DARE-EEG: A Foundation Model for Mining Dual-Aligned Representation of EEG
Yang Shao, Peiliang Gong, Qun Dai, Daoqiang Zhang
Abstract
Foundation models pre-trained through masked reconstruction on large-scale EEG data have emerged as a promising paradigm for learning generalizable neural representations across diverse brain-computer interface applications. However, a critical yet overlooked challenge is that EEG encoders must learn representations invariant to incomplete observations-when different masked views of the same signal have minimal overlap, existing methods fail to constrain them to a consistent latent subspace, leading to degraded transferability. To address this, we propose DARE-EEG, a self-supervised foundation model that explicitly enforces the mask-invariance property through dual-aligned representation learning during pre-training. Specifically, we introduce mask alignment that constrains representations from multiple masked views of the same EEG sample via contrastive learning, complementing anchor alignment that aligns masked representations to momentum-updated complete features for semantic stability. Additionally, we propose conv-linear-probing, a parameter-efficient strategy that adapts pre-trained representations to heterogeneous electrode configurations and sampling rates through decoupled spectro-spatial projections. Extensive experiments across diverse EEG benchmarks demonstrate that DARE-EEG consistently achieves state-of-the-art in accuracy performance while maintaining relatively low parameter complexity and superior cross-dataset portability compared to existing methods. Furthermore, DARE-EEG contributes to effectively discovering and utilizing the rich potential representations in EEG.
AI Impact Assessments
(1 model)Scientific Impact Assessment: DARE-EEG
1. Core Contribution
DARE-EEG addresses a specific limitation in masked autoencoder-based EEG foundation models: the lack of explicit constraints ensuring that different masked views of the same EEG signal map to consistent latent representations. The paper terms this the "Mask-Invariance Property" (MIP) and proposes a dual-alignment framework combining: (a) Anchor Alignment (AA) — aligning masked representations to momentum-updated complete features (borrowed from Data2Vec/EEGPT), and (b) Mask Alignment (MA) — a novel contrastive learning objective that explicitly enforces consistency between different masked views of the same sample. Additionally, the paper introduces Conv-Linear-Probing (CLP), a lightweight adaptation module using decoupled 1×1 (channel) and large-kernel (temporal) convolutions to handle heterogeneous electrode configurations and sampling rates during downstream transfer.
The core novelty lies in the combination of MA with AA for EEG foundation models. The MA component is conceptually related to multi-crop strategies in self-supervised vision (e.g., DINO, SwAV), but its application to masked EEG reconstruction and the specific framing around mask invariance is contextually novel.
2. Methodological Rigor
Architecture and Training: The model is pre-trained on ~264K EEG samples (~917 hours) spanning five paradigms, with a 103M parameter deep variant. The use of Global Learnable Tokens (GLTs) embedded in the channel dimension, RoPE for temporal ordering, and a separate predictor/reconstructor pipeline is well-motivated. The masking strategy (50% temporal patch masking with 80% channel masking within visible patches) is aggressive, making the MIP argument more compelling.
Theoretical Analysis: The paper provides propositions in Appendix C arguing that AA alone cannot guarantee MIP (via a counterexample construction where mask-dependent residuals are orthogonal to the target representation), while MA explicitly optimizes for inter-view consistency. While these arguments are intuitive and directionally correct, they are more illustrative than rigorous proofs — the counterexample shows a *possible* failure mode but doesn't characterize when it *will* occur in practice.
Experimental Design: Evaluation covers seven downstream benchmarks across clinical (TUAB, TUEV), motor imagery (BCIC-2A/2B), emotion (SEEDIV), sleep staging (SleepEDF), and cognitive workload (MMWM) tasks. The comparison includes both classical supervised models (EEGNet, DeepCNN, Conformer) and recent foundation models (BENDR, BIOT, LaBraM, EEGPT). Leave-one-subject-out cross-validation is used where appropriate.
Concerns:
3. Potential Impact
The paper addresses a genuine practical challenge in EEG foundation models — the need for robust representations under partial observations. The CLP module is a practical contribution for cross-dataset transfer, as electrode configuration heterogeneity is a persistent barrier in EEG research. The analysis in Appendix B.3 showing learned channel projections for motor imagery vs. cognitive workload tasks provides useful interpretability insights.
If the improvements hold broadly, this work could influence how EEG foundation models are pre-trained, potentially becoming a standard component alongside masked reconstruction. The parameter efficiency argument (6M base model achieving competitive results) is attractive for resource-constrained BCI applications.
However, the impact may be somewhat bounded: the MA component is essentially applying established contrastive learning between augmented views (a well-known technique) to the specific setting of masked EEG views. The conceptual leap, while valid, is incremental.
4. Timeliness & Relevance
EEG foundation models are an active and rapidly growing area, with BENDR, BIOT, LaBraM, and EEGPT all appearing in the 2023-2024 timeframe. The paper directly builds on and competes with these recent works. The identified MIP problem is timely — as masking ratios increase (following MAE trends from vision), the consistency of masked representations becomes more critical. The need for cross-dataset portability in EEG is a persistent bottleneck that CLP partially addresses.
5. Strengths & Limitations
Key Strengths:
Notable Limitations:
Overall Assessment
DARE-EEG makes a solid, well-executed contribution to EEG foundation models by identifying and addressing mask-invariance as a specific failure mode of masked reconstruction approaches. The dual-alignment framework and CLP adaptation strategy are well-motivated and produce consistent (if sometimes modest) improvements. The work is thorough in evaluation and presentation but represents an incremental advance built on well-established techniques (contrastive learning, momentum encoders) applied to a specific domain. The paper's impact will likely be strongest within the EEG/BCI community rather than broadly across machine learning.
