arXiv subjects

Mingru Yang

Publications and source records attributed to Mingru Yang.

12 recordsLinked to original sources

arXiv2026

Summary of DCASE 2026 Task 5: Audio-Dependent Question Answering

DCASE~2026 Task~5 introduces Audio-Dependent Question Answering (ADQA), which tests whether large audio-language models answer from the audio rather than from textual priors. An Audio-Dependency Filtering (ADF) pipeline combines silent-audio probing, per-option perplexity, a large language model (LLM) commonsense check, and human review to remove items solvable from text alone. The 3000 items that pass form the ADQA-Bench evaluation set, spanning music, speech, and environmental audio. The inaugural edition draws 14 teams and 36 submissions across two tracks defined by total parameter count (up to 100B and under 10B). A Chung-Ang University ensemble of MOSS-Audio-8B-Thinking and Qwen3-Omni-30B reaches the top overall accuracy at \pct{58.33}, and a MOSS-only configuration from the same team leads the sub-10B track at \pct{57.30}. Across the 30 submissions with a comparable development score, evaluation accuracy falls by 11.91 percentage points (pp) on average (median 10.91\,pp) on the hidden evaluation split, which is designed to be harder than the development split. The most common building blocks are: the MOSS-Audio-8B-Thinking backbone (13 of 36 submissions), Low-Rank Adaptation (LoRA) fine-tuning on AudioMCQ-StrongAC, and preference or reinforcement-learning objectives -- Group Relative Policy Optimization (GRPO) in five teams, Group reward-Decoupled Normalization Policy Optimization (GDPO) in two. At test time, prompt engineering is near-universal, and majority or choice-permutation voting is common. Every system misses the same set of 233 evaluation items.

eess.AS↗

arXiv2026

TriA Pipeline: A Large-Scale Automatic Audio Annotation Pipeline For Audio Classification In Specific Scenarios

There are some datasets of varying scales for audio classification (AC) applied to different tasks. However, annotated data is limited for most scenarios, such as domestic environments. To address this challenge, we propose an $\textbf{A}$utomatic $\textbf{A}$udio $\textbf{A}$nnotation Pipeline--TriA Pipeline, which can efficiently convert audio from various scenarios into high-quality training data with audio event annotations. A TriA dataset was constructed with the TriA Pipeline, over 2130 hours of audio covering 431 audio classes. Furthermore, we partitioned a prior-knowledge-guided subset (TriA$_{\mathrm{GK}}$) from TriA and conduct comparative experiments on three domestic AC tasks. Comparing the result on manually annotated data only and that on manually annotated data combines TriA$_{\mathrm{GK}}$, TriA$_{\mathrm{GK}}$ could achieve average relative gains of 3.97% in accuracy and 3.35% in Macro-F1, validating the effectiveness of TriA$_{\mathrm{GK}}$ and the TriA Pipeline.

eess.AS↗

arXiv2025

Extracting conserved operators from a projected entangled pair state

Given a tensor network state, how can we determine conserved operators (including Hamiltonians) for which the state is an eigenstate? We answer this question by presenting a method to extract geometrically $k$-local conserved operators that have the given infinite projected entangled pair state (iPEPS) in 2D as an (approximate) eigenstate. The key ingredient is the evaluation of the static structure factors of multi-site operators through differentiating the generating function. These generating functions define a manifold of the given tensor network state deformed by some parameters, endowed with a quantum geometry, where conserved operators correspond to vanishing fidelity susceptibility. Despite the approximation errors, we show that our method is still able to extract from exact or variational iPEPS to good precision both frustration-free and non-frustration-free parent Hamiltonians that are beyond the standard construction and obtain better locality. In particular, we find a 4-site-plaquette local Hamiltonian that approximately has the short-range RVB state as the ground state. Moreover, we find a Hamiltonian for which the deformed toric code state at arbitrary string tension is an excited eigenstate with the same energy, thereby potentially realizing quantum many-body scars.

quant-ph↗

arXiv2025

Excited states from local effective Hamiltonians of matrix product states and their entanglement spectrum transition

Solving excited states is a challenging task for interacting systems. For one-dimensional critical systems, however, excited states can be directly accessed from the eigenvectors of the local effective Hamiltonian that is constructed from the ground state obtained by variational matrix product state (MPS) optimization. Despite its numerical success, the theoretical mechanism underlying this method has remained largely unexplored. In this work, we provide a conformal field theory (CFT) perspective that helps elucidate this connection. The key insight is that this construction effectively uses a truncated basis of ground-state Schmidt vectors to represent excited states, where the contribution of each Schmidt vector can be expressed as a CFT correlation function and shown to decay with increasing Schmidt index. The CFT analysis further predicts an entanglement-spectrum transition of excited states as the ratio of the subsystem size to the total system size is varied. Our numerical results support this picture and demonstrate a reorganization of the entanglement spectrum into distinct conformal towers as this ratio changes.

