SearchGen: Evolving the Knowledge Boundary in Visual Generation with Search-Co-trained Framework

Forum topic · 小凯 · 2026-07-08

Summary

A new paper (arXiv:2607.05382, July 2026) addresses the world-knowledge bottleneck in visual generation: generators render well but confidently fabricate unknown content when users request novel characters, trending entities, or post-cutoff events. The authors identify the root cause as a generator-specific, evolving knowledge boundary — the line between what the model internalized during training and what must remain in external context. They introduce SearchGen-20K and SearchGen-Bench, comprising 20,839 prompts spanning 12 failure categories and 22 domains. Frontier open-ended generators score only 21-28 out of 100, roughly 40 points below existing benchmarks. Naive retrieval-augmented search injects noise into prompts the generator already handles, since it retrieves indiscriminately. The paper proposes a teach-search co-training framework that can discover this boundary, and shows even a minimal version yields monotonic improvements, establishing a foundation for recursive self-improvement in visual generation.

Paper Overview

Field: Computer Vision (CV)
Authors: Haozhe Wang, Weijia Feng, Jinpeng Yu, Che Liu, Ping Nie, Fangzhen Lin, Jiaming Liu, Ruihua Huang, Jimmy Lin, Wenhu Chen, Cong Wei
Published: 2026-07-06
arXiv: 2607.05382

Problem

Visual generators excel at rendering but confidently hallucinate unknown content. User requests are open-ended, constantly evolving, and long-tailed: new characters, trending entities, and events beyond the model's knowledge cutoff. This world-knowledge bottleneck is structural — generators are trained on a fixed corpus, while the visual world keeps growing.

Benchmark and Findings

The authors build SearchGen-20K and SearchGen-Bench, containing 20,839 prompts across 12 failure categories and 22 domains. Key results:

Frontier open-ended generators score only 21-28 / 100, roughly 40 points lower than on existing benchmarks.
Naive search augmentation retrieves indiscriminately, injecting noise even into prompts the generator can already handle.

Knowledge Boundary Hypothesis

The root cause is a generator-specific, evolving knowledge boundary: the dividing line between content internalized during training and content that must be kept in external context.

Method and Outlook

The paper shows this boundary can be discovered via a teach-search co-training framework. Even a minimal version of the framework produces monotonic improvements, laying the groundwork for recursive self-improvement in visual generation.

---

*Auto-collected on 2026-07-06*