Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation

Hugging Face Daily Papers Papers
visual-generation knowledge-boundary agentic-generation co-training long-tailed-requests benchmark search-tools

Summary

This paper addresses the knowledge boundary problem in visual generation by introducing the SearchGen-20K benchmark and SearchGen-Corpus-1M, and proposes a teach-then-search co-training framework to handle evolving, long-tailed user requests beyond a generator's training data.

Visual generators excel at rendering, but they confidently fabricate what they do not know. User requests are unbounded, evolving, and deeply long-tailed: new characters, trending entities, post-cutoff events, and more. This world-knowledge bottleneck is structural: generators are trained on fixed corpora, but the visual world is open-ended. We construct SearchGen-20K and SearchGen-Bench, with 20,839 prompts spanning twelve failure categories and twenty-two domains, paired with a pre-executed multimodal SearchGen-Corpus-1M to support offline, reproducible research. On SearchGen-Bench, frontier open generators score only 21 to 28 out of 100, a 40-point collapse invisible to existing benchmarks. The natural remedy is to employ search tools, enabling agentic visual generation. However, we find that naive search fails: it retrieves indiscriminately, injecting noise into prompts the generator already handles. We trace the root cause to a generator-specific, evolving knowledge boundary: the divide between what a generator can internalize through training and what must remain in external context. Although this boundary is hard to specify in advance, we show that it is discoverable through a teach-then-search co-training framework. Even a minimal version of this co-training recipe produces monotonic improvement, laying the foundation for recursive self-improvement in visual generation that can meet world-knowledge-grounded requests. We release the full dataset, co-training corpus, and search corpus as a replayable harness for tool-augmented, world-knowledge-grounded visual generation.
Original Article
View Cached Full Text

Cached at: 07/15/26, 04:22 PM

Paper page - Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation

Source: https://huggingface.co/papers/2607.05382 Authors:

,

,

,

,

,

,

,

,

,

Abstract

Visualgeneratorsexcelatrendering,buttheyconfidentlyfabricatewhattheydonotknow.Userrequestsareunbounded,evolving,anddeeplylong-tailed:newcharacters,trendingentities,post-cutoffevents,andmore.Thisworld-knowledgebottleneckisstructural:generatorsaretrainedonfixedcorpora,butthevisualworldisopen-ended.WeconstructSearchGen-20KandSearchGen-Bench,with20,839promptsspanningtwelvefailurecategoriesandtwenty-twodomains,pairedwithapre-executedmultimodalSearchGen-Corpus-1Mtosupportoffline,reproducibleresearch.OnSearchGen-Bench,frontieropengeneratorsscoreonly21to28outof100,a40-pointcollapseinvisibletoexistingbenchmarks.Thenaturalremedyistoemploysearchtools,enablingagenticvisualgeneration.However,wefindthatnaivesearchfails:itretrievesindiscriminately,injectingnoiseintopromptsthegeneratoralreadyhandles.Wetracetherootcausetoagenerator-specific,evolvingknowledgeboundary:thedividebetweenwhatageneratorcaninternalizethroughtrainingandwhatmustremaininexternalcontext.Althoughthisboundaryishardtospecifyinadvance,weshowthatitisdiscoverablethroughateach-then-searchco-trainingframework.Evenaminimalversionofthisco-trainingrecipeproducesmonotonicimprovement,layingthefoundationforrecursiveself-improvementinvisualgenerationthatcanmeetworld-knowledge-groundedrequests.Wereleasethefulldataset,co-trainingcorpus,andsearchcorpusasareplayableharnessfortool-augmented,world-knowledge-groundedvisualgeneration.

View arXiv pageView PDFProject pageGitHub6Add to collection

Get this paper in your agent:

hf papers read 2607\.05382

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2607.05382 in a model README.md to link it from this page.

Datasets citing this paper3

#### JasperHaozhe/SearchGen-20K Viewer• Updated3 days ago • 20.2k • 50 • 1 #### JasperHaozhe/SearchGen-Corpus-1M Viewer• Updated3 days ago • 1.74M • 32 #### JasperHaozhe/SearchGen-Bench Viewer• Updated3 days ago • 751 • 25

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2607.05382 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

Self-Evolving Visual Questioner

Hugging Face Daily Papers

This paper introduces a self-evolving framework for vision-language models to improve their question-generation capabilities without external supervision, enhancing both question quality and answerer performance.

EvoBrowseComp: Benchmarking Search Agents on Evolving Knowledge

arXiv cs.CL

This paper introduces EvoBrowseComp, a dynamic benchmark of 400 English and 400 Chinese complex questions that are synthesized via live-web traversal to evaluate search agents without test-set contamination, ensuring robustness against parametric memorization.