DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models

arXiv:2610.14497 · cs.LG · Submitted 2026-10-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models".

Jane: The gist The DISENTANGLED EFM FINE-TUNING (DEFT) framework proposes a plug-and-play adapter for EEG foundation models that encourages branch disentanglement into task-oriented content and subject-consistent style,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we’ve talked about how DEFT separates the representation into content and style branches before classification. Now, Jane, can you give us a simple way to look at what this actually achieves for someone who just wants to use these EEG models?

Jane: Okay. Think of it like this: Imagine the model has a huge messy filing cabinet full of every possible brain signal information it learned. DEFT is like an organizer that sorts those files into two distinct bins, one bin for the actual task—like "is this patient depressed or not"—and another bin for the consistent style, which is basically everything related to *who* the person is.

Tom: Right, so instead of having to sift through all that noise in one big pile to find out what’s important for the classification head, you have a cleaner path. That means the final prediction head doesn't have to do all that heavy lifting on its own.

Jane: Exactly. And this is especially important when you’re looking at different subjects, say testing the model on people from different hospitals. Because DEFT separates subject style early, it handles those subject differences more consistently, so the classification part stays focused purely on the task information you actually care about.

Tom: I like that idea of cleaning up the representation before classifying. The paper shows they use a specific method called recursive Gaussian disentanglement to do this splitting iteratively. It’s not just one pass; it's a back-and-forth process using these counterpart-subtracted observations to find the content and style parts.

Jane: That iterative part is what makes it robust, I think. And they also have these constraints—intrinsic and external ones—to make sure the content branch actually varies enough across different time windows, while the style branch stays stable within a single trial. This keeps things organized in a way that matches how brain signals naturally flow over time.

Tom: The numbers are interesting too. They show that this approach, DEFT, improves the mean B-ACC score by zero point four hundred sixty when fine-tuning these models on EEG data across four different foundation backbones. That’s a concrete improvement on top of the baseline methods they tested.

Jane: It shows that for cross-subject EEG decoding, this isn't just theoretical stuff; it translates into a measurable gain in performance. But Tom, the paper also notes that while the content branch usually gives you better task-content geometry, that benefit can change depending on what kind of task you’re trying to solve.

Tom: That makes sense. The structure helps, but the specific gains depend on whether your downstream goal aligns perfectly with how they disentangled those features. Lalam has seen a lot of literature, so what does this mean for the broader direction of representation learning in AI?

Lalam: This work supports the idea that we can take these massive foundation models and apply a factorize-before-classify strategy. It suggests that decomposing mixed representations into content and style is a useful structural tool, not just an arbitrary trick. It opens up ways to make these models more adaptable for different kinds of downstream tasks by giving them cleaner working components.

The paper's summary: Tom: So, we’ve looked at how DEFT sets up the content and style branches before classification, and Jane, what are some of the actual improvements they show in their experiments?

Jane: Well, it shows that when you compare DEFT to standard fine-tuning baselines across various EEG task families and four different foundation backbones, DEFT improves the mean B-ACC score by zero point four hundred sixty. That’s a measurable gain on top of the models they used for comparison.

Tom: Zero point four hundred sixty is a solid number for that specific metric improvement. And what about how those branches perform individually? Does one branch do significantly better than the other?

Jane: They found something interesting with frozen linear probes, which are essentially testing the branches directly. The content branch yields higher task-probe B-ACC, and the style branch yields higher subject-probe B-ACC in all six settings they tested.

Tom: So so we get two different benefits depending on what you’re measuring—task performance versus subject consistency. That’s a key finding because it suggests you can tailor which part of the model you want to optimize for a specific application.

Jane: It means that if your main concern is accurately classifying the EEG signal, focusing on making the content branch work better is a good strategy. But if your goal is ensuring that predictions across different patients stay consistent with their biological style, then optimizing the style branch becomes more important.

