arXiv_SD
-
Grounded Decoding for Autoregressive Speech Enhancement via Adaptive Code-Space Grounding and Local LLM Refinement
2026-08-21 13:55:59Hao Shi, Yuan Gao, Zhaoheng Ni, Junyi Peng, Gongping Huang, Yu Tsao, Xugang LuAbstract
Large language model (LLM)-based autoregressive speech enhancement (SE) produces natural speech using learned clean-speech priors, but may hallucinate content unsupported by the input. Deterministic SE better preserves observation-coupled evidence, yet often retains residual noise or local distortion. We propose an evidence-grounded generative SE framework that uses a deterministic estimate as imperfect evidence. A Whisper-guided DPRNN produces an enhanced waveform, which is blended with the observation and tokenized into a discrete evidence sequence. The evidence conditions an autoregressive clean-speech token generator and is reused during decoding through Code-Space Grounding (CSG), which penalizes candidates according to their Hamming distance in the factorized finite-scalar-quantized (FSQ) space. Because the appropriate grounding strength depends on acoustic difficulty, we introduce SNR-Conditioned CSG (SNR-CSG), which maps a calibrated residual-SNR estimate to an utterance-level strength and constructs an adaptive grounded anchor. Although grounding improves content fidelity, the anchor may retain local acoustic defects inherited from the evidence. Since such defects are predominantly local in the FSQ space, nearby tokens may provide better acoustic realizations without large departures from the observation-supported trajectory. We therefore propose Grounded Neighborhood Refinement with LLM ranking (GNR-LLM). It performs one additional teacher-forced pass conditioned on the grounded-anchor history, intersects the LLM top-$K$ candidates with a local FSQ Hamming neighborhood. Experiments on in-domain, controlled-SNR, and DNS no-reverb conditions show that SNR-CSG provides robust automatic grounding, while GNR-LLM substantially improves low-SNR perceptual quality without sacrificing content fidelity.
Abstract (translated)
URL
https://arxiv.org/abs/2609.04245
PDF
-
AudioWorldSim: Realistic Binaural Audio Datasets For World Models
Abstract
This technical report presents AudioWorldSim, an open-source platform designed to generate realistic binaural audio datasets and advance research in audio-based machine learning, particularly world models. Built as a custom extension of Meta's SoundSpaces 2.0 platform, AudioWorldSim leverages their comprehensive acoustics framework, but focuses on the automatic rollout of random agent navigations, as well as implements crucial fixes to how continuous sound is composed. AudioWorldSim is made publicly available to the research community at this https URL to facilitate reproducibility.
Abstract (translated)
URL
https://arxiv.org/abs/2608.21075
PDF
-
Neural Multichannel Distant Speaker Diarization and Source Separation with Beta Speaker Activity Prior
2026-08-21 08:35:00Sicheng Mao, Mathieu Fontaine, Anthony Larcher, Roland BadeauAbstract
Distant speaker diarization remains challenging due to adverse acoustic conditions, varying numbers of speakers and overlapping speech. While data-driven approaches have shown strong performance, model-driven methods offer a compelling alternative by leveraging spatial information from multichannel recordings. This paper is motivated to propose a Bayesian diarization model for a model-driven method called neural FCASA to enhance its robustness. Specifically, we propose a beta prior over speaker activity and hence a variational lower bound objective that can be seen as a regularized continuous speaker activity score in place of the original cross-entropy loss to train the diarization model. Our experiments show significant improvements in terms of Diarization Error Rate by at least 3% (16% relatively) and Jaccard Error Rate by at least 4% (20% relatively) on the AMI dataset compared to the baseline.
Abstract (translated)
URL
https://arxiv.org/abs/2608.28661
PDF
-
A Regularized Block Diagonal RLS Algorithm for Acoustic Echo Cancellation
Abstract
While the recursive least square (RLS) algorithm is widely used in adaptive filtering applications like acoustic echo cancellation (AEC) due to its fast convergence rate, its high computational complexity severely limit its practical deployment for long filters. In this paper, a regularized block-diagonal RLS (RBD-RLS) algorithm is proposed to address these challenges. By approximating the inverse covariance matrix as a block-diagonal structure, RBD-RLS simplifies the update process into independent parallel computations of sub-blocks, effectively reducing the computational complexity. Additionally, Tikhonov regularization is applied to each sub-blocks for enhance numerical stability. A series of experimental results demonstrate that RBD-RLS maintains good convergence while significantly reducing computational complexity. Moreover, it still exhibits relative robustness in real-world scenarios.
