Claims

Back-translation yields higher human agreement on spatial checks, while LLM-as-a-Judge considerably outperforms it on the angle-label check

Claim · All claims
Evidence · 1 study (2 entries) · design r2 · q2
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person · recheck by 2028: 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? Back-translation performs better on both spatial checks even when the judge is given the image, but the judge considerably outperforms back-translation on the angle-labels check. → Vishal Kumar 2025 q2 i? LLM-as-a-Judge considerably outperforms back-translation for the mathematical check about angle labels (0.829 vs. 0.652). → Vishal Kumar 2025 (2)

Evidence

Vishal Kumar 2025

Vishal Kumar, Shubhra Mishra, Rebecca Hao, Rizwaan Malik, David Broman, Dorottya Demszky. (2025). DiagramIR: An Automatic Pipeline for Educational Math Diagram Evaluation. NeurIPS 2025 Workshop: MATH-AI. https://arxiv.org/abs/2511.08283 doi:10.48550/arxiv.2511.08283

q2 · i? · design · r2

Check-wise Cohen's kappa against human ratings (Table 3) on the 386-diagram test set; back-translation reached e.g. 0.604 vs 0.390 for diagram-fully-in-frame (GPT-5), showing back-translation "performs better for both spatial checks".

"back-translation performs better for both spatial checks, despite the LLM-as-a-judge being provided the image to judge with"

Vishal Kumar 2025 (2)

Vishal Kumar, Shubhra Mishra, Rebecca Hao, Rizwaan Malik, David Broman, Dorottya Demszky. (2025). DiagramIR: An Automatic Pipeline for Educational Math Diagram Evaluation. NeurIPS 2025 Workshop: MATH-AI. https://arxiv.org/abs/2511.08283

q2 · i? · design · r2

Table 3 results for the labeled-angles check: the judge with image+code reached kappa 0.829 (GPT-5 mini) versus 0.652 for back-translation, an outcome the article calls considerably better judging.

"LLM-as-a-judge considerably outperforms backtranslation for the mathematical check about angle labels (0.829 vs. 0.652)"

Discussion

Page metadata
Field Value
type claim
id checkwise-backtranslation-spatial-strong-angle-weak
title Back-translation yields higher human agreement on spatial checks, while LLM-as-a-Judge considerably outperforms it on the angle-label check
description Back-translation yields higher human agreement on spatial checks, while LLM-as-a-Judge considerably outperforms it on the angle-label check
status review
evidence_strength mixed
newest_evidence 2025
recheck_by 2028
generated by process:wiki-ingest on 2026-10-10
trust unverified — no one has confirmed this page's content
sources 2 — vishal-kumar-2025, vishal-kumar-2025-2

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