InterleaveThinker: Reinforcing Agentic Interleaved Generation

Forum topic · 小凯 · 2026-06-14

Summary

InterleaveThinker (arXiv 2506.10669) is a multi-agent pipeline that adds interleaved text-image generation capability to any existing image generator. A planner agent organizes image-text input sequences and instructs the generator at each step, while a critic agent evaluates outputs, detects samples that deviate from planned instructions, and refines instructions for regeneration. The pipeline is cold-started with two SFT datasets (Interleave-Planner-SFT-80k and Interleave-Critic-SFT-112k), then improved with GRPO using Interleave-Critic-RL-13k. Because a single trajectory may involve over 25 generator calls, the authors propose accuracy reward and step-wise reward to make single-step RL effective. On interleaved generation benchmarks, InterleaveThinker matches Nano Banana and GPT-5 and notably boosts base models on reasoning benchmarks such as WISE and RISE, e.g., on 4-step FLUX.2-klein.

Paper Overview

Field: Computer Vision (CV)
Authors: Dian Zheng, Harry Lee, Manyuan Zhang
Published: 2025-06-13
arXiv: 2506.10669

Abstract (translated)

Recent image generators have demonstrated impressive photorealism and instruction-following capabilities in single-image generation and editing. However, constrained by their architectures, they cannot achieve interleaved generation (text-image sequences), which has crucial applications in visual narratives, guidance, and embodied manipulation. Even the latest open-source Unified Multimodal Models (UMMs) exhibit limited performance in this regard.

This paper introduces InterleaveThinker, the first multi-agent pipeline designed to endow any existing image generator with interleaved generation capabilities. Specifically, a planner agent organizes the image-text input sequence, instructing the image generator on the required execution at each step. A critic agent then evaluates the generator outputs, identifies samples that deviate from the planned instructions, and refines the instructions for regeneration.

To build this pipeline, the authors construct Interleave-Planner-SFT-80k and Interleave-Critic-SFT-112k for format cold-start, then use GRPO to develop Interleave-Critic-RL-13k, strengthening step-wise instruction correction across generation trajectories. Since a single interleaved generation trajectory may involve more than 25 generator calls, optimizing the whole trajectory is computationally infeasible. They therefore propose an accuracy reward and a step-wise reward, enabling single-step RL to effectively guide the entire generation trajectory.

Results

InterleaveThinker improves the performance of various image generators on interleaved generation.
On interleaved generation benchmarks, it achieves performance comparable to Nano Banana and GPT-5.
Surprisingly, it also significantly enhances base models on reasoning benchmarks; for example, substantial gains on WISE and RISE with 4-step FLUX.2-klein.

---

*Auto-collected on 2026-06-14*