The Year in AI Papers
The year/Independent research

Paper 2607.05382

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

Haozhe Wang, Weijia Feng, Jinpeng Yu, Che Liu, Ping Nie, Fangzhen Lin, Jiaming Liu, Ruihua Huang, Jimmy Lin, Wenhu Chen, Cong Wei

Published
Jul 2026
Research lab
Independent
Citations
2
GitHub
115 stars

01 In brief

Summary

Visual generators fail on requests requiring world knowledge beyond their training data, such as new characters or recent events.

The authors introduce SEARCHGEN-20K, a dataset of 20,939 prompts with twelve failure categories, and SEARCHGEN-BENCH, where open generators score only 21–28 out of 100, a 40-point drop from standard benchmarks.

Naive search retrieval degrades performance by injecting noise.

The paper proposes a co-training framework: a noise-resistant agentic reasoner (gate-filter-integrate) controls when to search and how to integrate results, while online DPO teaches the generator to internalize stable knowledge and build noise robustness, followed by rejection finetuning to recalibrate the reasoner to the generator's shifted knowledge boundary.

This minimal recipe yields monotonic improvements, with an 8B reasoner matching a frontier oracle on a 4B generator.

The work releases the dataset, reasoning traces, and a cached search corpus for offline research.

02 From the paper

Abstract

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.