Gen-Searcher: Reinforcing Agentic Search for Image Generation

summary

The gist

Gen-Searcher presents Gen-Searcher as the first attempt to train a search-augmented image generation agent that performs multi-hop reasoning and search to collect textual knowledge and reference…

In short

Gen-Searcher is a search-augmented image generation agent designed to perform multi-hop reasoning and web searches to gather knowledge for grounded image creation. It uses supervised fine-tuning followed by agentic reinforcement learning with dual reward feedback, achieving significant performance gains on benchmarks like KnowGen.

Key concepts

Multi-hop Reasoning
This is the agent's ability to solve complex problems by chaining together multiple search steps or pieces of information. Instead of answering a simple question, the agent searches for one piece of data, uses that result to formulate a new search query, and repeats this process until it gathers enough context to create an accurate image.
Dual Reward Feedback
The training uses two types of feedback: an image-based reward (K-Score) and a text-based reward (Rtext). The text reward checks if the collected search information is sufficient to generate the target image, helping stabilize training by supervising the quality of gathered knowledge alongside visual output quality.
Grounding Generation
This refers to creating an image based on specific, verifiable real-world facts retrieved from external sources. Gen-Searcher actively searches the web for necessary details—like specific titles or data points—and uses this factual evidence to ensure the final generated image accurately reflects those external truths.

Terminology used across episodes

This episode discusses

The paper

Gen-Searcher: Reinforcing Agentic Search for Image Generation · Read on arXiv

Kaituo Feng, Manyuan Zhang, †Shawn Chen, Yunlong Lin, Kaixuan Fan, №Yilei Jiang, 了Hongyu Li, Dian Zheng, 发Chenyang Wang

MMLab, CUHK 2UCLA 3UC Berkeley, Meituan

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Gen-Searcher: Reinforcing Agentic Search for Image Generation".

