SearchGen-20K: Teaching Visual Generators to Search Beyond Their Knowledge Boundary

Forum topic · 小凯 · 2026-07-08

Summary

A new paper on arXiv (2607.05382) introduces SearchGen-20K and SearchGen-Bench, a benchmark of 20,839 prompts spanning 12 failure categories and 22 domains, designed to expose the world-knowledge bottleneck of visual generation models. While visual generators excel at rendering, they confidently hallucinate unknown content such as new characters, trending entities, and post-cutoff events. Frontier open generators score only 21-28 out of 100 on the benchmark, roughly 40 points lower than existing benchmarks. The authors argue that naive retrieval hurts performance by injecting noise into prompts the generator can already handle. The root cause is a generator-specific, evolving knowledge boundary separating what a model internalized during training from what must be retrieved externally. The paper proposes a teach-search collaborative training framework that can discover this boundary, showing that even minimal versions yield monotonic improvements and laying 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

Abstract

Visual generators are excellent at rendering, but they confidently fabricate unknown content. User requests are unbounded, constantly evolving, and long-tailed: new characters, trending entities, post-cutoff events, and more. This world-knowledge bottleneck is structural: generators are trained on a fixed corpus, while the visual world remains open.

This paper builds SearchGen-20K and SearchGen-Bench, containing 20,839 prompts spanning 12 failure categories and 22 domains. Frontier open generators score only 21-28 out of 100 — roughly 40 points below existing benchmarks.

Naive search retrieves indiscriminately, injecting noise into prompts the generator can already handle. The root cause is a generator-specific, evolving knowledge boundary: the dividing line between content the generator has internalized through training and content that must be kept in external context.

The paper demonstrates that this boundary can be discovered through a teach-search collaborative training framework; even a minimal version produces monotonic improvements, laying the groundwork for recursive self-improvement in visual generation.

Key Findings

New benchmark: SearchGen-20K / SearchGen-Bench with 20,839 prompts, 12 failure categories, 22 domains.
Frontier open generators achieve only 21-28/100, about 40 points below prior benchmarks.
Naive retrieval degrades performance by adding noise to already-mastered prompts.
A teach-search collaborative training framework can identify the generator's evolving knowledge boundary, enabling monotonic gains and recursive self-improvement.

--- *Auto-collected on 2026-07-06*