cond-mat.str-el↗

arXiv2025

BridgeCode: A Dual Speech Representation Paradigm for Autoregressive Zero-Shot Text-to-Speech Synthesis

Autoregressive (AR) frameworks have recently achieved remarkable progress in zero-shot text-to-speech (TTS) by leveraging discrete speech tokens and large language model techniques. Despite their success, existing AR-based zero-shot TTS systems face two critical limitations: (i) an inherent speed-quality trade-off, as sequential token generation either reduces frame rates at the cost of expressiveness or enriches tokens at the cost of efficiency, and (ii) a text-oriented supervision mismatch, as cross-entropy loss penalizes token errors uniformly without considering the fine-grained acoustic similarity among adjacent tokens. To address these challenges, we propose BridgeTTS, a novel AR-TTS framework built upon the dual speech representation paradigm BridgeCode. BridgeTTS reduces AR iterations by predicting sparse tokens while reconstructing rich continuous features for high-quality synthesis. Joint optimization of token-level and feature-level objectives further enhances naturalness and intelligibility. Experiments demonstrate that BridgeTTS achieves competitive quality and speaker similarity while significantly accelerating synthesis. Speech demos are available at https://test1562.github.io/demo/.

cs.SD↗

arXiv2025

Measuring Audio's Impact on Correctness: Audio-Contribution-Aware Post-Training of Large Audio Language Models

Large Audio Language Models (LALMs) represent an important frontier in multimodal AI, addressing diverse audio tasks. Recently, post-training of LALMs has received increasing attention due to significant performance improvements over foundation models. While single-stage post-training such as reinforcement learning (RL) has demonstrated promising results, multi-stage approaches such as supervised fine-tuning (SFT) followed by RL remain suboptimal. The allocation of data across multiple training stages to maximize LALM capabilities has not been fully explored, and large-scale, high-quality datasets for such research are also lacking. To address these problems, we firstly present AudioMCQ, a comprehensive audio multiple-choice question dataset comprising 571k samples with two kinds of chain-of-thought annotations. Secondly, we investigate the prevalent zero audio-contribution phenomenon in LALMs, where models derive correct answers solely from textual information without processing audio content. We propose Audio-Contribution Filtering to partition data into weak and strong audio-contribution subsets. Based on these insights, we develop two effective post-training paradigms: Weak-to-Strong (SFT on weak audio-contribution data followed by RL on strong audio-contribution data) and Mixed-to-Strong (SFT on mixed audio-contribution data followed by RL on strong audio-contribution data). We achieve first place in the DCASE 2025 Audio-Question-Answering challenge by using AudioMCQ. Additionally, leveraging our dataset with different training strategies, we achieve 78.2\% on MMAU-test-mini, 75.6\% on MMAU, 67.1\% on MMAR, and 70.7\% on MMSU, establishing new state-of-the-art performance.

eess.AS↗

arXiv2025

Generalizable Audio Deepfake Detection via Hierarchical Structure Learning and Feature Whitening in Poincar\'e sphere

Audio deepfake detection (ADD) faces critical generalization challenges due to diverse real-world spoofing attacks and domain variations. However, existing methods primarily rely on Euclidean distances, failing to adequately capture the intrinsic hierarchical structures associated with attack categories and domain factors. To address these issues, we design a novel framework Poin-HierNet to construct domain-invariant hierarchical representations in the Poincar\'e sphere. Poin-HierNet includes three key components: 1) Poincar\'e Prototype Learning (PPL) with several data prototypes aligning sample features and capturing multilevel hierarchies beyond human labels; 2) Hierarchical Structure Learning (HSL) leverages top prototypes to establish a tree-like hierarchical structure from data prototypes; and 3) Poincar\'e Feature Whitening (PFW) enhances domain invariance by applying feature whitening to suppress domain-sensitive features. We evaluate our approach on four datasets: ASVspoof 2019 LA, ASVspoof 2021 LA, ASVspoof 2021 DF, and In-The-Wild. Experimental results demonstrate that Poin-HierNet exceeds state-of-the-art methods in Equal Error Rate.

cs.SD↗

arXiv2025

Beyond fragmented dopant dynamics in quantum spin lattices: Robust localization and non-Gaussian diffusion