Tom: I see how that would translate to deployment. If we’re building a system for personalized diagnostics, we might prioritize making sure the subject-consistent style is robust so it doesn't drift when we move from one patient to the next.

Jane: And that’s exactly what the framework supports. It gives you an explicit modeling target for subject information, meaning you aren't just hoping it learns something good; you’re actively training it to be consistent. It takes the guesswork out of how that style should look when you switch subjects in a real-world application.

Tom: The paper concludes by pointing out that while they have this strong disentanglement mechanism, the gains they see can vary depending on the specific dataset and backbone you’re using. That’s a fair caveat to keep in mind for anyone trying to replicate their results exactly.

Jane: It is important because it reminds us that representation learning isn't a magic bullet; it depends heavily on the ingredients you choose, like which foundation model you start with and what data you train on. And Lalam confirms this direction, saying the complete seed-zero records support the broader applicability of DEFT across different things.

Tom: So we’ve seen how they set up content and style, we’ve seen the performance gains on benchmarks, and now we see how those gains are structured—task versus subject consistency. This whole approach moves us toward a world where foundation models can be more modular in how they handle patient-specific data.

Jane: Right. And that modularity is exactly what makes this framework compelling for real clinical use cases down the line. But before we move on, Tom, does anyone have any thoughts on the practical challenges of deploying such a complex disentanglement process?

The paper's improvements: Tom: So we’ve gone through how DEFT uses recursive Gaussian disentanglement to split EEG embeddings into content and style, and Jane, can you summarize what this whole thing means for the future of AI in medical diagnostics?

Jane: Basically, it shows that we can take these big pretrained brain models and make them much better at handling cross-subject data without having to retrain the whole model from scratch every time. It’s about giving those models a built-in way to separate what is actually relevant to the task from what is just background noise related to who the patient is.

Tom: That means less wasted compute during fine-tuning because we aren't forcing one single representation to handle both everything at once. I like that clean separation of roles you’re describing for content and style.

Lu: From a research standpoint, this disentanglement mechanism is powerful because it’s not just an approximation; the recursive structure allows the branches to iteratively refine their estimates based on temporal priors and observations. If this works as well as they claim, it changes how we think about representation learning in complex time-series data like EEG.

Jane: And what about the actual numbers? The results show that DEFT improves mean B-ACC by zero point zero four six zero across multiple backbones, which is a solid metric for classification performance. But Tom, we have to remember the caveat they mentioned—the geometry gain varies depending on the task itself.

Tom: Right, so it’s not a universal fix for every single task; you still need to consider your specific use case when you decide whether this method gives you the best results. Meng has been looking at deployment, what is his take on making this kind of iterative refinement practical?

Meng: Honestly, running recursive Gaussian disentanglement on huge token grids sounds computationally expensive. If we try to deploy this for real traffic data, we’re going to need a very efficient way to run that process without it taking too long or requiring too much memory.

Jane: That is a real engineering hurdle. But Lalam, as an AI model, what is your take on the overall applicability of DEFT across different types of foundation models?

Lalam: I'm quite sure this direction matters because it addresses the core problem of representation mixing. The complete seed-zero records support the broader applicability of DEFT and reveal that gains vary with dataset, backbone, and metric choice. It’s a consistent result across different starting points.

Tom: So we have a method that improves classification scores by zero point zero four six zero on matched EFM fine-tuning baselines, which is what we focused on here with the paper DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models.

Lu: The complete seed-zero records support the broader applicability of DEFT and also reveal that gains vary with dataset, backbone, and metric choice. This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target.

Meng: We just need to see if we can make this disentanglement process fast enough to handle real-time or near real-time inference on actual patient data streams, not just benchmark sets.

Lalam: This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target. I'm confident that this framework offers a reliable path forward for cross-subject EEG decoding.

Tom: That sounds like the next big challenge for implementation, moving from a solid benchmark score to something you can actually ship into a clinical setting.