Abstract (translated)
URL
https://arxiv.org/abs/2608.20693
PDF
-
Tracking the Trend in How Speech Synthesizers Deceive People
2026-08-20 12:27:37Milan \v{S}alko, Anton Firc, Kamil Malinka, Vojt\v{e}ch Stan\v{e}k, Martin Pere\v{s}ini, Filip Ple\v{s}ko, Jakub Re\v{s}Abstract
Advances in speech synthesis have made deepfake audio highly realistic. Earlier studies reported 70-80% human detection accuracy, but relied primarily on older synthesizers. We compare human detection for three selected voice synthesis tools released in 2019, 2022, and 2024 with 82 IT professionals, and benchmark humans against six pretrained detectors on the same material. For fully synthetic speech (full spoofs), the F1 score drops from about 90% for RTVC and YourTTS to 48% for ElevenLabs, although listeners were explicitly warned that deepfakes were present. For partial spoofing, where only one sentence of an utterance is altered, strict accuracy falls to 9%, and listeners classify the synthetic sentence as bona fide 77% of the time. Humans and detectors fail in complementary ways, and neither reliably localizes short manipulations. Additionally, listeners increasingly mislabel bona fide speech as fake, eroding trust in unmanipulated audio. These findings show that human perception alone is unreliable for the selected modern and partial-spoof conditions and motivate procedural verification, provenance, watermarking, and segment-level detection.
Abstract (translated)
URL
https://arxiv.org/abs/2608.19959
PDF
-
Unified Music Identification for Tracks and Versions
2026-08-20 11:35:33R. Oguz Araz, Joan Serr\`a, Yuki Mitsufuji, Xavier Serra, Dmitry BogdanovAbstract
Given a music database, track identification (TI) retrieves the exact track matching an audio excerpt, whereas version identification (VI) retrieves its musical versions. Traditionally, the two tasks have been addressed separately. However, as every track is its own closest version, we investigate whether VI can subsume TI. This requires VI systems to be robust to both signal manipulation and audio degradation. We therefore propose a unified benchmark that evaluates accuracy and robustness on each task. Comparing seven existing models on this benchmark, we show that none of them are both accurate and robust on both tasks. We then train a baseline model targeting both tasks and show that a unified system is possible with 10 s TI queries. Lastly, we characterize the two retrieval constraints that limit our model's TI performance. We envision extending this unification to other music identification tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2608.19919
PDF
-
Fourier is Frontier: Frequency-Aware Autoencoding for High-Fidelity Music Reconstruction
Abstract
Continuous-latent audio autoencoders form the backbone of latent music generators, yet decoders at high compression rates commonly exhibit three failure modes: high-frequency loss, phase incoherence, and stereo-image collapse. These share a structural root: waveform autoencoders lack an explicit frequency axis, leaving no handle for targeted per-band correction. Among five matched-budget representations, the complex STFT achieves the lowest full-band and high-frequency spectral distances, providing direct access to magnitude and phase at every bin. Building on this, we present ear-VAE2, a complex-spectral autoencoder with cross-channel interaction. Spec-SnakeBeta learns a periodic activation per frequency bin with frequency-dependent initialization, outperforming other activation variants while using fewer parameters than the fully independent variant. Duplex-Aware Refiner applies band-specific corrections to magnitude and phase following duplex theory of sound localization. On the 546-track Song Describer Dataset, ear-VAE2 achieves the best point estimates on five of seven reconstruction metrics. The Duplex-Aware Refiner reduces Mel Distance by 19.4% and uses ~45% fewer residual-output dimensions than the Unconstrained Refiner, while also lowering spectral distances, spatial-cue errors, and receiving higher ratings from professional engineers. The downstream generator using ear-VAE2 latents achieves better point estimates on all 12 automatic this http URL page is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2608.19843
PDF
-
Dancing Through Soundscapes: Designing a Low-Cost, Sound-Based Device for Sensing and Interpreting Movement and Dance
2026-08-20 09:29:00Swen E. Gaudl, Silvia Carderelli-GronauAbstract
When we move through space, we often rely on multiple senses beyond vision to perceive and act in that environment: we ``feel'' the presence of others; we build internal representations and models and recall them to navigate the environment. We also leave traces and impressions that others pick up on. The traces include echoes, heat, the displacement of objects such as furniture or footprints, air movement close to the face of another, smells such as perfume, but also the immediate sounds we make when we move and breathe. Movement is a spatial and temporal activity, and dance as a form of movement practice requires coordination of oneself in relation to others, the space and a potential score. When rehearsing dance, dancers have to relate to others often not just by looking but more often by feeling and imagining or remembering where others are based on experience and shared practice. So how can we approach technology-mediated movement and dance? Why should we explore it? How can we support spatial and temporal practice meaningfully and joyfully? In this work, we focus on sound traces; we present the design and rationale for a sound-based artefact that translates movement-based sound into layered, explorable, generative soundscapes. The work contributes a novel artefact for exploring movement-based activities with a audio-first approach, with a focus on the spatial performative experience. The paper further reflects on observations from workshops and public sharings, including how participants used repetition, stillness, environmental sound, and call-and-response to understand and improvise with the soundscape.
