The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning
Training-inference mismatch in reinforcement learning for large language models leads to instability, which is addressed through a new policy optimization objective and framework t…
Hugging Face · Daily Papers
·Jing Liang, Hongyao Tang
·
·▲ 53 upvotes
Este artigo está em destaque na seleção diária de papers do Hugging Face, curada pela comunidade de pesquisa em IA.
Autores: Jing Liang, Hongyao Tang, Yi Ma, Yancheng He, Weixun Wang, Xiaoyang Li
- 53 upvotes da comunidade
- Temas: reinforcement learning, large language models, training-inference mismatch, off-policy, policy optimization, policy improvement
Resumo
Resumo original (em inglês), extraído do paper:
Training-inference mismatch in reinforcement learning for large language models leads to instability, which is addressed through a new policy optimization objective and framework that ensures consistent policy improvements between training and inference phases.Onde ler
// relacionados
Leia também
Editorial
AutoMem: e se lembrar fosse uma habilidade que a IA aprende, não um sistema que a gente projeta?
Editorial
LongCat-2.0: a Meituan abre um modelo de 1,6 trilhão de parâmetros treinado só em chips chineses
Blog
Amazon sunsets Mechanical Turk, the original "Artificial Artificial Intelligence"
Blog