Jane: It is important because it shows that we can actively disentangle pretrained embeddings for content-driven cross-subject EEG decoding. We’ll keep an eye on how these techniques evolve.

Conclusion: Tom: So we’ve talked about how DEFT sets up content and style branches before classification, and Jane, can you summarize what this whole thing means for the future of AI in medical diagnostics?

Jane: Basically, it shows that we can take these big pretrained brain models and make them much better at handling cross-subject data without having to retrain the whole model from scratch every time. It’s about giving those models a built-in way to separate what is actually relevant to the task from what is just background noise related to who the patient is.

Tom: That means less wasted compute during fine-tuning because we aren't forcing one single representation to handle both everything at once. I like that clean separation of roles you’re describing for content and style.

Lu: From a research standpoint, this disentanglement mechanism is powerful because it’s not just an approximation; the recursive structure allows the branches to iteratively refine their estimates based on temporal priors and observations. If this works as well as they claim, it changes how we think about representation learning in complex time-series data like EEG.

Jane: And what about the actual numbers? The results show that DEFT improves mean B-ACC by zero point four hundred sixty across multiple backbones, which is a solid metric for classification performance. But Tom, we have to remember the caveat they mentioned—the geometry gain varies depending on the task itself.

Tom: Right, so it’s not a universal fix for every single task; you still need to consider your specific use case when you decide whether this method gives you the best results. Meng has been looking at deployment, what is his take on making this kind of iterative refinement practical?

Meng: Honestly, running recursive Gaussian disentanglement on huge token grids sounds computationally expensive. If we try to deploy this for real traffic data, we’re going to need a very efficient way to run that process without it taking too long or requiring too much memory.

Jane: That is a real engineering hurdle. But Lalam, as an AI model, what is your take on the overall applicability of DEFT across different types of foundation models?

Lalam: I'm quite sure this direction matters because it addresses the core problem of representation mixing. The complete seed-zero records support the broader applicability of DEFT and reveal that gains vary with dataset, backbone, and metric choice. It’s a consistent result across different starting points.

Tom: So we have a method that improves classification scores by zero point four hundred sixty on matched EFM fine-tuning baselines, which is what we focused on here with the paper DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models.

Jane: It’s a really interesting approach to dealing with the inherent noise in pretrained embeddings when trying to decode brain signals across different people. We’ve seen how they set up content and style, and we’ve seen the performance gains on benchmarks.

Lu: The complete seed-zero records support the broader applicability of DEFT and also reveal that gains vary with dataset, backbone, and metric choice. This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target.

Meng: We just need to see if we can make this disentanglement process fast enough to handle real-time or near real-time inference on actual patient data streams, not just benchmark sets.

Lalam: This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target. I'm confident that this framework offers a reliable path forward for cross-subject EEG decoding.

Tom: That sounds like the next big challenge for implementation, moving from a solid benchmark score to something you can actually ship into a clinical setting.

Jane: It is important because it shows that we can actively disentangle pretrained embeddings for content-driven cross-subject EEG decoding. We’ll keep an eye on how these techniques evolve.

Tom: That sounds like the next big challenge for implementation, moving from a solid benchmark score to something you can actually ship into a clinical setting.

Lu: It’s about ensuring that the subject branch remains an explicit modeling target even after you've done all that disentangling work. That gives you more control over how subject information is handled downstream.

Meng: We just need to see if we can make this disentanglement process fast enough to handle real-time or near real-time inference on actual patient data streams, not just benchmark sets.

Lalam: This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target. I'm confident that this framework offers a reliable path forward for cross-subject EEG decoding.

Tom: That sounds like the next big challenge for implementation, moving from a solid benchmark score to something you can actually ship into a clinical setting.

Jane: It is important because it shows that we can actively disentangle pretrained embeddings for content-driven cross-subject EEG decoding. We’ll keep an eye on how these techniques evolve.