Abstract (translated)
URL
https://arxiv.org/abs/2608.19827
PDF
-
Humanoid Musical Robots as Experimental Interfaces for Music-Evoked Emotion
2026-08-20 03:23:36Vincent K. M. Cheung, Jia-Yeu LinAbstract
Advances in technology have led to increasingly sophisticated musical humanoid robots. However, their use has largely been limited to performance and related research in human-robot interaction. In this position paper, we propose a novel perspective: musical humanoid robots as experimental interfaces for investigating music-evoked emotions. We argue that current research is constrained by paradigms relying on pre-recorded auditory stimuli, which fail to capture the multimodal, embodied, and interactive nature of real-world musical experience. Building on existing theories of music cognition and emotion, we identify mechanisms that require controlled manipulation of both acoustic and non-acoustic variables. We show that humanoid robots are well-suited as they enable parametric control of performance variables, reproducibility across trials, and the decoupling and recombination of auditory, visual, and interactive components. We illustrate the technical feasibility of this perspective through a case study of the WAseda Saxophonist Robot 5 (WAS-5), demonstrating reproducible control of acoustic and interaction variables that are prerequisites for future music-emotion experiments. Our work positions musical humanoid robots as a methodological platform that enables future controlled investigations of music-evoked emotions.
Abstract (translated)
URL
https://arxiv.org/abs/2608.20433
PDF
-
MultiVerse: A Creator-Centered Approach to Steering Context-Adaptive Lyrics
Abstract
Generative AI may enable new forms of context-aware creative expression by dynamically tailoring media content to its consumption context. For instance, AI systems could adapt song lyrics to the listener and their current activity. However, existing media adaptation systems primarily optimize for audience experience, often neglecting artists' intent, style, and preference. We address this challenge by introducing a novel creator-centered approach to adaptive media authoring and present MultiVerse, a system that instantiates this approach for steering adaptive lyrics. Our approach allows creators to explicitly author controls based on their intent, lyric structure, and audience context, and uses rule-based validations to ensure controls are followed. We conducted a study with 10 songwriters, comparing MultiVerse with a prompting-based workflow for composing adaptive lyrics. The comparison revealed that creators preferred to author how lyrics adapt by explicitly specifying relevant context and adaptation constraints, while recognizing trade-offs in flexibility and iteration speed. Interviews further revealed that creators viewed adaptive media as enabling new forms of audience connection, introducing a distinct creative process, favoring new compositional strategies, and reshaping notions of authorship.
Abstract (translated)
URL
https://arxiv.org/abs/2608.19350
PDF
-
Geometric Iterative Retrieval for Neural Audio Codec Resynthesis
2026-08-19 17:29:56Leo Schmidt-Traub, Fr\'ed\'eric Berdoz, Luca A. Lanzend\"orfer, Roger WattenhoferAbstract
Neural audio codecs based on Residual Vector Quantization (RVQ) have become the dominant discrete representation for token-based general audio generation, yet resynthesizing high-quality audio from coarse codec tokens remains an open problem and bounds the fidelity of every system that generates them. Prior work has framed resynthesis as a choice between discrete token prediction and continuous regression. We argue that this dichotomy is incomplete and introduce geometric iterative retrieval, a paradigm that uses the RVQ layer hierarchy itself as a natural iterative decomposition in continuous codebook space. Rather than classifying over discrete vocabularies or regressing to a single target vector, our method performs contrastive retrieval in the codebook's geometric space. We evaluate our method on codec restoration tasks across speech and music, and show improvements over both single-pass token prediction and one-step regression baselines.