Jane: Here's the gist of the paper. Gen-Searcher presents Gen-Searcher as the first attempt to train a search-augmented image generation agent that performs multi-hop reasoning and search to collect textual knowledge and reference images needed for grounded generation This work introduces Gen-Searcher, which trains a multimodal deep search agent for image generation using SFT followed by agentic reinforcement learning with dual reward feedback The method consistently improves image generation performance across different backbones, bringing around 16-point gains for Qwen-Image on KnowGen and 15 points on WISE Pruning redundant reasoning tokens The paper addresses the fundamental limitation of existing text-to-image models, which are constrained by frozen internal knowledge and struggle with knowledge-intensive or up-to-date information Prior RAG-based approaches are limited by the coverage and freshness of static databases, while prompt-based workflows lack adaptive planning capabilities The core idea is to train a search agent that actively performs multi-hop web search and reasoning to gather knowledge for grounded image generation How the training signal is built The authors construct a dedicated data pipeline consisting of four stages: text prompt construction, agentic trajectory generation, grounded image synthesis, and data filtering and curation Text prompts are constructed using two strategies: carefully designed prompt engineering to instruct Gemini 3 Pro to generate multi-hop search-intensive prompts across diverse categories, and converting existing deep research datasets into image-generation-oriented prompts The agentic trajectory generation involves employing Gemini 3 Pro with search tools (search, image search, browse) in a multi-turn manner to perform search, browsing, and reasoning before producing a final search-grounded prompt and reference images Ground-truth images are synthesized using the proprietary image generation model Nano Banana Pro Where the gains come from The training scheme involves two stages: supervised fine-tuning on Gen-Searcher-SFT-10k to equip the model with basic tool-use abilities, and agentic RL based on GRPO to optimize tool-calling trajectories A key innovation is the dual reward feedback design, combining an image-based reward (K-Score) and a text-based reward (Rtext) to provide more stable and informative feedback for optimizing the search and reasoning process This dual approach addresses the noise from pure image rewards by supervising the quality of gathered information via Rtext, which evaluates whether the collected prompt contains sufficient information to generate the target image Benchmarks and Evaluation The work introduces KnowGen, a comprehensive benchmark designed to evaluate search-grounded image generation under knowledge-intensive real-world scenarios KnowGen explicitly focuses on knowledge-intensive and search-dependent generation where solving the prompt often requires retrieving and aggregating evidence from the web Evaluation is conducted using K-Score, a metric that combines four dimensions: faithfulness, visual correctness, text accuracy, and aesthetics The final K-Score is computed as a weighted combination of these four dimensions where the weighting emphasizes the two most critical aspects of search-grounded image generation Gen-Searcher brings substantial gains across different image generation backbones on both KnowGen and WISE, while also exhibiting strong transferability across image generators The method demonstrates that Gen-Searcher can substantially compensate for the lack of built-in search capability in open-source image generators by actively gathering grounded textual evidence and visual references from the web The results suggest that Gen-Searcher is a general and powerful model for enhancing image generation in real-world knowledge-intensive scenarios The authors fully open-source their data, models, and code to serve as an open foundation for future research on search agents for real-world image generation The final performance of the full Gen-Searcher model reaches the best result of 31.52 on KnowGen The dual reward design is effective, with performance remaining consistently strong when alpha is set in the range of 0.3 to 0.6 Page 1 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Kaituo Feng1,4 Manyuan Zhang1,4† Shawn Chen2 Yunlong Lin1 Kaixuan Fan1 Yilei Jiang1 Hongyu Li4 Dian Zheng1 Chenyang Wang3 Xiangyu Yue1‡ Home: https://gen-searcher.vercel.app/ Page 2 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation and informative learning signals for GRPO training Experiments show that Gen-Searcher brings substantial gains, improving Qwen-Image by around 16 points on KnowGen and 15 points on WISE We hope this work can serve as an open foundation for search agents in image generation, and we fully open-source our data, models, and code Page 3 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Ours Previous paradigm Image Generator Image Generator Gen-Searcher multi-hop deep search Incorrect text and visual features Correct text and visual features PROMPT In In garden the Wuthering Waves Resonator who is the first featured 5-star of version 1.4 is shown A small shows 'Base HP' at Level 90 and Intro Skill’ Intro Figure 2: Our proposed Gen-Searcher enables search-grounded generation in real-world knowledge-intensive scenarios Page 4 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation We conduct extensive experiments to evaluate the effectiveness of Gen-Searcher Our method consistently improves image generation performance across different backbones, bringing around 16-point gains for Qwen-Image on KnowGen We further observe strong transferability: a Gen-Searcher trained with Qwen-Image can be directly applied to Seedream 4.5

15: and Nano Banana Pro

2: without additional training, yielding about 16-point and 3-point improvements, respectively Beyond our proposed KnowGen benchmark, we also evaluate on WISE

