# STEMMA: Song-to-Stem Multi-Audio Reasoning for Large Audio Language Models

arXiv: https://arxiv.org/abs/2610.11884 · HTML: https://mlsanity.com/paper/2610.11884/

## Key facts

- No citing papers in the map yet (data through 19 Oct 2026).
- Placed in Audio language models and spoken dialogue, one of 11 directions in Speech and audio language models.
- Builds on 22 of its 30 references that are in the map.
- First posted on arXiv 8 Oct 2026.

## Placed in

- Speech and audio language models / [Audio language models and spoken dialogue](https://mlsanity.com/speech/audio-language-models-and-spoken-dialogue/)

## Builds on (most cited first, 22 of 22)

- [Qwen2.5-Omni Technical Report](https://mlsanity.com/paper/2503.20215/) (2025)
- [Qwen2-Audio Technical Report](https://mlsanity.com/paper/2407.10759/) (2024)
- [MMAU: A Massive Multi-Task Audio Understanding and Reasoning Benchmark](https://mlsanity.com/paper/2410.19168/) (2024)
- [Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio Language Models](https://mlsanity.com/paper/2507.08128/) (2025)
- [FMA: A Dataset For Music Analysis](https://mlsanity.com/paper/1612.01840/) (2016)
- [MMAR: A Challenging Benchmark for Deep Reasoning in Speech, Audio, Music, and Their Mix](https://mlsanity.com/paper/2505.13032/) (2025)
- [MMAU-Pro: A Challenging and Comprehensive Benchmark for Holistic Evaluation of Audio General Intelligence](https://mlsanity.com/paper/2508.13992/) (2025)
- [Moisesdb: A dataset for source separation beyond 4-stems](https://mlsanity.com/paper/2307.15913/) (2023)
- [LLark: A Multimodal Instruction-Following Language Model for Music](https://mlsanity.com/paper/2310.07160/) (2023)
- [Music Flamingo: Scaling Music Understanding in Audio Language Models](https://mlsanity.com/paper/2511.10289/) (2025)
- [Are you really listening? Boosting Perceptual Awareness in Music-QA Benchmarks](https://mlsanity.com/paper/2504.00369/) (2025)
- [MOSS-Audio Technical Report](https://mlsanity.com/paper/2606.01802/) (2026)
- [SongFormer: Scaling Music Structure Analysis with Heterogeneous Supervision](https://mlsanity.com/paper/2510.02797/) (2025)
- [JamendoMaxCaps: A Large Scale Music-caption Dataset with Imputed Metadata](https://mlsanity.com/paper/2502.07461/) (2025)
- [SCNet: Sparse Compression Network for Music Source Separation](https://mlsanity.com/paper/2401.13276/) (2024)
- [ADIFF: Explaining audio difference using natural language](https://mlsanity.com/paper/2502.04476/) (2025)
- [Beyond Single-Audio: Advancing Multi-Audio Processing in Audio Large Language Models](https://mlsanity.com/paper/2409.18680/) (2024)
- [From Generation to Attribution: Music AI Agent Architectures for the Post-Streaming Era](https://mlsanity.com/paper/2510.20276/) (2025)
- [MUGEN: Evaluating and Improving Multi-audio Understanding of Large Audio-Language Models](https://mlsanity.com/paper/2603.09714/) (2026)
- [Automatic Music Sample Identification with Multi-Track Contrastive Learning](https://mlsanity.com/paper/2510.11507/) (2025)
- [Zero-shot Musical Stem Retrieval with Joint-Embedding Predictive Architectures](https://mlsanity.com/paper/2411.19806/) (2024)
- [Jamendo-MT-QA: A Benchmark for Multi-Track Comparative Music Question Answering](https://mlsanity.com/paper/2604.09721/) (2026)

## Abstract (arXiv)

Music understanding often requires comparing excerpts and reasoning about relationships among songs, sections, and stems. However, existing large audio-language models (LALMs) and music question-answering datasets typically operate on single recordings or compare independently sampled tracks with no known production relationship. We introduce STEMMA, a multi-audio music question-answering framework built around production provenance: whether excerpts originate from the same track or section, and which stems belong to which mixtures. Because such relations are sparse under conventional audio-first sampling, STEMMA adopts a relation-first construction strategy: it first specifies a target relation and then queries the catalog for excerpts that satisfy it and hard negatives that do not. Labels are determined directly from catalog provenance rather than generated by a language model from metadata. We build STEMMA-Bench for evaluation and a track-disjoint training set, STEMMA-Instruct. Fine-tuning two LALMs on STEMMA-Instruct improves multi-audio reasoning, with the largest gains on structural relations directly determined by the catalog, while preserving single-audio music understanding.