Back-translation evaluation outperforms LLM-as-a-Judge (code+image) in agreement with human raters across four models¶
Claim · All claims
Evidence · 1 study · designr2·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 achieves higher agreement with human ratings (Cohen's kappa 0.48-0.56) than LLM-as-a-Judge in its strongest setting (kappa 0.39-0.47) across four models. → Vishal Kumar 2025
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
Comparison on a 386-diagram test set of teacher-generated TikZ geometric figures, reported in Table 2: back-translation reached kappa 0.483-0.563 versus 0.388-0.498 for LLM-as-a-Judge with code+image, showing "comparable agreement with human raters" was exceeded by back-translation on every model.
"We find that back-translation outperforms LLM-as-a-Judge in its strongest setting (where it uses both code and image input), demonstrating comparable agreement with human raters"
Discussion¶
Related Claims¶
- Back-translation yields higher human agreement on spatial checks, while LLM-as-a-Judge considerably outperforms it on the angle-label check — related
- An LLM-based AI Evaluator agrees with expert human raters on collaboration transcripts at a level similar to inter-expert agreement — related
- A fine-tuned BERT model achieves almost perfect agreement with human coders when classifying student self-affirmation essays, matching human-human reliability — related
- GPT-4.1-Mini with back-translation matches the best LLM-judge (GPT-5) at 10.3x lower evaluation cost — a narrower finding that bears on this claim
- The cross-family LLM judge panel agrees with human domain experts on every judged axis, reaching the expert ceiling on answer-holding — related
- An open-source Llama3 model reached substantial agreement with a human coder when sentiment-labeling pre-extracted student reflection quotes — related
- The LLM-judge scoring pipeline for the two metrics aligns with human labels and human raters — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
backtranslation-outperforms-llm-judge-diagram-agreement |
title |
Back-translation evaluation outperforms LLM-as-a-Judge (code+image) in agreement with human raters across four models |
description |
Back-translation evaluation outperforms LLM-as-a-Judge (code+image) in agreement with human raters across four models |
status |
review |
evidence_strength |
moderate |
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 |
1 — vishal-kumar-2025 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.