Abstract (translated)
URL
https://arxiv.org/abs/2608.19141
PDF
-
Does Mapping Non-Maximal Probabilities to GMM Components Matter for S-JEPA Encoder Representations?
Abstract
S-JEPA uses soft Gaussian mixture model (GMM) posteriors instead of hard cluster labels to preserve uncertainty. It remains unclear whether the probability values alone are sufficient, or whether it also matters which GMM components receive the non-maximal probabilities. We test this with two matched controls. FIXED-RANDPERM keeps the top-1 component and probability together with the multiset of non-maximal probability values, but reassigns those non-maximal values using a mapping fixed for each physical frame. UNIFORM-TAIL keeps the top-1 component, its probability, and total non-maximal mass but distributes that mass uniformly. Across three independent seeds, REAL SOFT outperforms both controls on two frozen Encoder readouts. It provides better recovery of the original GMM tail and greater accessibility of spectral dynamics over short time scales after controlling for the complete spectrum of the current frame. In two exposure experiments, both readouts improved overall as more frames retained the original mapping. We also descriptively follow one Phase 2 trajectory after the switch to the online GMM. These results show that the numerical probability structure of the soft target does not fully determine the learned Encoder representation. The mapping of non-maximal probabilities to GMM components also matters.
Abstract (translated)
URL
https://arxiv.org/abs/2608.19084
PDF
-
Computational Features for Symbolic Melody Analysis
2026-08-19 16:01:00David M. Whyatt, Peter M. C. HarrisonAbstract
This paper addresses the general problem of extracting music-theoretic and psychological features from symbolically encoded melodies. We review existing melodic feature extraction toolboxes, enumerate their features, and organise them into a common taxonomy. We then describe a new software library that provides implementations of all of these features in a straightforward Python package. We then demonstrate the combined feature set on the Essen Folksong Collection, using the dataset to produce a series of style classification models. These models help us answer key questions about the interpretability and dimensionality of the feature set. Our results show excellent classification accuracy using the full feature set, and promising performance for an eight-dimensional factor-analytic solution that improves the interpretability of the classifier. We distribute our new toolbox as an open-source Python package, $\mathtt{melody-features}$, which can easily be used in various applications within music analysis, music psychology, and music information retrieval.
Abstract (translated)
URL
https://arxiv.org/abs/2608.19061
PDF
-
Sounds Uncertain: Exploring the Affective Aspects of Sonification for Uncertainty Visualization
2026-08-19 08:31:09Marcel-Simon Dutt, Sita A. Vriend, Elias Elmquist, Daniel WeiskopfAbstract
Affective visualization can influence how users perceive, interpret, and engage with data by embedding and conveying emotion through visual design. While sound is widely used in media to evoke emotions, little is known about how sonification can support affective visualization. In this work, we investigate how sonification can communicate emotion in uncertainty visualizations through a co-design study. Participants created two sonifications to accompany a visualization: one conveying the affective component of uncertainty and one conveying neutrality. Our findings show that uncertainty was commonly associated with wavy auditory qualities related to an ominous sentiment. On the other hand, neutrality was associated with clear and relaxing auditory qualities. These results provide insights for the design of visualizations that integrate sonification to communicate the affective component of uncertainty.
Abstract (translated)
URL
https://arxiv.org/abs/2608.18680
PDF
-
FM Synthesizer Audio-Parameter Shared Embeddings
Abstract
Given a target sound, finding the synthesizer preset that best reproduces it remains a core problem in sound design. Existing methods treat synthesis parameters as flat vectors, discarding the signal routing and parameter interactions that produce audio. We make two contributions. First, to learn a representation of parameters including their signal routing, we design a graph neural network whose message passing structure imitates FM signal processing. Second, we adapt the multimodal objective from SLAP to learn joint embeddings of audio and FM synthesizer parameters, enabling preset retrieval from a gallery. We focus on the Yamaha DX7, where six identical sinusoid operators interact according to one of 32 routing topologies. Our graph encoder's message passing weights are shared across all nodes and layers, enabling processing of arbitrary topologies of any size. When every topology is seen during training, the DX7-GNN and two baselines achieve strong audio-to-preset retrieval. When some topologies are held out for testing, the DX7-GNN substantially outperforms both baselines despite having the fewest parameters. Our ablations further support the claim that imitating FM signal flow in a parameter encoder improves generalization to unseen topologies.
