A hierarchical data pipeline and two-stage SFT+RL training with step-wise rewards enables unified models to adaptively handle simple-to-complex X2I tasks and outperform baselines.
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
UNVERDICTED 1representative citing papers
citing papers explorer
-
Breaking Dual Bottlenecks: Evolving Unified Multimodal Models into Self-Adaptive Interleaved Visual Reasoners
A hierarchical data pipeline and two-stage SFT+RL training with step-wise rewards enables unified models to adaptively handle simple-to-complex X2I tasks and outperform baselines.