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

2026-07-06

AI Summary

Existing vision generation models are constrained by static training data and struggle to handle open-domain, dynamic, or long-tailed user requests—such as novel characters or trending events. This work proposes a “train-then-search” co-training framework that explicitly identifies the generator’s dynamic knowledge boundary as detectable yet not predefinable, and adaptively expands it through multimodal retrieval. To support this approach, we introduce the SearchGen-20K dataset and the SearchGen-Bench evaluation benchmark. Experiments reveal that state-of-the-art open-source models achieve only 21–28/100 on this benchmark in their original form, whereas our method substantially outperforms naive retrieval strategies and yields consistent performance gains, laying the foundation for recursively self-improving vision generation systems.

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.

Problem

• visual generation
• world knowledge
• knowledge boundary
• open-ended requests
• hallucination

Innovation

• agentic visual generation
• knowledge boundary
• co-training
• search-augmented generation
• long-tailed visual requests