Beyond Retrieval: Training Agents to Reason for Visual Creation

Modern text-to-image models produce photorealistic outputs, yet they remain fundamentally constrained by frozen parametric knowledge acquired during pretraining. When prompted to depict recent public figures, newly released products, or specific scientific phenomena, these systems hallucinate or default to generic representations because their training data cutoff prevents access to evolving real-world facts. While retrieval-augmented generation offers a partial solution, static databases lack the freshness and coverage necessary for the dynamic visual world, and shallow similarity-based retrieval fails on complex queries requiring sequential discovery across multiple sources.

In Gen-Searcher: Reinforcing Agentic Search for Image Generation, Feng et al. propose a fundamental shift from passive retrieval to active agentic search. Rather than treating external knowledge as a simple lookup table, they train a multimodal agent to perform multi-hop web searches, iterative browsing, and reasoning to construct grounded prompts and gather visual references before generation. This approach moves beyond the limitations of both frozen models and brittle prompt-based workflows, establishing a trained system that optimizes its own search trajectories for visual creation.

From Static Databases to Active Search Pipelines

Traditional RAG methods for image generation rely on single-round retrieval from static knowledge bases. These approaches fail on complex queries requiring multi-hop reasoning, where the necessary information spans multiple sources or requires sequential refinement. For instance, generating an image of "the Wuthering Waves Resonator who is the first featured 5-star of version 1.4" requires identifying the character through version history, then retrieving specific visual attributes and statistics. No single database query suffices, and manually designed search workflows lack the adaptability to handle such compositional knowledge requirements.

The authors address this by constructing Gen-Searcher, an agent trained to use search tools actively through a carefully designed data pipeline. Using Gemini 3 Pro, they generated approximately 10,000 search-intensive prompts across 20 diverse categories including celebrities, anime, physics, chemistry, and art. They then employed an iterative trajectory generation process where the model performs searches, browses results, and reasons over evidence to construct detailed, fact-grounded prompts accompanied by relevant reference images. Nano Banana Pro generated the ground truth synthesis images, with Seed1.8 filtering for quality assurance.

This process yielded two distinct datasets: Gen-Searcher-SFT-10k for supervised fine-tuning and Gen-Searcher-RL-6k for reinforcement learning. They also introduced KnowGen, a comprehensive benchmark with a K-Score metric specifically designed to evaluate search-grounded image generation on knowledge-intensive prompts. Unlike standard benchmarks that test aesthetic quality or prompt adherence alone, KnowGen explicitly requires models to demonstrate factual correctness obtained through external search, creating a more rigorous evaluation environment for agentic systems.

Stabilizing Multimodal RL with Dual Rewards

Training agents for image generation presents unique reinforcement learning challenges that differ significantly from text-based tasks. Unlike question answering, where reward signals can be derived from verifiable facts, image generation involves high variance outputs where even factually correct information might produce poor visual results due to the limitations of open-source generators such as Qwen-Image. Pure image-based rewards become noisy and unstable, offering insufficient signal for training the search policy effectively.

The authors solve this through agentic reinforcement learning with dual reward feedback. After an initial supervised fine-tuning stage to establish basic tool-use capabilities, they apply Group Relative Policy Optimization (GRPO) with two distinct reward components. The first is a text-based reward that evaluates whether the collected search-grounded prompt contains sufficient factual information and detail to generate the target image. The second is an image-based reward assessing the final visual output quality against the ground truth. This combination provides stable learning signals; the text reward guides the search and reasoning process, while the image reward ensures the gathered information actually