arXiv subjects

Mateusz Modrzejewski

Publications and source records attributed to Mateusz Modrzejewski.

At least 19 recordsLinked to original sources

arXiv2026

How Python Audio Libraries Change Across Releases: Module Structure and Documentation

Developers who upgrade a Python audio library need to know what changed in the code they import and whether the release notes say so. Release notes and breaking changes have been mined for general software, but the module structure of audio libraries and its documentation have not been studied together. We mine 56 audio repositories used in machine learning workflows: the import graph of 49 packages at 1,395 releases, joined to 20,864 release-note and changelog entries, labeled by three open-weight language models and checked against two human coders. We find that (1) the set of a package's top-level modules is unchanged in 86.1% of consecutive release pairs, while the set of module files is unchanged in 61.4%. (2) A major increment changes the top level 5.2 times as often as another increment, and major and 0.x minor increments together account for 48.7% of the observed changes. (3) When a release removes modules, its notes name 8.9% of the removal units under a lexical match, an estimated 4.3% after a hand check of the matches. (4) Longer notes are more likely to contain an entry labeled as structural, and within each length band this proportion is also higher for releases with module removals. We intend to release the aligned dataset and the pipeline

cs.SE↗

arXiv2026

Poolwise: Evaluating Reward Models for Selection across Prompts

Reward models are used to select training data, but common benchmarks compare responses only within the same prompt. Selecting data across prompts also depends on how scores compare between prompts. We introduce Poolwise, a protocol that selects a fixed fraction of a response pool by score, measures the selected set's quality using external labels, and reports within-prompt ranking separately. In matched experiments on two rating sets and two model sizes, training each prompt's mean score to match its average response rating improves selection over permuted or omitted mean targets in every seed at 30 percent retention. In grouped HelpSteer experiments, the model with the highest within-prompt accuracy selects a lower-rated set than the correct-mean model. Subtracting prompt means estimated from separate reference responses preserves local rankings but lowers selected helpfulness for six open reward models and two closed judges on HelpSteer. Poolwise provides a direct way to compare reward models and score transformations for data selection.

cs.CL↗

arXiv2026

A Tutorial on Generative Models for Music Audio

Music audio generation spans instrumental clips, full songs with vocals, edits of existing recordings, and systems that play alongside a musician in real time. This monograph explains the methods behind these systems for readers familiar with neural-network training who are new to music generation. It introduces neural audio codecs and autoencoders, autoregressive and masked token models, diffusion and flow matching, sampling acceleration, and preference optimization in a shared notation. Worked examples connect training objectives and sampling procedures to published systems. The application sections relate the musical requirements of instrumental generation, songs, editing, and live performance to model design. Comparison tables record representations, methods, inputs, tasks, and release information documented in papers and official sources. Symbolic composition, speech, and sound effects are included where they connect to music audio. We explain how to interpret automatic metrics and learned judges, design and report listening tests, and assess evidence about data provenance, memorization, attribution, and output detection. Selected legal decisions and disputes illustrate the implications for training and release. The reviewed systems combine established generative paradigms with new representations, planning stages, and post-training procedures. Differences in data, scale, and evaluation limit what cross-system comparisons can establish about the sources of reported gains.

eess.AS↗

arXiv2026

Finite-Sample Sensitivity of Kernel Metrics for Generative Evaluation

Kernel distances such as the Kernel Inception Distance (KID) and the Kernel Audio Distance (KAD) compare generated and reference samples in an embedding space, and model comparisons rest on their values. Each ships with one fixed kernel, and a reported value does not say which changes that kernel detects at the sample size used. We run each default as a permutation two-sample test on real embeddings of symbolic music, audio and images, comparing a sample with an independent sample changed by a mean shift, an isotropic contraction, or a rotation that keeps the mean and the covariance trace and is matched to the shift in transport distance. At 780 samples per side, both defaults rarely detect the rotations in five evaluation spaces, while they detect the shifts in every run. The median-bandwidth Gaussian kernel of KAD detects the contraction and regains power against rotations on image embeddings with an order of magnitude more samples. The cubic kernel of KID contains second- and third-moment information but misses the contraction in most runs, and our experiments do not establish a high-power rotation regime for it over the tested sample sizes. Narrower Gaussian bandwidths restore sensitivity in music and audio, an $\ell_1$ Laplace kernel does so for images, and aggregated tests combine both at nominal level. We recommend that a reported kernel distance state its sample budget and the changes its configuration detects at that budget.

cs.LG↗

arXiv2026

MI-MIDI 2: From Observation to Control with Sparse Autoencoders in Text-to-MIDI Generation

