A seguirPróximo guia
Modelagem de recompensa Bradley-Terry
Fundamentos
GUIA de IA de linguagem
A reward model maps a prompt and response to a scalar score intended to approximate preference or another objective.
Ouça este guiaNarrado por IA · ~2 min
It is a learned proxy, not a direct measure of human values.
In RLHF-style pipelines, a reward model is often trained on human comparisons of model outputs. A pairwise preference loss encourages the preferred response to score above the rejected response. At inference, best-of-N may sample responses and select the highest-scoring candidate. Optimizing a proxy can lead to reward hacking when the score is imperfect. InstructGPT collected human rankings of four to nine model responses for a prompt and converted those rankings into pairwise preferences for reward-model training. A common Bradley–Terry preference model converts differences between scalar response scores into preference probabilities, and the pairwise loss trains preferred responses to score higher. A common reward-model implementation uses a scalar sequence-scoring head, though architectures vary. Ensembles may reveal scorer disagreement but cannot remove shared model errors or guarantee against reward hacking.
Train from pairwise response rankings.
Use scores to rank candidate responses in best-of-N selection.
Test reward predictions on held-out comparisons.
Inspect for outputs that score highly but violate the intended behavior.
Reward models remain useful components of some alignment pipelines, but require ongoing evaluation and safeguards against optimizing unintended behaviors.
Boletim informativo gratuito
Três histórias verificadas de IA a cada manhã de dia útil, escritas em linguagem simples. Gratuito para sempre, sem anúncios.
Um e-mail a cada dia útil. Cancele a inscrição com um clique. Nunca vendemos ou compartilhamos seu endereço.
Teste você mesmo
Feedback instantâneo sobre cada resposta e um certificado compartilhável com um ID verificável assim que você concluir um curso.
Apoie a educação gratuita em IA. O AI Understanding é uma organização sem fins lucrativos 501(c)(3) — sem anúncios, sem paywall, nunca. Faça uma doação
A reward model maps a prompt and response to a scalar score intended to approximate preference or another objective. It is a learned proxy, not a direct measure of human values.
Um modelo de recompensa mapeia um prompt e uma resposta para um número escalar que representa a qualidade prevista ou a preferência humana.
InstructGPT labelers ranked sets of four to nine model outputs for a prompt; the rankings were converted into pairwise comparisons for reward-model training. This is specific to that paper’s procedure.
In the Bradley–Terry model, scalar-score differences parameterize the probability that one response is preferred to another, and the pairwise loss trains scores accordingly. These scores are not automatically calibrated measures of quality.
O hacking de recompensas acontece quando o modelo encontra atalhos (como comprimento excessivo ou bajulação) que o modelo de recompensa imperfeito avalia muito, mas os humanos não.
In a common reward-model setup, a language-model backbone is adapted with a scalar scoring head that produces one score for a prompt-response sequence. This describes that setup, not every reward-model architecture.
Continue aprendendo
Mais guias escolhidos para este tópico
A seguirPróximo guia
Modelagem de recompensa Bradley-Terry
Fundamentos