Summary
This post summarizes the paper "Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in A..." (arXiv:2607.05382), in the computer vision field by Haozhe Wang, Weijia Feng, Cong Wei, Wenhu Chen, and colleagues. Visual generators render convincingly but confidently fabricate unknown content, since user requests are unbounded, continuously evolving, and long-tailed: new characters, trending entities, and post-cutoff events. The paper identifies this world-knowledge bottleneck as structural: generators are trained on a fixed corpus while the visual world is open. The authors build SearchGen-20K and SearchGen-Bench, comprising 20,839 prompts across 12 failure categories and 22 domains, where frontier open generators score only 21–28 out of 100, roughly 40 points below existing benchmarks. Naive search indiscriminately retrieves content, injecting noise into prompts the generator already handles. The root cause is a generator-specific, evolving knowledge boundary between what is internalized via training and what must remain in external context. The paper shows this boundary can be discovered via a teach-search collaborative training framework, with even minimal versions yielding monotonic improvement, laying groundwork for recursive self-improvement in visual generation.
Paper Overview
Field: CV (Computer Vision)
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
Summary
Visual generators excel at rendering, yet confidently fabricate content they do not know. User requests are unbounded, continuously evolving, and long-tailed: new characters, trending entities, and events beyond a training cutoff. This world-knowledge bottleneck is structural rather than incidental: generators are trained on a fixed corpus, while the visual world is open-ended.
Benchmark: SearchGen-20K / SearchGen-Bench
- 20,839 prompts
- 12 failure categories
- 22 domains
- Frontier open generators score only 21–28 out of 100, roughly 40 points below existing benchmarks
Key Findings
- Naive search fails: indiscriminate retrieval injects noise into prompts the generator can already handle well.
- Root cause: a generator-specific, continuously evolving knowledge boundary — the dividing line between content internalized through training and content that must be kept in external context.
- Solution direction: the authors demonstrate that this boundary can be discovered through a teach–search collaborative training framework. Even a minimal version of the framework produces monotonic improvements.
- This establishes a foundation for recursive self-improvement in visual generation.
---
*Auto-collected on 2026-07-06. Original forum post at zhichai.net.*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178346219