Sparse autoencoders (SAEs) can expose features of a generative model without predefined labels, but in language models the evidence used to interpret these features has proven an unreliable guide to which of them steer the output. Whether interpretations of SAE features predict control in music generators has not, to our knowledge, been tested. We develop an end-to-end protocol for discovering and evaluating SAE features in text-to-MIDI generation and apply it to text2midi and MIDI-LLM. The protocol characterizes features through correlations with musical metrics, shared instrument programs, and automated descriptions, each with its own control: note-density adjustment, matched inactive excerpts, and permutation calibration. It then tests candidates by intervention, measured directly in the generated MIDI against unmodified generation and a matched random direction, ranks them in screens, and confirms selected features on new prompts. Evidence that survives these controls points to features that steer the output. Descriptions exceed permutation controls for instrumentation, and a feature whose description names a synth-lead program introduces that program in a prespecified confirmation and in a replication on disjoint prompts and reappears in independently trained dictionaries. Features selected by density-adjusted correlations show a small positive shift of their target metric as a family, whereas features selected by raw correlations show none. An atlas provides MIDI, piano-roll, and audio examples for every feature of the targeted screen.

cs.SD↗

arXiv2026

Prompt Sensitivity in Tool-Augmented Model Comparisons

Language models are compared on question-answering benchmarks that pose each question in one wording, and a reworded question can change which of two models scores higher. We ask whether giving the models tools makes such comparisons more stable. An accuracy gain cannot answer this, because tools change several steps between the question and the scored answer at once. We propose a matched protocol that separates those steps: the answer readout, reasoning, the decision to call, the tool's answer, and the report supplied without a call. We apply it to six open models in five domains under five wordings. We find that tools raise accuracy where a program answers the question and widen the wording sensitivity of many pairwise comparisons. The widening follows how often a model calls, and call frequency follows the wording for some models and not others. It also depends on the interface and the instructions: a native tool-calling interface and a second tool description leave no detected widening in molecules, and sensitivity persists where tools return intermediate results. These findings show that deterministic reports do not guarantee stable comparisons. We recommend reporting call frequency and comparison stability under several wordings beside accuracy.

cs.CL↗

arXiv2026

What Do Audio-Language Judges of Generated Music Listen For? Four Probes with Controls from Expert Ratings

Open audio-language models are used as judges of generated music: given a clip and its text prompt, the model rates the clip, and the judge is validated by how well its ratings correlate with expert ratings. The correlation does not show what the judge responds to: the music, the text prompt, or the way the question is asked. We propose four probes that test each of these, three of them read against a control computed from the human ratings: removing the audio, replacing the clip with another model's clip for the same prompt, swapping the order of two clips in pairwise judging, and rewording the instruction. We apply them to five open judges on all 2,748 clips of MusicEval, and the last two to the SongEval songs. Without audio, no judge exceeds what the prompt alone predicts from the human ratings, and for prompt alignment every judge falls below it. With audio, every judge follows the clip it hears when rating prompt alignment, two do when rating musical quality, and the strongest judge's quality score tracks the experts' alignment rating more than their quality rating. The pairwise verdicts of the four smaller judges follow the position of a clip in the pair (its slot) rather than its quality on both rating sets. For musical quality, rewording the instruction moves the correlation of most judges as much as changing the judge. MusicEval's prompt assignment bounds any prompt-only judge at the model level. We recommend reporting the four probes with every music judge.

cs.SD↗

arXiv2026

TADA! Tuning Audio Diffusion Models through Activation Steering

Audio diffusion models can synthesize high-fidelity music from text, yet achieving fine-grained control over specific musical attributes remains challenging, as their internal mechanisms for representing high-level concepts are poorly understood. In this work, we use activation patching to demonstrate that recent audio diffusion architectures exhibit a semantic bottleneck, where a small, shared subset of consecutive attention layers controls distinct musical concepts, such as the presence of specific instruments, vocals, or genres. Building on this, we systematically evaluate a broad spectrum of steering paradigms, comparing activation steering against prompt-level, score-space, and weight-space interventions, analyzing the interaction between the steering mechanism and the intervention site. Our new benchmark, supported by an extensive user study, demonstrates that localized activation steering establishes a new state-of-the-art in audio concept modulation.

cs.SD↗

arXiv2026

DINO-A: Adapting Self-Distillation Vision Transformers to General Audio Representation Learning

