Gen-Searcher: Reinforcing Agentic Search for Image Generation

Dimensions overall score 9.0

• Gen-Searcher is the first trained search-augmented image generation agent that performs multi-hop reasoning and search to collect textual knowledge and reference images for grounded generation.method 95%
  Evidencepartial

  we present Gen-Searcher, as the first attempt to train a search-augmented image generation agent, which performs multi-hop reasoning and search to collect the textual knowledge and reference images needed for grounded generation.

  Implicationpartial

  Explicitly stated in the abstract as a first attempt, with the method described in detail.

  Verificationpartial

  partial

• Gen-Searcher improves Qwen-Image by around 16 points on the KnowGen benchmark.result 90%
  Evidencepartial

  Experiments show that Gen-Searcher brings substantial gains, improving Qwen-Image by around 16 points on KnowGen and 15 points on WISE.

  Implicationpartial

  Direct numeric result stated in the abstract and supported by experimental data in the paper.

  Verificationpartial

  partial

• The system's quality heavily depends on the accuracy of retrieved web content, and incomplete or incorrect web data can lead to erroneous image outputs.limitation 85%
  Evidencepartial

  The system's quality heavily depends on the accuracy of retrieved web content. Incomplete or incorrect web data can lead to erroneous image outputs.

  Implicationpartial

  Explicitly stated as a caveat in the analysis section, indicating a known limitation.

  Verificationpartial

  partial

• Gen-Searcher uses agentic reinforcement learning with dual reward feedback (text-based and image-based) to provide stable and informative learning signals for GRPO training.method 80%
  Evidencepartial

  we train Gen-Searcher with SFT followed by agentic reinforcement learning with dual reward feedback, which combines text-based and image-based rewards to provide more stable and informative learning signals for GRPO training.

  Implicationpartial

  Described in the abstract and analysis as a core method, though specific implementation details are summarized.

  Verificationpartial

  partial

• The authors constructed two high-quality datasets, Gen-Searcher-SFT-10k and Gen-Searcher-RL-6k, containing diverse search-intensive prompts and corresponding ground-truth synthesis images.technical 90%
  Evidencepartial

  we construct a tailored data pipeline and curate two high-quality datasets, Gen-Searcher-SFT-10k and Gen-Searcher-RL-6k, containing diverse search-intensive prompts and corresponding ground-truth synthesis images.

  Implicationpartial

  Explicitly stated in the abstract as part of the data pipeline, with dataset names and purposes provided.

  Verificationpartial

  partial

• Standard text-to-image systems perform poorly on knowledge-intensive and search-grounded image generation, as shown by low K-Scores (around 9 to 15), while proprietary models perform substantially better.result 85%
  Evidencepartial

  around 9 to 15 K-Score, showing that knowledge-intensive and search-grounded image generation remains far beyond the capability of standard text-to-image systems. In contrast, proprietary models perform substantially better

  Implicationpartial

  Directly stated in the results section with comparative performance metrics.

  Verificationpartial

  partial

• Gen-Searcher is equipped with three search tools, including a web text search tool, to perform multi-turn reasoning and information gathering.technical 80%
  Evidencepartial

  Search Tools.Gen-Searcher is equipped with three search tools. The first is search, which performs web text search

  Implicationpartial

  Described in the method section and illustrated in an inference example, though tool specifics are briefly mentioned.

  Verificationpartial

  partial

• The model was trained using 8 NVIDIA H800 GPUs, with Qwen3-VL-8B-Instruct as the base model, following a two-stage process of supervised fine-tuning and agentic RL training.method 90%
  Evidencepartial

  We train Gen-Searcher-8B using 8 NVIDIA H800 GPUs, with Qwen3-VL-8B-Instruct as the base model. We first perform supervised fine-tuning on Gen-Searcher-SFT-10k, and then further conduct agentic RL training on Gen-Searcher-RL-6k.

  Implicationpartial

  Specific training details are provided in the experiments section, including hardware, base model, and training stages.

  Verificationpartial

  partial

Competitive landscape

Gen-Searcher leverages agentic reinforcement learning for search-augmented image generation, delivering contextually relevant and high-fidelity visual content.