Abstract (translated)
URL
https://arxiv.org/abs/2608.18226
PDF
-
Target Speaker Identification: A Low-Latency Streaming Pipeline
2026-08-18 16:24:49Patrick S. Burke (Children's National Hospital), Satyam Raj (Arizona State University), Sean Kinahan (Arizona State University)Abstract
We present a real-time pipeline of open source, pretrained models for streaming identification of a target speaker, motivated by hearing-aid applications where latency as low as 10 ms can be perceptible. We formulate a two-step approach in which incoming audio is first segmented by speaker using low-latency streaming diarization, followed by speaker verification against a registered target speaker. To emulate conversational speech while minimizing overlap, we use the This American Life Podcast Transcripts dataset and select the host as a consistent target speaker. We benchmark offline diarization with Pyannote and LIUM using diarization error rate (DER) and select Pyannote based on baseline performance and compatibility with streaming. We then evaluate speaker verification using Pyannote and TitaNet-Large and generate ROC curves to select an operating region. We integrate Diart and tune clustering parameters to reduce DER while maintaining real-time operation. We pair Diart with Pyannote verification and evaluate system-level performance by converting predicted and ground-truth speech regions into 100 ms binary masks. Across 17 evaluation episodes, the system achieves greater than 0.90 median accuracy with high specificity (0.95-0.98) at cosine distance thresholds of 0.7-0.75, demonstrating a practical proof of concept for downstream low-latency selective amplification.
Abstract (translated)
URL
https://arxiv.org/abs/2608.17972
PDF
-
UniVerse: Benchmarking and Enhancing LALMs on Culturally Inclusive Low-Resource Music Understanding
2026-08-18 14:47:48Ziya Zhou, Shangda Wu, Shenyang Xu, Yutong Zheng, Dafang Liang, Suin Chung, Danbinaerin Han, Junyan Jiang, Yongyi Zang, Ruibin Yuan, Rongxiu Zhong, Shilei Zhang, Junlan Feng, Jinglei Liu, Haotian Zhou, Zijin Li, Dasaem Jeong, Wei Xue, Yike GuoAbstract
Recent advances in large audio-language models (LALMs) have significantly improved performance in tasks such as music captioning, genre classification, and sound event detection. However, limited attention has been paid to improving their adaptability across diverse musical traditions, particularly folk music rooted in distinct cultural contexts. Folk-music traditions are typically resource-scarce, unevenly represented across regions, and poorly documented. Even when such samples appear in large-scale pre-training, LALMs often fail to capture their structural and stylistic characteristics, partly due to the absence of dedicated evaluation protocols and training solutions. To address these limitations, we introduce UniVerse, a reproducible solution for low-resource music understanding. Specifically, we propose UniVerseBench, a benchmark of 5,042 Q&A pairs across more than 38 cultural and linguistic entities, constructed via an expert-guided yet highly automated pipeline. In parallel, we construct a fully automated, model-generated multi-turn dialogue training dataset UniVerseSet. By training LALMs on UniVerseSet, we systematically adapt and investigate representative multimodal imbalance learning strategies across both dense and Mixture-of-Experts (MoE) architectures. Experimental results indicate that fully automated data curation combined with imbalance-aware training yields non-trivial improvements, but models still struggle to capture fine-grained acoustic features, indicating a gap between surface-level alignment and deep musical comprehension.