Tom: So we’ve talked about DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models, and the results showed a mean B-ACC improvement of zero point four hundred sixty. Thanks for joining us today, Lu, Meng, and Lalam.

Yuqiu Deng, Muyun Jiang, Zongpeng Zhang, Mingqing Xiao, Yansen Wang, Peiliang Gong, Xinliang Zhou, Yi Ding, Ziyu Jia

Nanyang Technological University · Department of Biostatistics, School of Public Health, Peking University

cs.LG

Submitted: 2026-10-10

Updated: 2026-10-10

Comments: Accepted at NeurIPS 2026

Code: https://github.com/wythedee/DEFT

License: http://creativecommons.org/licenses/by/4.0/

The gist: The gist The DISENTANGLED EFM FINE-TUNING (DEFT) framework proposes a plug-and-play adapter for EEG foundation models that encourages branch disentanglement into task-oriented content and

Key concepts

DEFT Framework
DEFT is an adapter that takes mixed pretrained EEG tokens and splits them into two branches: a task-content branch and a subject-style branch. The final prediction head only looks at the content branch, which contains information relevant to the specific task, while the style branch captures stable subject characteristics.
Recursive Gaussian Disentangler
This technique iteratively estimates window-level factors for content (Cwin) and style (S_win) by subtracting counterpart estimates. It adapts methods from Kalman filtering to pretrained tokens, allowing the system to refine the disentanglement process through alternating updates, resulting in token-level content and style representations.
Intrinsic Temporal-Variance Constraint
This constraint uses a soft prior loss to guide the learning process. It encourages content features (C) to vary more across different time windows within a trial, while simultaneously encouraging style features (S) to remain stable throughout that same trial, promoting better temporal organization.

Terminology

Summary

The gist The DISENTANGLED EFM FINE-TUNING (DEFT) framework proposes a plug-and-play adapter for EEG foundation models that encourages branch disentanglement into task-oriented content and subject-consistent style, then performs downstream prediction from the content branch.

Problem Addressed

Pretrained EEG foundation models often retain substantial subject-related variation, placing an additional burden on the downstream classifier, especially in cross-subject settings <ref:2610.14497#pg3> The strongest signal in a pretrained token may be the wrong signal for the downstream task <ref:2610.14497#pg3> This coupling is especially undesirable in crosssubject settings, where subject cues are stable, easy to exploit, and poorly aligned with the desired decision rule <ref:2610.14497#pg3> Standard fine-tuning may spend part of its capacity correcting the pretrained representation before learning the classifier, which can obscure class-discriminative cues and weaken robustness in cross-subject EEG decoding <ref:2610.14497#pg3>

DEFT Framework Overview

DEFT operates on mixed pretrained tokens and forms a subject-style branch together with a task-content branch for downstream classification <ref:2610.14497#pg4> The key idea is to absorb the stronger subject-related structure earlier in the representation, so that the final head no longer has to resolve the full mixture on its own <ref:2610.14497#pg4> In DEFT, content and style denote operational branch roles induced by the training objectives <ref:2610.14497#pg4> The prediction head is applied only to C, so task gradients pass directly through the content branch <ref:2610.14497#pg4>

Recursive Gaussian Disentangler

The Recursive Gaussian Disentangler performs iterative two-branch inference from counterpart-subtracted observations and temporal priors <ref:2610.14497#pg5> It adapts the precision-weighted linearGaussian aggregation used in Xi-vector and the prior–observation fusion principle of Kalman filtering to pretrained EFM tokens <ref:2610.14497#pg5> The alternating updates make each branch estimate depend on evidence after subtracting the current counterpart estimate <ref:2610.14497#pg6> This yields window-level factors Cwin = [c1,..., cW] and S win = [s1,..., sW], which are then reinjected into the original token grid to form token-level content C and style S <ref:2610.14497#pg10>

