 of 85 required견고성80 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.예시 이름과 예시 값으로, 실제 기록이나 운영 중인 시스템이 아닌 설계안입니다.
04영역 전반

Evaluation must reflect the domain where a model is used.

Evaluations are designed for the Labs 증거 시리즈. 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.
05방법과 근거

How model evaluation is modelled, measured and tested.

모형, 그 가정, 데이터, 불확실성을 과학자나 엔지니어가 확인할 형태로 제시합니다. 구조는 발행된 방법을 따릅니다. 제시된 값은 예시이며 결과가 아닙니다.

모형, 방정식 및 가정

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

요소 i의 리스크Ri = f(Hi, Ei, Vi, Ci)재해, 노출, 취약성, 대처 역량
연간 기대 손실EAL = ∑k pk · Lk연간 확률이 pk
결합 시스템xt+1 = A xt + B ut + εtA carries the couplings, u the shock, ε the error
가정만드는 이유시험 방법
한 사건 안에서 재해와 노출은 서로 독립적임손실 적분을 다루기 쉽게 유지함결합 기록에 적합한 코퓰러와 비교
결합은 운전점 근처에서 선형이다충격이 A를 통해 전파되게 함선형 범위를 넘어 섭동을 주고 비교하십시오
취약성 곡선은 장소 간에 옮겨 쓸 수 있다지역 곡선은 거의 없다관측된 손실이 있는 곳에서는 이에 대해 검증
투입 자료
출처 정보가 있는 버전 관리 데이터세트
산출물
분포이며 결코 단일 숫자가 아님
코드
모델 카드 및 시험과 함께 공개
재현
같은 데이터, 코드, 환경과 시드
표준 리스크 및 상태공간 정식화로, 나열된 가정은 해당 방법에 일반적인 것이며 검토를 위해 명시한 것으로 결과가 아닙니다.
FAIR 지침 원칙 · GO FAIR와 FORCE11W3C PROV · 월드 와이드 웹 컨소시엄불확실성의 일관된 처리에 관한 지침 · IPCCCEOS 분석 준비 데이터(Analysis Ready Data) · 지구관측위성위원회
06지표

측정되는 것과 그 정확도.

각 지표에는 단위, 주기, 그 뒤의 출처 종류, 값에 붙을 근거 표시가 있습니다. 값은 표시하지 않습니다. 값은 프로그램이 자체 권한으로 제공합니다.

지표단위주기출처 유형라벨
Benchmark score with uncertaintyscore with confidence intervalper evaluationPublished benchmark results보고됨
Contamination checks reportedpercent of evaluations reporting checksper evaluationEvaluation reports신고됨
Red-team findings by severitycount by severity classper exerciseRed-team reports보고됨
Performance gap between subgroupsdifference in error ratesper evaluationDisaggregated evaluation관측
Post-deployment incident reportscount by type월별Incident monitoring보고됨
보정 오차expected calibration errorper evaluationEvaluation on held-out data관측
07출처

근거의 출처.

이름이 명시된 공개 데이터셋과 기관 기록으로, 각각 라이선스 등급이 있습니다. 협약에 따른 출처는 그 조건에 따라서만 사용하며 결코 재게시하지 않습니다.

출처제공하는 것라이선스 등급
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)
국제 AI 안전 보고서Synthesis 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 characteristics공개, CC BY 4.0
08결정

누가 결정하며 근거는 무엇에 쓰이는지.

인공지능과 자율 시스템 prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.

결정 주체질문이 페이지가 제공하는 근거
개발자Whether 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

이 주제가 생산하는 산출물

  • Evaluation protocols with uncertainty and contamination checks
  • Disaggregated evaluation designs for subgroup performance
  • Red-team methods for public sector use cases
  • Evaluation reporting templates
09핵심 질문

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)?
10메커니즘

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 분포 변화, 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.

11방법과 데이터

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 확률과 통계. 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 사회 및 행동 과학 methods for rater agreement. Human judgement is costly and can be inconsistent.

12방법

연구되는 방식과 방법이 말할 수 없는 것.

  1. AI 평가 및 보증 공학AI 시스템을 위한 보증 사례 · 구조화된 레드 티밍 · 배치 후 모니터링이곳에서의 적용
  2. 확률과 통계벤치마크 점수의 불확실성 · 보정 평가 · 감사 표본 추출이곳에서의 적용
  3. 인공지능과 머신러닝분포 변화 분석 · 특정 용도를 위한 미세 조정 · 해석 가능성 방법이곳에서의 적용
  4. 사회 및 행동과학인간 감독 연구 · 영향받는 지역사회 연구 · 직장 도입 연구이곳에서의 적용
13Where this page’s references

Where this page’s references lead.

이 페이지를 읽을 때 기준이 되는 기구와 문서. 상태와 판은 바뀌며, 각 링크는 발행 기구로 연결됩니다.

15질문

질문과 답변.

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 모델 and in publications once reviewed, each with its scope and limits.
보고서기여하기

이 페이지에서 증거를 기여하십시오.

보고서 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.

GCRI인공지능과 자율 시스템기하급수 기술Labs 근거
인공지능과 자율 시스템

방법, 데이터세트 또는 질문을 가져오십시오.

정보와 문의는 오늘부터 열려 있습니다. 문의에는 참조 번호가 부여되고 다음 단계가 안내됩니다. 신청, 결제, 컴퓨팅 할당, 등록은 해당 팀, 약관, 심사 역량이 갖추어졌을 때만 열립니다.