GUIDA ALL'AI linguistica

Chiedere ai LLM quanto sono sicuri

Verbalized confidence is a score or phrase a language model reports about its own answer, such as “I am 80% confident.” The number is generated as text and is not automatically a calibrated probability of correctness.

Ascolta questa guidaNarrato dall'IA · ~3 min

In questa pagina3 minuti di lettura
  1. Panoramica
  2. Immersione profonda
  3. Implementazione nel mondo reale
  4. The Future of Asking LLMs How Confident They Are
  5. Metterlo in pratica
  6. Continua a esplorare
  7. Domande frequenti

Panoramica

Studies find useful signals in some settings but also substantial miscalibration and overconfidence.

Immersione profonda

A verbal confidence score is an answer the model generates in natural language. Asking for a percentage does not by itself create a measurement process that knows whether the answer is correct. The score can still correlate with correctness on a particular task, but a team must test that relationship. Calibration asks whether predictions assigned a confidence level are correct at a corresponding rate over a relevant set of cases; it is assessed across examples, not by inspecting one answer. Research findings are mixed because results depend on models, prompts, and evaluation data. A 2023 study of human-feedback-tuned models found verbalized confidence better calibrated than token probabilities on TriviaQA, SciQ, and TruthfulQA in its experiments. A 2024 evaluation using different language and vision-language models and difficult uncertainty tasks found high calibration error and overconfidence in its tested settings. These findings are not contradictory rules for every modern model: they show why the confidence signal needs evaluation on the intended task and current model version. Before using a score to escalate or approve decisions, collect labeled examples that resemble production traffic. Compare confidence buckets with observed accuracy, choose a threshold against the costs of wrong answers and unnecessary review, and monitor drift. Expected Calibration Error is one summary of mismatch across bins, but it can conceal uneven performance across subgroups or tasks. Token log probabilities describe likelihoods of generated tokens, not a direct probability that the full answer is correct. Repeated-answer agreement can add evidence about stability, but repeated samples may share the same misconception. Treat verbal confidence as one input to a validated decision process, not a substitute for external evidence or expert review.

Implementazione nel mondo reale

A support team asks a model to label tickets and report confidence, then checks the score against a labeled sample before using it to route tickets.

A researcher compares verbal confidence with actual correctness on answerable trivia questions and finds that calibration depends on the benchmark and prompt.

A model reports 90% confidence on several answers; reviewers resist interpreting the same round number as a measured 90% success rate.

A service tests whether multiple independently worded answers agree, treating agreement as another signal rather than proof that the answer is true.

The Future of Asking LLMs How Confident They Are

Uncertainty reporting is likely to use several signals, including model outputs, repeated-sample agreement, retrieval evidence, and task-specific calibration. Better tooling may make it easier to evaluate these signals, but calibration can shift when models, prompts, or input populations change. Teams should refresh labeled tests and keep human review for decisions where a false confident answer carries meaningful harm. A score that worked last quarter may no longer fit current users. Reassessment should be part of routine deployment maintenance for every model release.

Metterlo in pratica

Perché è importante

  • Velocità e scala. I flussi di lavoro linguistici possono muoversi più velocemente senza sacrificare la coerenza.
  • Accedere e raggiungere. Espande l'accesso attraverso lingue e stili di comunicazione.
  • Decisioni più chiare. I team possono dedicare più tempo al giudizio mentre l'automazione gestisce la ripetizione.

Attenzione

  • Fatti allucinati possono tranquillamente entrare nei rapporti, nei flussi di supporto o nei risultati della ricerca.
  • La sensibilità tempestiva può creare risultati incoerenti tra richieste simili.
  • I dati di testo sensibili potrebbero essere esposti se i controlli di accesso sono deboli.

Prossimi passi

  1. Definisci il formato di output, il tono e gli standard di qualità prima dell'implementazione.
  2. Risposte concrete con fonti attendibili ogni volta che la precisione è importante.
  3. Mantenere un checkpoint di revisione umana per i risultati ad alto rischio.
  4. Tieni traccia dei modelli di errore e riqualifica regolarmente le richieste o i flussi di lavoro.

Continua a esplorare

Newsletter gratuita

Ricevi il briefing quotidiano sull'IA

Tre storie verificate sull'IA ogni mattina nei giorni feriali, scritte in un inglese semplice. Gratuito per sempre, senza pubblicità.

Un'email ogni giorno feriale. Disiscriviti con un clic. Non vendiamo né condividiamo mai il tuo indirizzo.

Mettiti alla prova

Take the Asking LLMs How Confident They Are quiz

Feedback immediato su ogni risposta e un certificato condivisibile con un ID verificabile una volta superato un corso.

Inizia il quiz

Sostieni l'istruzione gratuita sull'IA. AI Understanding è un'organizzazione non profit 501(c)(3) — nessun annuncio, nessun paywall, mai. Fai una donazione

Domande frequenti

What is Asking LLMs How Confident They Are?

Verbalized confidence is a score or phrase a language model reports about its own answer, such as “I am 80% confident.” The number is generated as text and is not automatically a calibrated probability of correctness. Studies find useful signals in some settings but also substantial miscalibration and overconfidence.

What does a model-generated “80% confident” statement establish by itself?

The guide explains the score is generated as text and is not automatically calibrated against correctness.

How should calibration of a confidence score be assessed?

Calibration compares confidence and observed accuracy across examples from the intended task.

What does the guide say about research on verbalized confidence?

The guide cites studies with differing results in their tested models and benchmarks.

What does ECE summarize?

ECE compares average stated confidence with observed accuracy within bins.

What do generated-token log probabilities directly describe?

Token log probabilities describe model likelihoods over output tokens and are not direct correctness probabilities.