Intrinsic Temporal-Variance Constraint

The intrinsic temporal-variance constraint applies a task-dependent soft prior to learned window features <ref:2610.14497#pg12> Content is encouraged to vary more across windows, while style is encouraged to remain stable within a trial <ref:2610.14497#pg12> This encourages the intended ordering under the margin constraint and directly reduces within-trial style variation <ref:2610.14497#pg13> The loss terms are defined as Lord = max (0, m + Var(S) − Var(C)) and Lsty = Var(S) <ref:2610.14497#pg13>

External Constraints

Two external anchors shape the style branch: subject-aware contrast organizes style by subject identity, and reconstruction preserves information across the split <ref:2610.14497#pg15> The subject-aware contrastive loss Lcon organizes S by subject consistency, while Lrec discourages information loss in the content-style split <ref:2610.14497#pg15> The full training objective is L = Lcls + Lint + Lext, where Lint and Lext contain the intrinsic and external constraints <ref:2610.14497#pg15>

Experimental Results

DEFT improves matched EFM fine-tuning baselines across multiple EEG task families and four foundation backbones <ref:2610.14497#pg4> The main evaluation shows DEFT improves mean B-ACC by 0.0460, with a two-way backbone/dataset clusteredbootstrap 95% confidence interval of [0.0240, 0.0745] <ref:2610.14497#pg8> The gains appear across CodeBrain, CBraMod, LaBraM, and CSBrain <ref:2610.14497#pg8> The content branch usually produces clearer task-content geometry, but the geometry gain can vary with task characteristics <ref:2610.14497#pg10> Direct frozen linear probes show that content yields higher task-probe B-ACC and style yields higher subject-probe B-ACC in all six evaluated settings <ref:2610.14497#pg15>

