LLM-generated responses outperform self-augmentation when augmenting small educational classification datasets¶
Claim · All claims
Evidence · 1 study · causalr2·q2
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person
Design limits · Keith Cochran 2025: the intervention came bundled with another change (read from the study by two models, 2026-10-27)
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? Maximum classification performance was achieved with LLM-generated responses rather than self-augmentation (data replication) for all seven concepts, and self-augmented training was less stable at high augmentation levels. → Keith Cochran 2025
Evidence¶
Keith Cochran 2025¶
Keith Cochran, Clayton Cohn, Jean-François Rouet, Peter Hastings. (2025). Comparing Text Augmentation by GPT-3.5 and Llama3 for Evaluating Student Responses. International Journal of Artificial Intelligence in Education. https://doi.org/10.1007/s40593-025-00473-x
q2 · i? · causal · r2
Head-to-head comparison of six augmentation methods (self plus five LLM/temperature combinations) across four BERT-based classifiers on French and English student-response datasets, measured by micro-F1 averaged over ten seeds. Generated texts were "longer and more complicated" than copies.
"This hypothesis was supported by our results which showed that maximum performance was achieved using generated responses for all seven concepts. LLM-generated responses outperformed self-augmentation because the generated texts were longer and more complicated."
Discussion¶
Related Claims¶
- Fine-tuning BERT-based classifiers on LLM-augmented student-response data outperforms unaugmented training for all tested concepts — related
- Augmentation levels below 20x produce unreliable classification performance in small educational datasets — related
- Higher LLM temperature settings yield the best English augmentation performance, with Llama temperature 2 the top generator overall — related
- Datasets with entropy below 0.5 are difficult to augment enough to beat majority-label guessing — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
llm-generated-beats-self-augmentation |
title |
LLM-generated responses outperform self-augmentation when augmenting small educational classification datasets |
description |
LLM-generated responses outperform self-augmentation when augmenting small educational classification datasets |
status |
review |
evidence_strength |
moderate |
newest_evidence |
2025 |
generated |
by process:wiki-ingest on 2026-10-27 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — keith-cochran-2025 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.