Benchmarking and Confidence Evaluation of LALMs For Temporal Reasoning

TL;DR

This work targets temporal reasoning in audio by introducing the TREA dataset, which probes event duration, ordering, and counting through audio MCQA. It benchmarks open-source LALMs in zero-shot settings and reveals a sizable gap to human performance, with accuracy often below 50% on several tasks. Beyond accuracy, the paper proposes a test-time perturbation–based uncertainty measure and calibration analysis, showing that accuracy and uncertainty/calibration do not always align. The findings advocate for incorporating temporal reasoning benchmarks and uncertainty metrics into LALM development to support robust, trustworthy audio understanding in real-world tasks.

Abstract

The popular success of text-based large language models (LLM) has streamlined the attention of the multimodal community to combine other modalities like vision and audio along with text to achieve similar multimodal capabilities. In this quest, large audio language models (LALMs) have to be evaluated on reasoning related tasks which are different from traditional classification or generation tasks. Towards this goal, we propose a novel dataset called temporal reasoning evaluation of audio (TREA). We benchmark open-source LALMs and observe that they are consistently behind human capabilities on the tasks in the TREA dataset. While evaluating LALMs, we also propose an uncertainty metric, which computes the invariance of the model to semantically identical perturbations of the input. Our analysis shows that the accuracy and uncertainty metrics are not necessarily correlated and thus, points to a need for wholesome evaluation of LALMs for high-stakes applications.

Figures (2)

• Figure 1: Examples from TREA dataset illustrating the sub-tasks in event order, count and duration based MCQA.
• Figure 2: Examples of original sample (from order task), and two perturbed samples, one with audio modification and one with text modification.