arXivAgentsMethodImportance 8arXiv1mo ago

InterleaveThinker: Reinforcing Agentic Interleaved Generation

Dian Zheng, Harry Lee, Manyuan Zhang, Kaituo Feng, Zoey Guo, Ray Zhang, Hongsheng Li

0 CitationsPopularity 0.7

AI analysis

TL;DR

InterleaveThinker is a multi-agent pipeline that enables existing image generators to produce interleaved text-image sequences by leveraging planner and critic agents to plan, evaluate, and refine the generation process.

Problem

Recent image generators excel at single-image generation and editing but cannot perform interleaved generation (text-image sequences) due to architectural limitations. Existing Unified Multimodal Models (UMMs) also show limited performance on this task.

Approach

Planner Agent: Analyzes the input image-text sequence and instructs the image generator on the required action at each step.

Critic Agent: Evaluates the generator's outputs, identifies samples deviating from planned instructions, and refines instructions for regeneration.

Data Construction: Builds Interleave-Planner-SFT-80k and Interleave-Critic-SFT-112k for format cold-start via supervised learning.

Reinforcement Learning: Uses Interleave-Critic-RL-13k with GRPO to reinforce step-wise instruction correction within a generation trajectory. Introduces accuracy reward and step-wise reward to enable single-step RL to effectively guide the entire trajectory.

Results & Contribution

InterleaveThinker improves interleaved generation performance across various image generators (e.g., FLUX, SD3). It achieves performance comparable to Nano Banana and GPT-5 on interleaved generation benchmarks. Additionally, it significantly enhances the base model on reasoning-based benchmarks (WISE, RISE).