Selected work

Representative Papers

The methodology of Constructing the Large-Scale Dataset for Detecting Presuicidal and Anti-Suicidal Signals in Social Media Texts in Russian

Aug 01, 2026

This study addresses the significant challenge of efficiently identifying suicidal ideation and anti-suicidal求助 signals within the vast, noisy textual data of Russian-language social media—a task critical for timely intervention yet hindered by linguistic and contextual complexity. To this end, the authors propose a systematic methodology encompassing fine-grained category definitions, detailed annotation guidelines, and a multi-stage human annotation and validation pipeline. As a key contribution, they construct and publicly release the first large-scale Russian dataset of pre-suicidal and anti-suicidal signals, comprising over 50,000 annotated posts. The project also provides full code and documentation, and demonstrates the dataset’s utility through baseline classification models evaluated at multiple levels of granularity, thereby establishing a reproducible benchmark for future research on automated detection of suicide-related content in Russian.

0 citationsRead paper

AGORA: Adversarial Generation Of Real-time Animatable 3D Gaussian Head Avatars

Dec 06, 2025

Generating high-fidelity, animatable 3D facial avatars remains challenging due to slow rendering and dynamic inconsistency in implicit representations (e.g., NeRF), and the lack of dynamic controllability in 3D Gaussian Splatting (3DGS). This work proposes the first real-time drivable 3D Gaussian facial avatar method. It introduces a lightweight FLAME-conditioned deformation branch to predict residual Gaussian point displacements for fine-grained expression modeling; designs a dual-discriminator GAN framework to enhance motion realism; and jointly optimizes via parametric mesh supervision and synthetic rendering losses. Our approach achieves the first CPU-based real-time inference (~9 FPS) and GPU inference exceeding 250 FPS—significantly outperforming NeRF-based baselines. It delivers concurrent advances in identity preservation, expression accuracy, and rendering efficiency.

0 citationsRead paper

Speak, Edit, Repeat: High-Fidelity Voice Editing and Zero-Shot TTS with Cross-Attentive Mamba

Oct 06, 2025

This work addresses the challenge of simultaneously achieving high naturalness, speaker consistency, memory efficiency, and low latency in high-fidelity speech editing and zero-shot text-to-speech (TTS). We propose MAVE—a novel architecture that unifies these tasks by synergistically integrating the structured state-space model Mamba with cross-modal cross-attention. MAVE enables context-aware speech editing and high-quality zero-shot TTS without task-specific training, leveraging autoregressive modeling, text-acoustic alignment, and efficient long-sequence processing. On the RealEdit benchmark, 57.2% of edited utterances are indistinguishable from originals; in zero-shot TTS, it surpasses VoiceCraft in naturalness and speaker similarity while reducing inference memory by ~6× and maintaining comparable latency. The core innovation lies in the principled co-design of Mamba and cross-modal attention, yielding significant improvements in the performance–efficiency trade-off.

0 citationsRead paper

COSPADI: Compressing LLMs via Calibration-Guided Sparse Dictionary Learning

Sep 26, 2025

To address the rigidity of low-rank weight approximations and substantial accuracy degradation in post-training compression of large language models (LLMs), this paper proposes CoSpaDi—a training-free LLM compression framework that requires no fine-tuning. Its core innovation replaces conventional low-rank decomposition with sparse dictionary learning: using a small calibration dataset, it jointly optimizes a dense dictionary and a column-wise sparse coefficient matrix, enabling flexible, structured sparse representations of weight columns in heterogeneous subspaces. Additionally, an output activation matching strategy is introduced to preserve functional fidelity. CoSpaDi natively supports synergistic sparse computation and quantization. Experiments on Llama and Qwen families demonstrate that CoSpaDi consistently outperforms state-of-the-art low-rank methods across 20%–50% compression ratios, achieving lower perplexity and higher task accuracy.

0 citationsRead paper

Share Your Attention: Transformer Weight Sharing via Matrix-based Dictionary Learning

Aug 06, 2025

Large language models (LLMs) suffer from prohibitive computational and memory overhead, hindering deployment. Existing compression methods primarily focus on intra-layer optimization, overlooking cross-layer redundancy inherent in Transformer’s repetitive architectural structure. Method: We propose MASA—the first lightweight method to introduce matrix dictionary learning for cross-layer sharing of attention projection matrices in Transformers. MASA learns a shared atomic basis matrix and reconstructs layer-wise weights via linear combinations, enabling structured parameter compression without architectural modification, knowledge distillation, or reliance on key-value caching. Contribution/Results: On LLMs ranging from 100M to 700M parameters, MASA achieves a 66.7% parameter reduction while outperforming Grouped Query Attention (GQA) and low-rank baselines. In Vision Transformers (ViTs), it reduces attention parameters by two-thirds with no accuracy loss. This work is the first to empirically validate dictionary learning for cross-layer weight sharing, establishing a novel paradigm for efficient, pre-trained model compression.