Abstract (translated)
URL
https://arxiv.org/abs/2608.17852
PDF
-
ChiroEcho: extending automated bat vocalisation classification beyond the learned taxonomy
2026-08-18 12:38:51Burooj Ghani, Welmoed Eversteijn, Milan van Hirtum, Juan Sebasti\'an Ca\~nas, Vincent J. Kalkman, Dan Stowell, A. Leonie BaierAbstract
Bats are key indicators of ecosystem health and are protected throughout Europe, making reliable population monitoring a conservation priority. Their cryptic nocturnal lifestyle makes passive acoustic monitoring essential, yet automated identification remains difficult as echolocation calls vary with behaviour and environment and overlap among species. We present a deep learning framework that jointly predicts species and genus and combines genus predictions with geographic species distributions at inference. When only one species of a predicted genus occurs in a region, the framework can resolve species absent from the learned taxonomy. This reframes geographic information as a means of extending, rather than constraining, a classifier's effective taxonomy. Using recordings spanning 35 European bat species, we evaluate closed-set classification, examine the instability of performance estimates for sparsely represented species, and conduct a controlled held-out proof-of-principle experiment. The rare-species analysis shows how limited evaluation data can obscure species-level performance, while the held-out experiment shows that genus predictions and location can recover labels unavailable to the species head. Geographic resolution extends operational coverage from 35 to 41 of the 48 native European bat species, increasing coverage from 73% to 85%. To our knowledge, this is the broadest operational coverage reported for automated European bat classification. More broadly, the bat framework provides proof of principle for resolving unseen fine-grained classes by combining coarse predictions with transparent external constraints.
Abstract (translated)
URL
https://arxiv.org/abs/2608.18191
PDF
-
The Last Mile of Deepfake Speech Detection: An Industry-Academia Experience Report
2026-08-18 09:50:02Anton Firc, Kamil Malinka, Vojt\v{e}ch Stan\v{e}k, Miroslav Hlav\'a\v{c}ek, Marek Barto\v{n}Abstract
Synthetic speech detection benchmarks now report sub-1% error rates on some in-domain evaluations, yet performance degrades under unseen attacks, channel mismatch, and distribution shift. Based on a three-year effort with Phonexia, a commercial speaker-recognition vendor, we report barriers encountered while building and deploying a detector. Many public benchmarks are not licensed for commercial model development. Real inputs are not four-second clean clips but long, codec-degraded, sometimes partially synthetic recordings. And when a calibrated system returns a log-likelihood ratio of 2.5, no one can tell the customer what it means for their decision. Rather than proposing a new model, we connect these barriers to concrete research and coordination proposals: shared standards for commercially usable datasets, realistic deployment benchmarks, and scores that non-experts can act on. These observations come from one project and should be tested in other settings.
Abstract (translated)
URL
https://arxiv.org/abs/2608.17585
PDF
-
FireRedTTS3: Unified Speech Generation and Editing with Semantically Enriched Speech Representations
2026-08-18 08:19:46Feiyu Shen, Kun Xie, Yichen Wu, Ziqi Dai, Yichen Han, Junjie Li, Xuelong Geng, Fenglong Xie, Lei Xie, Xu Tang, Yao HuAbstract
Recent continuous autoregressive TTS models operate directly on continuous speech representations, preserving rich acoustic details while leveraging the instruction-following capabilities of text LLMs. This paradigm opens new possibilities for voice cloning, instruction-controlled voice design, and speech editing, but remains susceptible to error accumulation during autoregressive generation. Existing solutions often require additional semantic modules, multi-stage tokenizer training pipelines, or complex autoregressive architectures. In this work, we propose FireRedTTS3, a simple yet effective speech generation and editing framework that mitigates error accumulation at the representation level. Specifically, we leverage a frozen Audio Encoder trained on diverse speech understanding tasks as a semantic teacher to regularize the audio feature space. This improves text-speech alignment and stabilizes autoregressive generation while keeping the overall system simple. FireRedTTS3 provides two variants: FireRedTTS3-Base for multilingual and multi-dialect zero-shot voice cloning, and FireRedTTS3-Instruct for unified voice cloning, instruction-controlled voice design, and speech editing. Experiments show that FireRedTTS3-Base achieves the best average speech intelligibility and speaker similarity among compared systems on Seed-TTS-Eval and MiniMax-MLS-Test, while FireRedTTS3-Instruct outperforms competing systems on InstructTTSEval and Ming-Freeform-Audio-Edit. These results demonstrate that semantically enriched continuous speech representations, combined with a simple architecture, enable stable, controllable, and high-fidelity speech generation and editing. Code and models are available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2608.17492
PDF
