Wenhao Yu, Zhengcheng Wang, Jiahong Liu, Zexuan Qiu, Minda Hu, Shaohang Wei, Jiafei Lyu, Junyan Lyu · 2026 · arXiv
DOI: 10.48550/arXiv.2610.29965 · open-access PDF
O aprendizado de reforço conjunto e a autodestilação combinam as recompensas de tarefas com a orientação densa do professor para o treinamento de agentes, mas os seus gradientes podem ser parcialmente cancelados. Um coeficiente de mistura controla a escala relativa sem remover seletivamente componentes opostos. apresentamos o CORD, um quadro geral para a otimização consciente de conflitos de dois objetivos. O CORD separa a escolha de um objetivo protegido da condição que desencadeia a projeção e combina os gradientes uma vez por mini-batch do otimizador.
Machine translation (Meta M2M100); the English original follows.
Joint reinforcement learning and self-distillation combines task rewards with dense teacher guidance for agent training, but their gradients can partly cancel. A mixing coefficient controls relative scale without selectively removing opposing components. We present CORD, a general framework for conflict-aware optimization of two objectives. CORD separates the choice of protected objective from the condition that triggers projection, and combines the gradients once per optimizer mini-batch. A nearest-point formulation yields a closed-form update and characterizes its first-order effect on the p
Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese
GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.29965, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txtSources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange