Claims

Manually verified numerical accuracy of generated visuals is lowest under pre-generation control (12.2%), intermediate under mid-generation control (59.83%), and highest under post-generation control (97.85%)

Claim · All claims
Evidence · 1 study (3 entries) · causal r2 · q3
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims

Reading the evidence codes

q is the design tier, i the impact, n the sample, and the kind with its r the rigour judged within that kind: what each code means.

This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.

q3 i? Manual verification of generated images found only 11 of 90 pre-gen images contained the correct number of objects (accuracy 12.2%). → Zhengxu Li 2026 q3 i? Manual verification found 70 of 117 mid-gen images contained the correct number of objects (accuracy 59.83%). → Zhengxu Li 2026 (2) q3 i? Manual verification found only 2 of 93 post-gen images were numerically incorrect (accuracy 97.85%). → Zhengxu Li 2026 (3)

Evidence

Zhengxu Li 2026

Zhengxu Li, Junling Wang, and April Yi Wang. (2026). When Should Teachers Control AI Generation for Mathematics Visuals? arXiv preprint. https://arxiv.org/abs/2605.10672 doi:10.48550/arxiv.2605.10672

q3 · i? · causal · r2

Manual verification of images generated in the 24-teacher within-subject study. The authors report that only 11 of 90 pre-gen images "contained the correct number of objects specified in the prompt (accuracy: 12.2%)". No effect size printed.

"Based on manual verification of the 90 images generated using pre-gen control across all participants, only 11 images contained the correct number of objects specified in the prompt (accuracy: 12.2%)."

Zhengxu Li 2026 (2)

Zhengxu Li, Junling Wang, and April Yi Wang. (2026). When Should Teachers Control AI Generation for Mathematics Visuals? arXiv preprint. https://arxiv.org/abs/2605.10672

q3 · i? · causal · r2

Manual verification of mid-gen images in the same study: 70 of 117 images "contained the correct number of objects (accuracy: 59.83%)". Fifteen participants appreciated that layouts helped encode structure but still verified final images. No effect size printed.

"Based on manual verification, 70 out of the 117 images generated using mid-gen control contained the correct number of objects (accuracy: 59.83%)."

Zhengxu Li 2026 (3)

Zhengxu Li, Junling Wang, and April Yi Wang. (2026). When Should Teachers Control AI Generation for Mathematics Visuals? arXiv preprint. https://arxiv.org/abs/2605.10672

q3 · i? · causal · r2

Manual verification of post-gen images in the same study: of 93 images, "only 2 were numerically incorrect (accuracy: 97.85%)". Post-gen visuals are deterministic SVG objects edited directly, so no further synthesis occurs. No effect size printed.

"In contrast, post-gen control exhibited substantially higher numerical accuracy: of the 93 images generated using post-gen control, only 2 were numerically incorrect (accuracy: 97.85%)."

Discussion

Page metadata
Field Value
type claim
id manual-verification-accuracy-by-control-timing
title Manually verified numerical accuracy of generated visuals is lowest under pre-generation control (12.2%), intermediate under mid-generation control (59.83%), and highest under post-generation control (97.85%)
description Manually verified numerical accuracy of generated visuals is lowest under pre-generation control (12.2%), intermediate under mid-generation control (59.83%), and highest under post-generation control (97.85%)
status review
evidence_strength moderate
newest_evidence 2026
recheck_by 2029
generated by process:wiki-ingest on 2026-10-11
trust unverified — no one has confirmed this page's content
sources 3 — zhengxu-li-2026, zhengxu-li-2026-2, zhengxu-li-2026-3

Frontmatter is the page's machine-readable record — see the Schema & Guide.