We present DINO-A, an adaptation of self-distillation from vision to general audio representation learning. While DINO has become a canonical method in self-supervised vision and prior audio work has explored latent prediction (BYOL-A) and masked modeling (Audio-MAE, BEATs), no prior work has brought canonical DINO to general audio classification in the way BYOL-A brought BYOL. DINO-A retains DINO's multi-crop, EMA teacher, and high-dimensional projection, replacing only the input modality and augmentations with log-mel spectrograms and the BYOL-A v2 augmentation block. We pretrain three backbones, two Vision Transformers with 8x8 and 16x16 patches and a convolutional encoder, on FSD50K and evaluate them with linear probing on ESC-50, Speech Commands v2, UrbanSound8K, and GTZAN. Three findings characterize the resulting representations. Patch resolution within the Vision Transformer family has consistent effect on representation quality, with smaller patches winning across all four tasks. The choice between Vision Transformer and convolutional backbone interacts with task type: convolutional networks lead on speech while Vision Transformers lead on environmental sounds and music. Under identical pretraining and evaluation conditions, DINO-A and BYOL-A v2 differ by 11.96 percentage points on average, and we trace this difference to two mechanisms: the interaction between DINO's high-dimensional projection space and FSD50K's limited scale, and the additional cost of multi-crop augmentation, which DINO uses but BYOL-A v2 does not. The high-dimensional projection space, central to DINO's success in vision, becomes a liability at FSD50K scale.

cs.SD↗

arXiv2026

MI-MIDI: Mechanistic Interpretability of Text-to-MIDI Generation Models via Probing, Lenses and Steering

Mechanistic interpretability of music generation has concentrated on audio models, leaving symbolic models largely unexplored. We analyze two public text-to-MIDI systems of contrasting design: the purpose-built encoder--decoder text2midi and MIDI-LLM, a Llama~3.2~1B model extended with MIDI tokens using linear probing, the logit and tuned lenses, activation patching and difference-in-means steering. Across these methods, we recover musically meaningful structure and show how architecture shapes its formation and control. Pitch, instrumentation, harmony and texture are linearly decodable in both models. text2midi refines predictions gradually across depth, whereas MIDI-LLM works largely in its inherited textual basis before a sharp late rotation into the musical vocabulary; patching identifies a matching late attenuation of prompt-driven instrument transfer. Steering produces bidirectional changes in register and polyphony in both systems, and in tempo/energy in MIDI-LLM. Our two-orientation protocol isolates directional control and shows that all-layer interventions are robust in text2midi but accumulate disruptively in MIDI-LLM. Together, the results provide a practical toolkit for tracing and controlling musical concepts in symbolic generators. Audio examples are available on a demo website.

cs.SD↗

arXiv2026

The Perceived Fragility of Explanations in Audio Models: Manipulation of Attribution with Unchanged Predictions

This paper investigates the fragility of post-hoc explanation methods in audio deepfake detection. While previous work on explanation manipulation focused on images using standard $L_p$ metrics, we introduce a psychoacoustic framework that optimizes inaudible perturbations to decouple model attributions from final classifications. We evaluate this vulnerability across state-of-the-art architectures under strict prediction-preserving constraints. By evaluating the manipulation cost through domain-specific perceptual audio quality metrics alongside explanation alignment criteria, our framework demonstrates that an adversary can systematically distort automated explanation heatmaps while preserving the predicted deepfake label. Full code available at: https://github.com/cncPomper/Audio-XAI

cs.SD↗

arXiv2026

Deep Dreams Are Made of This: Visualizing Monosemantic Features in Diffusion Models

This paper proposes latent visualization by optimization (LVO), a mechanistic interpretability technique that extends feature visualization by optimization - originally developed for convolutional neural networks - to latent diffusion models. LVO employs sparse autoencoders (SAEs) to disentangle polysemantic layer representations into monosemantic features. Key contributions include latent-space optimization, time-step activity analysis, schedule-matched noise injection, prior initialization through feature steering, and suitable regularization strategies. We demonstrate the method on Stable Diffusion 1.5 fine-tuned on the Style50 dataset, showing that SAE features produce clear visualizations of recognizable concepts - including diagonal compositions, human figures, roses, cables, and waterfall foam - that correlate with dataset examples, while the baseline without disentanglement produces less coherent results. We further show that regularization techniques from pixel-space feature visualization transfer to the latent domain, though they require different configurations for the raw-layer and SAE variants. Compared to dataset examples and steering, LVO provides complementary insights by directly revealing what activates a feature rather than its downstream effects.

