Gold text descriptions improve VLM correctness-and-error assessment, but performance still lags other question types¶
Claim · All claims
Evidence · 1 study · causalr2·q2
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.
q2 i? Appending gold natural language descriptions improves VLM performance on correctness & errors QA, yet captioned performance remains below captionless performance on all other question categories for the same images. → Li Lucy 2026
Evidence¶
Li Lucy 2026¶
Li Lucy, Albert Zhang, Nathan Anderson, Ryan Knight, Kyle Lo. (2026). The Aftermath of DrawEduMath: Vision Language Models Underperform with Struggling Students and Misdiagnose Errors. https://arxiv.org/abs/2603.00925 doi:10.48550/arxiv.2603.00925
q2 · i? · causal · r2
Experiment re-evaluating four representative VLMs on correctness & errors QA with gold teacher captions appended, after removing 262 of 2,030 images whose captions contained correctness content; a caption-then-answer setup with model-generated captions approached but did not match gold captions (Figure 7). No effect size printed.
"Results are in the expected direction, in that VLM performance on correctness & errors QA improves with natural language support (Figure 7). However, this improved performance on correctness & errors QA with captions still lags behind VLMs' captionless performance in all other question categories for the same set of images."
Discussion¶
Related Claims¶
- Some VLMs perform near chance on binary judgments of student correctness, and error-assessment behavior is idiosyncratic across models — related
- The performance gap between erroneous and non-erroneous student images persists after images are digitally redrawn to remove noise — related
- VLMs achieve higher accuracy on non-erroneous than erroneous student math responses even when controlling for the math problem — related
- Gemini frequently corrects board-written math errors with correction logs, but still shows case-consistency and notation-substitution failures — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
captions-improve-correctness-qa-but-lag |
title |
Gold text descriptions improve VLM correctness-and-error assessment, but performance still lags other question types |
description |
Gold text descriptions improve VLM correctness-and-error assessment, but performance still lags other question types |
status |
review |
evidence_strength |
moderate |
newest_evidence |
2026 |
recheck_by |
2029 |
generated |
by process:wiki-ingest on 2026-10-10 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — li-lucy-2026 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.