Selected work

Representative Papers

From Affect to Complex Behavior: Advancing Multimodal Human-Centered AI at the 10th ABAW Workshop & Competition

May 24, 2026

This work addresses the modeling and understanding of human affect and complex behaviors in real-world, unconstrained settings, tackling core challenges such as continuous and discrete emotion recognition alongside high-level behavioral analysis. Through a dual-track approach combining competitions and publications, it introduces novel tasks—including emotion mimicry intensity estimation, detection of hesitation and ambivalence, and fine-grained violence recognition—thereby extending the frontiers of traditional affective computing. Leveraging a large-scale in-the-wild multimodal dataset, the study establishes a standardized benchmark by integrating pose and motion estimation, robust modeling techniques, and fairness-aware evaluation protocols. This effort provides an authoritative platform for advancing research in affective and behavioral understanding, fostering collaborative innovation and practical deployment of multimodal human-centered AI systems.

0 citationsRead paper

NeuroRVQ: Multi-Scale EEG Tokenization for Generative Large Brainwave Models

Oct 14, 2025

Existing EEG foundation models suffer from insufficient reconstruction fidelity due to neural tokenizers’ inability to preserve high-frequency dynamic information. To address this, we propose NeuroRVQ—a multiscale tokenization framework for full-spectrum EEG—integrating multiscale feature extraction, hierarchical residual vector quantization (RVQ), and a phase-amplitude-aware loss function. NeuroRVQ is the first method to achieve cross-band, high-fidelity EEG signal reconstruction. It substantially reduces reconstruction error and consistently outperforms state-of-the-art large-scale EEG models across diverse downstream tasks, demonstrating superior representational capacity and generalizability. By providing a high-quality, strongly prior-guided discrete representation, NeuroRVQ establishes a robust foundation for generative EEG foundation models, enabling advanced applications such as neural decoding and multimodal biosignal fusion.

0 citationsRead paper

Are Large Brainwave Foundation Models Capable Yet? Insights from Fine-tuning

Jul 01, 2025

Large EEG foundation models (LBMs) suffer from high fine-tuning costs and marginal performance gains on BCI tasks such as memory decoding and sleep stage classification. Method: This work introduces Low-Rank Adaptation (LoRA) to EEG foundation modeling for the first time and proposes a multi-component collaborative adaptation strategy to improve parameter efficiency and physiological interpretability. Contribution/Results: Full-parameter fine-tuning yields only +0.9–1.2% accuracy improvement over conventional deep models, despite LBMs having three orders of magnitude more parameters. In contrast, LoRA reduces trainable parameters by >90% with no performance degradation. Ablation studies reveal that current LBM architectures are ill-suited to the time-frequency characteristics of EEG signals, highlighting the need for neurophysiologically grounded architectural redesign. This study provides empirical evidence and methodological guidance for developing efficient, interpretable EEG foundation models.

0 citationsRead paper

Advancing Brainwave Modeling with a Codebook-Based Foundation Model

May 22, 2025

Existing EEG pre-trained models inadequately model neural oscillatory features, limiting generalization and performance across BCI tasks. To address this, we propose LaBraM++, the first foundational EEG model integrating signal-processing priors with codebook-enhanced representation learning. LaBraM++ introduces band-guided learnable vector quantization, time-frequency domain self-supervised pre-training, and a lightweight adaptation head—collectively overcoming representational capacity bottlenecks and significantly enhancing oscillatory information capture. Evaluated across diverse BCI paradigms—including motor imagery, SSVEP, and ERP classification—LaBraM++ consistently outperforms state-of-the-art baselines, achieving an average accuracy improvement of 4.2% and a 32% reduction in training time. It establishes new open-source SOTA performance among EEG foundation models.

0 citationsRead paper

Recent publications

Latest Papers

From Affect to Complex Behavior: Advancing Multimodal Human-Centered AI at the 10th ABAW Workshop & Competition

May 24, 2026

This work addresses the modeling and understanding of human affect and complex behaviors in real-world, unconstrained settings, tackling core challenges such as continuous and discrete emotion recognition alongside high-level behavioral analysis. Through a dual-track approach combining competitions and publications, it introduces novel tasks—including emotion mimicry intensity estimation, detection of hesitation and ambivalence, and fine-grained violence recognition—thereby extending the frontiers of traditional affective computing. Leveraging a large-scale in-the-wild multimodal dataset, the study establishes a standardized benchmark by integrating pose and motion estimation, robust modeling techniques, and fairness-aware evaluation protocols. This effort provides an authoritative platform for advancing research in affective and behavioral understanding, fostering collaborative innovation and practical deployment of multimodal human-centered AI systems.

0 citationsRead paper

NeuroRVQ: Multi-Scale EEG Tokenization for Generative Large Brainwave Models

Oct 14, 2025

Existing EEG foundation models suffer from insufficient reconstruction fidelity due to neural tokenizers’ inability to preserve high-frequency dynamic information. To address this, we propose NeuroRVQ—a multiscale tokenization framework for full-spectrum EEG—integrating multiscale feature extraction, hierarchical residual vector quantization (RVQ), and a phase-amplitude-aware loss function. NeuroRVQ is the first method to achieve cross-band, high-fidelity EEG signal reconstruction. It substantially reduces reconstruction error and consistently outperforms state-of-the-art large-scale EEG models across diverse downstream tasks, demonstrating superior representational capacity and generalizability. By providing a high-quality, strongly prior-guided discrete representation, NeuroRVQ establishes a robust foundation for generative EEG foundation models, enabling advanced applications such as neural decoding and multimodal biosignal fusion.

0 citationsRead paper

Are Large Brainwave Foundation Models Capable Yet? Insights from Fine-tuning

Jul 01, 2025

Large EEG foundation models (LBMs) suffer from high fine-tuning costs and marginal performance gains on BCI tasks such as memory decoding and sleep stage classification. Method: This work introduces Low-Rank Adaptation (LoRA) to EEG foundation modeling for the first time and proposes a multi-component collaborative adaptation strategy to improve parameter efficiency and physiological interpretability. Contribution/Results: Full-parameter fine-tuning yields only +0.9–1.2% accuracy improvement over conventional deep models, despite LBMs having three orders of magnitude more parameters. In contrast, LoRA reduces trainable parameters by >90% with no performance degradation. Ablation studies reveal that current LBM architectures are ill-suited to the time-frequency characteristics of EEG signals, highlighting the need for neurophysiologically grounded architectural redesign. This study provides empirical evidence and methodological guidance for developing efficient, interpretable EEG foundation models.

0 citationsRead paper

Advancing Brainwave Modeling with a Codebook-Based Foundation Model

May 22, 2025

Existing EEG pre-trained models inadequately model neural oscillatory features, limiting generalization and performance across BCI tasks. To address this, we propose LaBraM++, the first foundational EEG model integrating signal-processing priors with codebook-enhanced representation learning. LaBraM++ introduces band-guided learnable vector quantization, time-frequency domain self-supervised pre-training, and a lightweight adaptation head—collectively overcoming representational capacity bottlenecks and significantly enhancing oscillatory information capture. Evaluated across diverse BCI paradigms—including motor imagery, SSVEP, and ERP classification—LaBraM++ consistently outperforms state-of-the-art baselines, achieving an average accuracy improvement of 4.2% and a 32% reduction in training time. It establishes new open-source SOTA performance among EEG foundation models.

0 citationsRead paper