16: , where Gen-Searcher improves Qwen-Image from 0.62 to 0.77, demonstrating its strong generalization in knowledge-intensive image generation tasks Page 5 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Figure 3: An illustration of our data curation pipeline Text Prompt Construction Gemini 3 Pro 20 Domain Categories DeepResearch QA Dataset Prompt Engineering Multi-hop Search Prompts for Gen (Diverse Scenarios) DeepResearch QA Dataset Prompt Engineering Multi-hop Search Prompts for Gen (Diverse Scenarios) Page 6 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Agentic Trajectory Generation Gemini 3 Pro Agentic Search multi-turn loop Multi-turn reasoning & information seeking and analysis Text Search from Web Image Search via Text Query Analyze Content in Webpage Agentic Trajectory Grounded Prompt Reference Images (visual features) Page 7 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Ground-Truth Image Synthesis Nano Banana Pro (Proprietary Image Generation Model) 30K Raw Samples GT images Figure 3: An illustration of our data curation pipeline Page 8 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Data Filtering & Benchmark Construction Data Filtering by Seed1.8 Scoring and Rules 17k High-quality Samples KnowGen Bench Gen-Searcher-SFT-10k Gen-Searcher-RL-6k Page 9 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Figure 4: Overview of the KnowGen benchmark Page 10 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Category Anime A live-action cosplay photo in a gothic cathedral features Black Swan from Honkai: Star Rail, the white-haired saintess of the Kjerag region from Arknights, and the Natlan scholar Citlali from Genshin Impact comparing religious artifacts on a stone altar an embossed plaque on the wall must correctly display the official religious title of the Kjerag saintess (must be correct) Category Chemistry An educational infographic comparing the heat of vaporization of the element Astatine and the halogen with the largest stable atomic number The chart must correctly show their symbols, their heat of vaporization values in kJ/mol, and the mass number of the most stable isotope of Astatine; all data must be correct Category Culture A close-up of a Bizen-yaki ceramic vase from Okayama, Japan, showing the characteristic Hidasuki firecord markings A small wooden sign shows the specific kiln temperature in Celsius used for the firing and the total number of days it remained in the kiln; all values must be correct Category Celebrities In a bright, glass-walled office, the researchers Akiko Iwasaki and Eric Topol are looking at a tablet showing an immunological study with the public health official Ashish Jha Iwasaki is pointing at a specific data point on the screen while the others take notes in shared journals Category Architecture A photograph of the 'Isbjerget' (The Iceberg) in Aarhus, Denmark; a resident's handbook on a balcony table shows the total number of apartments in the complex and the name of the French architecture firm that collaborated on the design, both of which must be correct Category Geography An educational poster illustrating the 'Bora' wind flow over the Velebit mountain range arrows indicate the cold air descending toward the Adriatic Sea, and a small data table shows the peak wind speed recorded at the Senj station during a specific historically significant winter event Page 11 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Category Anime A live-action cosplay photo in a gothic cathedral features Black Swan from Honkai: Star Rail, the white-haired saintess of the Kjerag region from Arknights, and the Natlan scholar Citlali from Genshin Impact comparing religious artifacts on a stone altar an embossed plaque on the wall must correctly display the official religious title of the Kjerag saintess (must be correct) Page 12 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Category Chemistry An educational infographic comparing the heat of vaporization of the element Astatine and the halogen with the largest stable atomic number The chart must correctly show their symbols, their heat of vaporization values in kJ/mol, and the mass number of the most stable isotope of Astatine; all data must be correct Page 13 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Category Culture A close-up of a Bizen-yaki ceramic vase from Okayama, Japan, showing the characteristic Hidasuki firecord markings A small wooden sign shows the specific kiln temperature in Celsius used for the firing and the total number of days it remained in the kiln; all values must be correct Page 14 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Category Celebrities In a bright, glass-walled office, the researchers Akiko Iwasaki and Eric Topol are looking at a tablet showing an immunological study with the public health official Ashish Jha Iwasaki is pointing at a specific data point on the screen while the others take notes in shared journals Page 15 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Category Architecture A photograph of the 'Isbjerget' (The Iceberg) in Aarhus, Denmark; a resident's handbook on a balcony table shows the total number of apartments in the complex and the name of the French architecture firm that collaborated on the design, both of which must be correct Page 16 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation Category Geography An educational poster illustrating the 'Bora' wind flow over the Velebit mountain range arrows indicate the cold air descending toward the Adriatic Sea, and a small data table shows the peak wind speed recorded at the Senj station during a specific historically significant winter event Page 17 --- Gen-Searcher: Reinforcing Agentic Search for Image Generation K-Score Evaluation Prompt You are a strict and professional expert evaluator for AI-generated image grounded with world knowledge (MODEL EVALUATION) You will receive 1) A task prompt (what the image must show).

Tom: 2) Image 1: the generated image (model output to be evaluated).

Jane: 3) Image 2: the ground-truth reference image (a strong reference implementation) All the input images are AI-generated.