Generated May 19, 2026
Comparison History (25)
DARE-EEG addresses a fundamental challenge in EEG foundation models with broad applicability across brain-computer interface applications. It introduces a principled self-supervised learning framework (dual-aligned representation learning) with demonstrated state-of-the-art results across diverse benchmarks, showing strong cross-dataset portability. The work has broader impact potential across neuroscience, clinical diagnostics, and BCI fields. In contrast, DemoEvolve addresses a narrower problem (harness evolution for LLM agents in games) with experiments limited to two game environments, and its contributions, while interesting, are more incremental in scope.
Paper 1 likely has higher scientific impact due to broader cross-domain relevance and stronger foundational contribution: improving self-supervised EEG foundation models via explicit mask-invariance (dual alignment) plus practical adaptation to heterogeneous sensor setups, with extensive benchmarking and clear applicability to many BCI/neuroscience tasks. Paper 2 addresses an important and timely evaluation bias for LLM-based finance backtesting, but its impact is narrower (primarily finance/LLM evaluation), depends on a specific threat model (parametric memorization), and may be more sensitive to protocol/design choices. Overall, Paper 1’s methods generalize across datasets and applications more widely.
DARE-EEG addresses a fundamental challenge in EEG foundation models—mask-invariant representation learning—with broader scientific impact. It introduces a novel self-supervised framework applicable across diverse brain-computer interface applications, with cross-dataset portability addressing a critical limitation in neuroscience/clinical AI. The conv-linear-probing strategy for heterogeneous electrode configurations is practically significant. While DecoR offers useful LLM routing improvements, it addresses a more incremental engineering problem. DARE-EEG's contributions span neuroscience, clinical applications, and representation learning methodology, giving it wider interdisciplinary reach.
Paper 1 presents a novel foundation model (DARE-EEG) with concrete technical innovations (dual-aligned representation learning, conv-linear-probing) validated through extensive experiments across diverse EEG benchmarks, demonstrating state-of-the-art results. It addresses a fundamental challenge in EEG representation learning with broad applicability across brain-computer interface applications. Paper 2, while practically useful, is primarily a methodological framework/catalog for LLM agent architectures without empirical validation at scale—it organizes existing concepts rather than introducing fundamentally new scientific contributions. Paper 1's rigorous experimental methodology and novel technical contributions give it higher scientific impact potential.
DARE-EEG addresses a fundamental challenge in EEG foundation models—mask-invariance in self-supervised learning—with broad applicability across brain-computer interface tasks. Foundation models for biosignals are a rapidly growing area with significant real-world medical and neuroscience applications. The paper introduces novel dual-alignment pre-training and a practical conv-linear-probing strategy for cross-dataset portability, demonstrating state-of-the-art results across diverse benchmarks. Paper 2 on multi-task unlearning is a solid contribution but addresses a more niche problem with narrower scope. EEG foundation models have greater potential for cross-field impact in neuroscience, healthcare, and AI.
Paper 2 likely has higher scientific impact: it proposes a new self-supervised objective (dual alignment for mask-invariance) plus a practical adaptation method (conv-linear-probing) and reports broad, state-of-the-art gains across diverse EEG benchmarks—suggesting strong methodological contribution and real-world BCI/health applications. Its ideas may generalize to other masked-view representation learning settings. Paper 1 is valuable and timely as a diagnostic/negative-result study of LLM agents in hardware-aware optimization, but its impact is more specialized and primarily characterizes limitations rather than delivering a broadly reusable method.
Paper 2 introduces a foundation model for EEG data with broad, transformative applications in Brain-Computer Interfaces, neuroscience, and healthcare. By solving masked reconstruction challenges and handling heterogeneous configurations, it addresses critical bottlenecks in neural representation learning. While Paper 1 identifies an important methodological flaw in LLM-based educational assessments, its scope is relatively narrow. Paper 2's potential to enable widespread advancements across multiple scientific and medical domains gives it a higher estimated scientific impact.
DARE-EEG addresses a fundamental challenge in EEG foundation models with a novel dual-aligned representation learning approach, demonstrating state-of-the-art results across diverse benchmarks. It has broader impact potential through brain-computer interface applications, medical diagnostics, and neuroscience research. The methodological contributions (mask-invariance, conv-linear-probing) are technically deeper and more generalizable. BLINKG, while useful as a benchmark for LLM-based knowledge graph construction, has a narrower scope and primarily evaluates existing capabilities rather than introducing transformative methodology.
Paper 2 (SceneCode) has higher impact potential due to its broader cross-field relevance (generative modeling, embodied AI, robotics, simulation, program synthesis), strong real-world applicability (editable, executable, simulator-ready articulated scenes), and timeliness as agentic/code-based generation becomes central to interactive 3D content creation. Its programmatic representation and execution-guided repair offer a novel, controllable alternative to static-mesh pipelines with clear downstream utility. Paper 1 advances EEG foundation modeling with solid rigor and BCI relevance, but its impact is more domain-specific and likely narrower in breadth.
While both papers present significant advancements, Paper 2 (ChemVA) targets a critical bottleneck in AI-driven scientific discovery: extracting and reasoning over chemical reaction diagrams. By enabling open-weight LLMs to rival proprietary systems with a massive 20% performance gain, it democratizes automated chemical reasoning. This has immediate, high-value real-world applications in drug discovery, material science, and automated literature digitization. Furthermore, the introduction of a new benchmark (OCRD-Bench) provides a lasting resource for the AI4Science community, offering slightly broader cross-disciplinary impact than the specialized BCI advancements in Paper 1.