Unified multimodal models gain self-adaptive modes (direct generation, self-reflection, multi-step planning) trained via SFT and RL with step-wise rewards to close the understanding-generation gap in anything-to-image tasks.
If Step 1 turns a 'cat' into a 'tiger', Step 2 must refer to it as 'the tiger', not 'the cat'
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
Unified multimodal models gain self-adaptive modes (direct generation, self-reflection, multi-step planning) trained via SFT and RL with step-wise rewards to close the understanding-generation gap in anything-to-image tasks.