Tom: All human in the images are AI-generated too.

Paper summary: Tom: So looking at the title "Gen-Searcher: Reinforcing Agentic Search for Image Generation," it really captures what they did here—it's not just about generating an image; it's about forcing the generation process to reinforce itself through active searching.

Jane: It’s about making the search part of the generation loop, which is a big shift from just using a static prompt. They built this entire system around multi-hop reasoning and searching to collect necessary knowledge and reference images before synthesizing anything.

Lu: What matters is that they didn't just stop at training an agent; they introduced that dual reward feedback mechanism with the K-Score and Rtext, which specifically targets the quality of the information being gathered. That’s what separates this from earlier attempts where you might just be optimizing for a final image score without caring about the search path.

Meng: From a deployment standpoint, if this works as advertised across different backbones like Qwen-Image or Seedream four point five, that means we have a more flexible agent framework that doesn't need custom training for every single model architecture we want to use for production <ref:2603.28767#pg3>. That’s useful information when you’re trying to ship something reliable.

Tom: And the numbers back up the claim; they saw about sixteen-point gains on KnowGen and fifteen points on WISE, which is substantial when you compare it against models that are frozen in their internal knowledge <ref:2603.28767#pg2,on KnowGen and 15 points on WISE>. They proved that this method can significantly improve generation performance by actively gathering external evidence.

Jane: It's definitely a model for enhancement in real-world knowledge-intensive scenarios because it shows the agent can actually compensate for what the base image generator is missing. The authors also mentioned that the dual reward design remains effective when alpha is set between zero point three and zero point six during optimization, which gives some stability to that process.

Lu: I think the most important thing they show is that we can train these agents to be robust not just on aesthetics but on verifiable facts found through search, and they’ve made the data pipeline open so others can study how this agentic trajectory works in practice.

Meng: Yeah, as a production engineer, I'm interested in the fact that they focused on keeping tool calls efficient; they capped the global tool-call budget at eight per item and limited calls per round to just one. That’s important because it shows they aren't just making the agent run wild with searches.

Tom: So in short, Gen-Searcher is their first attempt to create a search-augmented image generation agent that performs multi-hop reasoning and search to collect textual knowledge and reference images for grounded generation. It’s a general model for enhancing image generation when you need real-world context.

Jane: And the authors fully open the data, models, and code, which sets up an open foundation for future work on these types of search agents in image generation. It’s definitely something worth watching as a way to augment existing generative tools.

Conclusion: Tom: So we’re wrapping up this look at Gen-Searcher, which is about training an agent to search for information to make better images.

Jane: Right, so we're talking about how they built this whole system of multi-hop searching and reasoning just to ground the image generation in real knowledge.

Tom: Exactly. The authors are calling it a way to prune redundant reasoning tokens, which is basically trying to stop the agent from just rambling around without finding what it needs.

Jane: And they did this by constructing a data pipeline that goes from writing prompts, through the agent’s search and browsing steps, all the way to synthesizing that final image.

Tom: The core thing here is their training signal—they used supervised fine-tuning followed by agentic reinforcement learning with a dual reward feedback system.

Jane: That dual reward system, combining an image-based score and a text-based score, is what they say gives them more stable feedback for optimizing that whole search process.

Tom: And the results show substantial gains across different image backbones on benchmarks like KnowGen and WISE.

Jane: They saw about sixteen points on Qwen-Image on KnowGen and fifteen points on WISE, which shows how much it can boost performance compared to models that don't have built-in search.

Tom: It really demonstrates that this agent can actively gather external textual evidence and visual references from the web to fill in gaps.

Jane: And they’ve made the data, models, and code open so other researchers can look at how this agentic trajectory actually works.

Tom: This suggests Gen-Searcher is a general model for enhancing image generation in knowledge-intensive scenarios.

Jane: It opens up a new direction for how we think about augmenting generative tools with actual search capabilities.

More episodes

← Home