0 citationsRead paper

Recent publications

Latest Papers

The methodology of Constructing the Large-Scale Dataset for Detecting Presuicidal and Anti-Suicidal Signals in Social Media Texts in Russian

Aug 01, 2026

This study addresses the significant challenge of efficiently identifying suicidal ideation and anti-suicidal求助 signals within the vast, noisy textual data of Russian-language social media—a task critical for timely intervention yet hindered by linguistic and contextual complexity. To this end, the authors propose a systematic methodology encompassing fine-grained category definitions, detailed annotation guidelines, and a multi-stage human annotation and validation pipeline. As a key contribution, they construct and publicly release the first large-scale Russian dataset of pre-suicidal and anti-suicidal signals, comprising over 50,000 annotated posts. The project also provides full code and documentation, and demonstrates the dataset’s utility through baseline classification models evaluated at multiple levels of granularity, thereby establishing a reproducible benchmark for future research on automated detection of suicide-related content in Russian.

0 citationsRead paper

AGORA: Adversarial Generation Of Real-time Animatable 3D Gaussian Head Avatars

Dec 06, 2025

Generating high-fidelity, animatable 3D facial avatars remains challenging due to slow rendering and dynamic inconsistency in implicit representations (e.g., NeRF), and the lack of dynamic controllability in 3D Gaussian Splatting (3DGS). This work proposes the first real-time drivable 3D Gaussian facial avatar method. It introduces a lightweight FLAME-conditioned deformation branch to predict residual Gaussian point displacements for fine-grained expression modeling; designs a dual-discriminator GAN framework to enhance motion realism; and jointly optimizes via parametric mesh supervision and synthetic rendering losses. Our approach achieves the first CPU-based real-time inference (~9 FPS) and GPU inference exceeding 250 FPS—significantly outperforming NeRF-based baselines. It delivers concurrent advances in identity preservation, expression accuracy, and rendering efficiency.

0 citationsRead paper

Speak, Edit, Repeat: High-Fidelity Voice Editing and Zero-Shot TTS with Cross-Attentive Mamba

Oct 06, 2025

This work addresses the challenge of simultaneously achieving high naturalness, speaker consistency, memory efficiency, and low latency in high-fidelity speech editing and zero-shot text-to-speech (TTS). We propose MAVE—a novel architecture that unifies these tasks by synergistically integrating the structured state-space model Mamba with cross-modal cross-attention. MAVE enables context-aware speech editing and high-quality zero-shot TTS without task-specific training, leveraging autoregressive modeling, text-acoustic alignment, and efficient long-sequence processing. On the RealEdit benchmark, 57.2% of edited utterances are indistinguishable from originals; in zero-shot TTS, it surpasses VoiceCraft in naturalness and speaker similarity while reducing inference memory by ~6× and maintaining comparable latency. The core innovation lies in the principled co-design of Mamba and cross-modal attention, yielding significant improvements in the performance–efficiency trade-off.

0 citationsRead paper

COSPADI: Compressing LLMs via Calibration-Guided Sparse Dictionary Learning

Sep 26, 2025

To address the rigidity of low-rank weight approximations and substantial accuracy degradation in post-training compression of large language models (LLMs), this paper proposes CoSpaDi—a training-free LLM compression framework that requires no fine-tuning. Its core innovation replaces conventional low-rank decomposition with sparse dictionary learning: using a small calibration dataset, it jointly optimizes a dense dictionary and a column-wise sparse coefficient matrix, enabling flexible, structured sparse representations of weight columns in heterogeneous subspaces. Additionally, an output activation matching strategy is introduced to preserve functional fidelity. CoSpaDi natively supports synergistic sparse computation and quantization. Experiments on Llama and Qwen families demonstrate that CoSpaDi consistently outperforms state-of-the-art low-rank methods across 20%–50% compression ratios, achieving lower perplexity and higher task accuracy.

0 citationsRead paper

Share Your Attention: Transformer Weight Sharing via Matrix-based Dictionary Learning

Aug 06, 2025

Large language models (LLMs) suffer from prohibitive computational and memory overhead, hindering deployment. Existing compression methods primarily focus on intra-layer optimization, overlooking cross-layer redundancy inherent in Transformer’s repetitive architectural structure. Method: We propose MASA—the first lightweight method to introduce matrix dictionary learning for cross-layer sharing of attention projection matrices in Transformers. MASA learns a shared atomic basis matrix and reconstructs layer-wise weights via linear combinations, enabling structured parameter compression without architectural modification, knowledge distillation, or reliance on key-value caching. Contribution/Results: On LLMs ranging from 100M to 700M parameters, MASA achieves a 66.7% parameter reduction while outperforming Grouped Query Attention (GQA) and low-rank baselines. In Vision Transformers (ViTs), it reduces attention parameters by two-thirds with no accuracy loss. This work is the first to empirically validate dictionary learning for cross-layer weight sharing, establishing a novel paradigm for efficient, pre-trained model compression.

0 citationsRead paper