Skip to content

GCRI · Inteligência artificial e sistemas autônomos

Avaliação de modelos

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.

02A pergunta

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.

Mecanismo

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.

Mecanismo

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.

Mecanismo

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.

Mecanismo

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.

03Monitoramento

Registro de modelos.

A tela que uma equipe usaria para acompanhar esta matéria: suas medidas, estados e decisões. Os nomes são exemplos e os valores são ilustrativos.

Registo de modelos

Model evaluation · Model registry · Release gate

Um modelo bloqueado na etapa de controleGCRI · Inteligência artificial e sistemas autônomos

Modelos

M-184 triage assistantServiço públicoElevado19 of 24 suitesBloqueado
M-320 flood nowcastCiênciaMédio12 de 12Lançado
M-157 document searchInternoBaixo8 of 8Lançado
M-563 agent plannerAgênticoElevado12 of 16Avaliando

Ficha do modelo

Uso previsto
apoio à triagem, decisão humana
Fora do âmbito
decisões de elegibilidade
Limites conhecidos
menor precisão em casos raros
Proprietário
Responsável pelo modelo

Registo de incidentes

  1. D-21INCCitação alucinada no resumo
  2. D-9MITAncoragem de recuperação adicionada
  3. D-2INCInjeção de prompt por arquivo enviado
  4. D-1MITPermissão da ferramenta reduzida

Evaluation scores

vs release threshold

Task accuracy87 of 85 requiredRobustez80 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.Nomes de exemplo e valores ilustrativos; um projeto, não um registro nem um sistema ativo.
04Em todos os domínios

Evaluation must reflect the domain where a model is used.

Evaluations are designed for the série Labs Evidence. 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.
05Métodos e evidência

How model evaluation is modelled, measured and tested.

O modelo, os seus pressupostos, os seus dados e a sua incerteza, na forma que um cientista ou engenheiro verificaria. As estruturas seguem métodos publicados. Os valores representados são ilustrativos, não conclusões.

Modelo, equações e pressupostos

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

Risco do elemento iRi = f(Hi, Ei, Vi, Ci)ameaça, exposição, vulnerabilidade e capacidade de enfrentamento
Perda anual esperadaEAL = ∑k pk · Lksobre os eventos k com probabilidade anual pk
Sistemas acopladosxt+1 = A xt + B ut + εtA carries the couplings, u the shock, ε the error
PressupostoPorque é feitoComo é testado
A ameaça e a exposição são independentes dentro de um eventoMantém tratável o integral de perdasComparar com um ajuste de cópula sobre registros conjuntos
Os acoplamentos são lineares perto do ponto de funcionamentoPermite que os choques se propaguem através de APerturbar para além do intervalo linear e comparar
As curvas de vulnerabilidade transferem-se entre lugaresAs curvas locais raramente estão disponíveisValidar face às perdas observadas quando existam
Entradas
conjuntos de dados versionados com proveniência
Resultados
distribuições, nunca números isolados
Código
publicado com a ficha de modelo e os testes
Reprodução
mesmos dados, código, ambiente e semente
Formulações padrão de risco e de espaço de estados; os pressupostos listados são típicos do método, indicados para revisão, não resultados.
Princípios orientadores FAIR · GO FAIR e FORCE11W3C PROV · World Wide Web ConsortiumOrientação sobre o tratamento consistente das incertezas · IPCCCEOS Analysis Ready Data · Committee on Earth Observation Satellites
06Indicadores

O que é medido e com que qualidade.

Cada indicador leva sua unidade, sua periodicidade, o tipo de fonte por trás dele e a etiqueta de evidência que um valor levaria. Nenhum valor é mostrado: um programa os fornece sob sua própria autoridade.

IndicadorUnidadeCadênciaTipo de fonteRótulo
Benchmark score with uncertaintyscore with confidence intervalper evaluationPublished benchmark resultsREPORTADO
Contamination checks reportedpercent of evaluations reporting checksper evaluationEvaluation reportsDECLARADO
Red-team findings by severitycount by severity classper exerciseRed-team reportsREPORTADO
Performance gap between subgroupsdifference in error ratesper evaluationDisaggregated evaluationOBSERVADO
Post-deployment incident reportscount by typemensalIncident monitoringREPORTADO
Erro de calibraçãoexpected calibration errorper evaluationEvaluation on held-out dataOBSERVADO
07Fontes

De onde vêm as evidências.

Conjuntos de dados públicos nomeados e registros institucionais, cada um com sua classe de licença. As fontes obtidas por acordo só são usadas nos seus termos e nunca são republicadas.

FonteO que ofereceClasse de licença
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)
Relatório internacional sobre segurança da IASynthesis 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 characteristicsAberto, CC BY 4.0
08Decisões

Quem decide, e a que servem as evidências.

Estúdio de Inteligência artificial e sistemas prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.

Quem decideA perguntaEvidências a que esta página serve
DesenvolvedoresWhether 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

Produtos que este tema gera

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

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)?
10Mecanismos

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 mudança de distribuição, 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.

11Métodos e dados

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 probabilidade e estatística. 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 ciências sociais e comportamentais methods for rater agreement. Human judgement is costly and can be inconsistent.

13Where this page’s references

Where this page’s references lead.

Os organismos e documentos à luz dos quais esta página é lida. O estado e as edições mudam; cada ligação leva ao organismo emissor.

15Perguntas

Perguntas e respostas.

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 do país and in publications once reviewed, each with its scope and limits.
RelatóriosContribuir

Contribua com evidências a partir desta página.

Relatórios 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.

GCRIInteligência artificial e sistemas autônomosTecnologia exponencialEvidências de Labs
Inteligência artificial e sistemas autônomos

Traga um método, um conjunto de dados ou uma pergunta.

A informação e os pedidos estão abertos hoje. Um pedido recebe uma referência e indica o próximo passo; candidaturas, pagamentos, alocações de computação e registros só abrem quando existem a equipa, os termos e a capacidade de revisão.