SAR-LM: Symbolic Audio Reasoning with Large Language Models

TL;DR

SAR-LM introduces a symbolic audio reasoning pipeline that converts audio into structured, human-readable features, enhancing interpretability and enabling transparent error analysis while maintaining competitive performance on multiple benchmarks.

Contribution

The paper presents SAR-LM, a novel symbolic audio reasoning approach that improves interpretability over existing dense embedding methods by using structured, human-readable features.

Findings

Achieves competitive results on MMAU, MMAR, and OmniBench benchmarks.

Enables transparent error analysis by tracing failures to specific features.

Prioritizes interpretability without sacrificing performance.

Abstract

Large language models (LLMs) have advanced in text and vision, but their reasoning on audio remains limited. Most existing methods rely on dense audio embeddings, which are difficult to interpret and often fail on structured reasoning tasks. Caption-based approaches, introduced in recent benchmarks such as MMAU, improve performance by translating audio into text, yet still depend on dense embeddings as input, offering little insight when models fail. We present SAR-LM, a symbolic audio reasoning pipeline that builds on this caption-based paradigm by converting audio into structured, human-readable features across speech, sound events, and music. These symbolic inputs support both reasoning and transparent error analysis, enabling us to trace failures to specific features. Across three benchmarks, MMAU, MMAR, and OmniBench, SAR-LM achieves competitive results, while prioritizing interpretability as its primary contribution.