Evaluating acoustic language models through the lens of Bloom's taxonomy
INTER-NOISE and NOISE-CON Congress and Conference Proceedings · 20.07.2026 · 273 (2) · 663-669
DOI: 10.3397/nc_2026_0108 ↗Kurz gefasst
Fachartikel in der Zeitschrift INTER-NOISE and NOISE-CON Congress and Conference Proceedings, erschienen am 20.07.2026. Verfasst von Jennifer Celeste und 1 weiteren Autor:innen.
Abstract (Originalsprache)
The efficacy of Large Language Models (LLMs) in specialized technical fields such as acoustics remains under-evaluated due to a lack of domain-specific benchmarks. General-domain research suggests that many LLM benchmarks lack higher-order cognitive tasks and that training data emphasizing higher-order questions may improve performance across Bloom's levels. This study investigates whether these findings extend to the development of LLM-based assistants for acoustics education. Using a two-model synthetic data generation pipeline, we generated 1,800 synthetic question-and-answer pairs from a standard acoustics textbook, evenly distributed across the six levels of Bloom's Taxonomy. These pairs were split into training and benchmark subsets, providing both fine-tuning data and a preliminary Bloom's-aligned evaluation set. We fine-tuned Llama-2 7B using Low-Rank Adaptation (LoRA) on different distributions of these Bloom's levels to determine whether prioritizing higher-order data produces a more capable acoustics-focused model than lower-order or balanced datasets. Results show that domain-specific fine-tuning substantially improved factual recall but did not improve, and often slightly reduced, performance on higher-order Bloom's categories relative to the base model. Among models trained on different Bloom's-level distributions, the mixed-order model achieved the highest accuracy on four of the six Bloom's categories, while the lower-order model achieved the highest overall benchmark accuracy; training exclusively on higher-order data did not improve higher-order generalization. This work introduces an initial framework for Bloom's-aligned evaluation of AI in acoustics education and suggests that for compact models fine-tuned on small synthetic datasets, the primary effect of domain-specifictraining is to strengthen recall rather than higher-order generalization.
Zitieren
APA
Celeste, J., & Sharma, B. (2026). Evaluating acoustic language models through the lens of Bloom's taxonomy. INTER-NOISE and NOISE-CON Congress and Conference Proceedings, 273(2), 663-669. https://doi.org/10.3397/nc_2026_0108
Vancouver
Celeste J, Sharma B. Evaluating acoustic language models through the lens of Bloom's taxonomy. INTER-NOISE and NOISE-CON Congress and Conference Proceedings. 2026;273(2):663-669. doi:10.3397/nc_2026_0108
BibTeX
@article{Celeste2026,
author = {Celeste, Jennifer and Sharma, Bhisham},
title = {Evaluating acoustic language models through the lens of Bloom's taxonomy},
journal = {INTER-NOISE and NOISE-CON Congress and Conference Proceedings},
year = {2026},
volume = {273},
number = {2},
pages = {663--669},
doi = {10.3397/nc_2026_0108}
}
EUVIMED