cs.LG↗

arXiv2026

ConceptCaps: a Distilled Concept Dataset for Interpretability in Music Models

Concept-based interpretability methods like TCAV require clean, well-separated positive and negative examples for each concept. Existing music datasets lack this structure: tags are sparse, noisy, or ill-defined. We introduce ConceptCaps, a dataset of 21k music-caption-tags triplets with explicit labels from a 200-attribute taxonomy. Our pipeline separates semantic modeling from text generation: a VAE learns plausible attribute co-occurrence patterns, a fine-tuned LLM converts attribute lists into professional descriptions, and MusicGen synthesizes corresponding audio. This separation improves coherence and controllability over end-to-end approaches. We validate the dataset through audio-text alignment (CLAP), linguistic quality metrics (BERTScore, MAUVE), and TCAV analysis confirming that concept probes recover musically meaningful patterns. Dataset and code are available online.

cs.SD↗

arXiv2025

Prevailing Research Areas for Music AI in the Era of Foundation Models

Parallel to rapid advancements in foundation model research, the past few years have witnessed a surge in music AI applications. As AI-generated and AI-augmented music become increasingly mainstream, many researchers in the music AI community may wonder: what research frontiers remain unexplored? This paper outlines several key areas within music AI research that present significant opportunities for further investigation. We begin by examining foundational representation models and highlight emerging efforts toward explainability and interpretability. We then discuss the evolution toward multimodal systems, provide an overview of the current landscape of music datasets and their limitations, and address the growing importance of model efficiency in both training and deployment. Next, we explore applied directions, focusing first on generative models. We review recent systems, their computational constraints, and persistent challenges related to evaluation and controllability. We then examine extensions of these generative approaches to multimodal settings and their integration into artists' workflows, including applications in music editing, captioning, production, transcription, source separation, performance, discovery, and education. Finally, we explore copyright implications of generative music and propose strategies to safeguard artist rights. While not exhaustive, this survey aims to illuminate promising research directions enabled by recent developments in music foundation models.

cs.SD↗

arXiv2025

M6(GPT)3: Generating Multitrack Modifiable Multi-Minute MIDI Music from Text using Genetic algorithms, Probabilistic methods and GPT Models in any Progression and Time Signature

This work introduces the M6(GPT)3 composer system, capable of generating complete, multi-minute musical compositions with complex structures in any time signature, in the MIDI domain from input descriptions in natural language. The system utilizes an autoregressive transformer language model to map natural language prompts to composition parameters in JSON format. The defined structure includes time signature, scales, chord progressions, and valence-arousal values, from which accompaniment, melody, bass, motif, and percussion tracks are created. We propose a genetic algorithm for the generation of melodic elements. The algorithm incorporates mutations with musical significance and a fitness function based on normal distribution and predefined musical feature values. The values adaptively evolve, influenced by emotional parameters and distinct playing styles. The system for generating percussion in any time signature utilises probabilistic methods, including Markov chains. Through both human and objective evaluations, we demonstrate that our music generation approach outperforms baselines on specific, musically meaningful metrics, offering a viable alternative to purely neural network-based systems.

cs.SD↗

arXiv2025

Evaluating Fake Music Detection Performance Under Audio Augmentations

With the rapid advancement of generative audio models, distinguishing between human-composed and generated music is becoming increasingly challenging. As a response, models for detecting fake music have been proposed. In this work, we explore the robustness of such systems under audio augmentations. To evaluate model generalization, we constructed a dataset consisting of both real and synthetic music generated using several systems. We then apply a range of audio transformations and analyze how they affect classification accuracy. We test the performance of a recent state-of-the-art musical deepfake detection model in the presence of audio augmentations. The performance of the model decreases significantly even with the introduction of light augmentations.

cs.SD↗

arXiv2025

Frechet Music Distance: A Metric For Generative Symbolic Music Evaluation

In this paper we introduce the Frechet Music Distance (FMD), a novel evaluation metric for generative symbolic music models, inspired by the Frechet Inception Distance (FID) in computer vision and Frechet Audio Distance (FAD) in generative audio. FMD calculates the distance between distributions of reference and generated symbolic music embeddings, capturing abstract musical features. We validate FMD across several datasets and models. Results indicate that FMD effectively differentiates model quality, providing a domain-specific metric for evaluating symbolic music generation, and establishing a reproducible standard for future research in symbolic music modeling.

cs.SD↗

arXiv2024

Symbotunes: unified hub for symbolic music generative models

Implementations of popular symbolic music generative models often differ significantly in terms of the libraries utilized and overall project structure. Therefore, directly comparing the methods or becoming acquainted with them may present challenges. To mitigate this issue we introduce Symbotunes, an open-source unified hub for symbolic music generative models. Symbotunes contains modern Python implementations of well-known methods for symbolic music generation, as well as a unified pipeline for generating and training.

cs.SD↗