MBPO uses branch-level, sibling-relative advantages in tree-structured rollouts to improve credit assignment when training multimodal large language models with reinforcement learning.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Multi-Branch Policy Optimization for Multimodal Large Language Models
MBPO uses branch-level, sibling-relative advantages in tree-structured rollouts to improve credit assignment when training multimodal large language models with reinforcement learning.