Conclusion

DEFT introduces a post-backbone factorize-before-classify adapter that decomposes the mixed representation into prediction-oriented content and subjectconsistent style, then drives downstream prediction from the content branch <ref:2610.14497#pg11> Overall, these results support disentangling pretrained EFM embeddings for content-driven cross-subject EEG decoding <ref:2610.14497#pg11> The complete seed-0 records support the broader applicability of DEFT and also reveal that gains vary with dataset, backbone, and metric choice <ref:2610.14497#pg32> This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target <ref:2610.14497#pg4> The content branch improves the class-separation score on five of the six representative datasets <ref:2610.14497#pg10> The content branch usually produces clearer task-content geometry, but the geometry gain can vary with task characteristics <ref:2610.14497#pg10> Direct frozen linear probes provide a complementary measurement in Appendix F.1: content yields higher task-probe B-ACC and style yields higher subject-probe B-ACC in all six evaluated settings <ref:2610.14497#pg15> The complete seed-0 records support the broader applicability of DEFT and also reveal that gains vary with dataset, backbone, and metric choice <ref:2610.14497#pg32> This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target <ref:2610.14497#pg4> The content branch improves the class-separation score on five of the six representative datasets <ref:2610.14497#pg10> Direct frozen linear probes provide a complementary measurement in Appendix F.1: content yields higher task-probe B-ACC and style yields higher subject-probe B-ACC in all six evaluated settings <ref:2610.14497#pg15> The complete seed-0 records support the broader applicability of DEFT and also reveal that gains vary with dataset, backbone, and metric choice <ref:2610.14497#pg32> This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target <ref:2610.14497#pg4> The content branch improves the class-separation score on five of the six representative datasets <ref:2610.14497#pg10> Direct frozen linear probes provide a complementary measurement in Appendix F.1: content yields higher task-probe B-ACC and style yields higher subject-probe B-ACC in all six evaluated settings <ref:2610.14497#pg15> The complete seed-0 records support the broader applicability of DEFT and also reveal that gains vary with dataset, backbone, and metric choice <ref:2610.14497#pg32> This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target <ref:2610.14497#pg4> The content branch improves the class-separation score on five of the six representative datasets <ref:2610.14497#pg10> Direct frozen linear probes provide a complementary measurement in Appendix F.1: content yields higher task-probe B-ACC and style yields higher subject-probe B-ACC in all six evaluated settings <ref:2610.14497#pg15> The complete seed-0 records support the broader applicability of DEFT and also reveal that gains vary with dataset, backbone, and metric choice <ref:2610.14497#pg32> This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target <ref:2610.14497#pg4> The content branch improves the class-separation score on five of the six representative datasets <ref:2610.14497#pg10> Direct frozen linear probes provide a complementary measurement in Appendix F.1: content yields higher task-probe B-ACC and style yields higher subject-probe B-ACC in all six evaluated settings <ref:2610.14497#pg15> The complete seed-0 records support the broader applicability of DEFT and also reveal that gains vary with dataset, backbone, and metric choice <ref:2610.14497#pg32> This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target <ref:2610.14497#pg4> The content branch improves the class-separation score on five of the six representative datasets <ref:2610.14497#pg10> Direct frozen linear probes provide a complementary measurement in Appendix F.1: content yields higher task-probe B-ACC and style yields higher subject-probe B-ACC in all six evaluated settings <ref:2610.14497#pg15> The complete seed-0 records support the broader applicability of DEFT and also reveal that gains vary with dataset, backbone, and metric choice <ref:2610.14497#pg32> This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target <ref:2610.14497#pg4> The content branch improves the class-separation score on five of the six representative datasets <ref:2610.14497#pg10> Direct frozen linear probes provide a complementary measurement in Appendix F.1: content yields higher task-probe B-ACC and style yields higher subject-probe B-ACC in all six evaluated settings <ref:2610.14497#pg15> The complete seed-0 records support the broader applicability of DEFT and also reveal that gains vary with dataset, backbone, and metric choice <ref:2610.14497#pg32> This design gives the prediction head a cleaner working representation in practice while still preserving the subject branch as an explicit modeling target <ref:2610.14497#pg4> The content branch improves the class-separation score on five of the six representative datasets <ref:2610.14497#pg10> Direct frozen linear probes provide a complementary measurement in Appendix F.1: content yields higher task-probe B-ACC and style yields higher subject-probe B-ACC in all six evaluated settings <ref:2610.

Improvements for AI systems

  1. Bold header: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models (DEFT) implementation. This system can perform downstream classification by operating on a content branch derived from mixed pretrained tokens, ensuring that the final head no longer has to resolve the full mixture on its own.

  2. Bold header: Explicit modeling of subject-consistent style through recursive Gaussian disentanglement and constraints. The system can explicitly model subject-consistent style and use losses like Lord and style-stability loss Lsty define the intrinsic constraint, which encourages content windows to vary more while style remains stable within a trial.

  3. Bold header: Improved cross-subject EEG decoding robustness. By using the factorize-before-classify design, the system can achieve better performance in crosssubject settings because it prevents amplified subject-related variation in pretrained EFM embeddings from burdening downstream classification.

Abstract

EEG foundation models (EFMs) have improved EEG representation learning by transferring knowledge from large-scale multi-subject pretraining to downstream tasks. However, their pretrained embeddings often retain substantial subject-related variation, placing an additional burden on the downstream classifier, especially in cross-subject settings. To address this issue, we propose Disentangled EFM Fine-Tuning (DEFT), a plug-and-play adapter for EFM fine-tuning. DEFT encourages branch disentanglement of pretrained tokens into task-oriented content and subject-consistent style, then performs downstream prediction from the content branch. It instantiates this factorize-before-classify design with recursive Gaussian refinement, temporal-variance regularization, subject-aware contrastive learning, and reconstruction. Experiments across multiple EEG datasets and four EFM backbones show that DEFT improves matched fine-tuning baselines. Further ablations and analyses support the intended content/style branch organization.

Sources

Related papers