The motion of dopants in magnetic spin lattices has received tremendous attention for at least four decades due to its connection to high-temperature superconductivity. Despite these efforts, we lack a complete understanding of their behavior, especially out of the equilibrium and at nonzero temperatures. In this paper, we take a significant step towards a much deeper understanding based on state-of-the-art matrix-product-state calculations. In particular, we investigate the non-equilibrium dynamics of a dopant in two-leg $t$--$J$ ladders with antiferromagnetic XXZ spin interactions. In the Ising limit, we find that the dopant is localized for all investigated nonzero temperatures due to an emergent disordered potential, with a localization length controlled by the underlying correlation length of the spin lattice, which increases exponentially with decreasing temperature. The dopant, hereby, only delocalizes asymptotically in the zero temperature limit. This greatly generalizes the localization effect discovered recently in Hilbert space fragmented models. In the presence of spin-exchange processes at rate $\alpha$, the dopant diffuses with a diffusion coefficient, $D_h$, depending non-monotonically on $\alpha$. It initially increases linearly as $D_h \propto \alpha$ for $\alpha \ll 1$ before dropping off as $\alpha^{-1}$ for $\alpha > 1$. Moreover, we show that the underlying spin dynamics at infinite temperature behaves qualitatively the same, albeit with important quantitative differences. We substantiate these findings by showing that the dynamics features self-similar scaling behavior, which strongly deviates from the Gaussian behavior of regular diffusion, especially for weak spin exchange. Finally, we show that the diffusion coefficient $D_h$ follows an Arrhenius relation at high temperatures, whereby it is exponentially suppressed upon cooling.

cond-mat.quant-gas↗

arXiv2023

Validating phase-space methods with tensor networks in two-dimensional spin models with power-law interactions

Using a recently developed extension of the time-dependent variational principle for matrix product states, we evaluate the dynamics of 2D power-law interacting XXZ models, implementable in a variety of state-of-the-art experimental platforms. We compute the spin squeezing as a measure of correlations in the system, and compare to semiclassical phase-space calculations utilizing the discrete truncated Wigner approximation (DTWA). We find the latter efficiently and accurately captures the scaling of entanglement with system size in these systems, despite the comparatively resource-intensive tensor network representation of the dynamics. We also compare the steady-state behavior of DTWA to thermal ensemble calculations with tensor networks. Our results open a way to benchmark dynamical calculations for two-dimensional quantum systems, and allow us to rigorously validate recent predictions for the generation of scalable entangled resources for metrology in these systems.

quant-ph↗

arXiv2022

Detecting emergent continuous symmetries at quantum criticality

New or enlarged symmetries can emerge at the low-energy spectrum of a Hamiltonian that does not possess the symmetries, if the symmetry breaking terms in the Hamiltonian are irrelevant under the renormalization group flow. In this letter, we propose a tensor network based algorithm to numerically extract lattice operator approximation of the emergent conserved currents from the ground state of any quantum spin chains, without the necessity to have prior knowledge about its low-energy effective field theory. Our results for the spin-1/2 $J$-$Q$ Heisenberg chain and a one-dimensional version of the deconfined quantum critical points (DQCP) demonstrate the power of our method to obtain the emergent lattice Kac-Moody generators. It can also be viewed as a way to find the local integrals of motion of an integrable model and the local parent Hamiltonian of a critical gapless ground state.

cond-mat.str-el↗

arXiv2020

Time Dependent Variational Principle with Ancillary Krylov Subspace

We propose an improved scheme to do the time dependent variational principle (TDVP) in finite matrix product states (MPS) for two-dimensional systems or one-dimensional systems with long range interactions. We present a method to represent the time-evolving state in a MPS with its basis enriched by state-averaging with global Krylov vectors. We show that the projection error is significantly reduced so that precise time evolution can still be obtained even if a larger time step is used. Combined with the one-site TDVP, our approach provides a way to dynamically increase the bond dimension while still preserving unitarity for real time evolution. Our method can be more accurate and exhibit slower bond dimension growth than the conventional two-site TDVP.

cond-mat.str-el↗

arXiv2018

Density Matrix Renormalization Group Study of a One Dimensional Diatomic Molecule beyond the Born-Oppenheimer Approximation

We study one dimensional models of diatomic molecules where both the electrons and nuclei are treated as quantum particles, going beyond the usual Born-Oppenheimer approximation. The continuous system is approximated by a grid which computationally resembles a ladder, with the electrons living on one leg and the nuclei on the other. To simulate DMRG efficiently with this system, a three-site algorithm has been implemented. We also use a compression method to treat the long-range interactions between charged particles. We find that 1D diatomic molecules with spin-1/2 nuclei in the spin-triplet state will unbind when the mass of the nuclei reduces to only a few times larger than the electron mass, while the molecule with nuclei in the singlet state always binds, given the two electrons in their singlet state in both cases.

cond-mat.str-el↗