Beyond Classification: A Cough Regression Benchmark for Respiratory Acoustic Foundation Models

Researchers have introduced two new benchmarks to evaluate the performance of foundation models in respiratory acoustic and defeasible abduction tasks.

A new benchmark for respiratory acoustic foundation models evaluates the ability of these models to predict continuous health quantities from cough audio^[1]. The benchmark assesses five foundation models across six targets on three datasets. MLP-small outperformed the mean-predictor baseline on all tasks and linear probing in 23 of 30 model x task cases^[1]. HeAR led within-dataset age regression on Coswara with a mean absolute error of 9.12 years. Separately, a new dataset and generation pipeline, DeFAb, has been introduced for testing defeasible abduction in foundation models. DeFAb includes 372,648+ instances across 33.75M materialized rules from 18 sources^[2]. The best frontier language model achieved 65% accuracy in defeasible abduction, while a rule-based logic solver achieved 100% accuracy. Four frontier models did not reliably internalize defeasible reasoning, with rendering-robust Level 2 accuracy ranging from 7.8-23.5%^[2].