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

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Haozhe Wang, Weijia Feng, Jinpeng Yu, Che Liu, Ping Nie, Fangzhen Lin, Jiaming Liu, Ruihua Huang, Jimmy Lin, Wenhu Chen, Cong Wei
• Venue: arXiv: cs.CV
• Published: 2026-07-06
• Collected: 2026-07-07
• URL: http://arxiv.org/abs/2607.05382v1
• PDF: https://arxiv.org/pdf/2607.05382v1
• DOI: UNKNOWN

Plain-Language Brief

The researchers created two new datasets, SearchGen-20K and SearchGen-Bench, to evaluate the performance of current visual generators in handling a wide range of requests that require knowledge beyond what they were trained on. They found that these generators perform poorly, scoring only between 21 and 28 out of 100. The researchers propose using search tools to expand the knowledge base of these generators, but find that naive search fails due to noise injection. They introduce a teach-then-search co-training framework as a solution, which shows improvement in visual generation.

What the Document Says

• The text discusses the limitations of current visual generators in handling requests that require knowledge beyond their training.
• Two new datasets, SearchGen-20K and SearchGen-Bench, are created to evaluate these generators.
• On SearchGen-Bench, current open generators perform poorly, scoring between 21 and 28 out of 100.
• The researchers propose using search tools to expand the knowledge base of these generators, but find that naive search fails.
• A teach-then-search co-training framework is introduced as a solution for improving visual generation.
• The text mentions the release of the full dataset, co-training corpus, and search corpus.

Key Concepts Mentioned

• visual generators
• world-knowledge bottleneck
• SearchGen-20K
• SearchGen-Bench
• teach-then-search co-training framework

Claims Made by the Document

• Current visual generators perform poorly on requests that require knowledge beyond their training.
• A teach-then-search co-training framework improves visual generation.

Evidence Used

• The poor performance of current open generators on SearchGen-Bench is stated.
• Monotonic improvement in visual generation using the teach-then-search co-training framework is mentioned.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.