acceptodds
Under review as a conference paper at ICLR 2027

The Saturation Regularity: Rethinking Concept-Aligned Supervision in EEG Learning

Abstract

Concept-aligned supervision aims to improve learning by making task-relevant knowledge explicit in a model's representations. However, a concept's relevance does not establish the value of further supervising it once the model already encodes it. Choosing auxiliary supervision therefore requires understanding when stronger alignment improves prediction and what information remains useful when it does not. In EEG emotion recognition, we identify the Saturation Regularity: further concept supervision can strengthen alignment without detectable predictive gains, and can even reduce accuracy. This pattern varies across FACED training configurations and CIFAR-10 training stages. Random-target controls further show that auxiliary-loss gains need not depend on semantic content. To examine what remains useful beyond class-mean alignment, we introduce ResDiag, a residual diagnostic framework combining representation decomposition, directional ablation, and ensemble-contribution analysis. ResDiag reveals that within-class concept-related residuals retain predictive information and that their encoding strength is associated with ensemble contribution. Motivated by these supervision and residual analyses, we develop SatEnsemble, a saturation-informed ensemble learning framework. It strengthens individual models with a refined training recipe, then averages checkpoint predictions to exploit complementary information. Experiments show that SatEnsemble outperforms leading EEG methods on FACED, setting a new state of the art and establishing an effective strategy for learning beyond concept-aligned supervision.

Primary area:
applications to neuroscience & cognitive science
Keywords:
LLMs, EEG

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dZUgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies when concept-aligned supervision is effective and how to improve predictive performance after its benefits saturate. The results show that stronger concept alignment does not necessarily improve performance and may even reduce accuracy. Some benefits of auxiliary supervision may also be independent of concept semantics. The authors introduce ResDiag, which combines representation decomposition, directional ablation, and ensemble-contribution analysis to show that within-class concept-related residuals still contain useful predictive information. They further propose SatEnsemble, which combines an improved training recipe with prediction averaging across checkpoints to use complementary information across models, and reports competitive results on FACED.

Critical strengths
  1. The research question is valuable. The authors examine whether further supervision of a concept still improves prediction once the model already encodes it. This addresses a scientifically meaningful question beyond simply pursuing higher accuracy. The paper also carefully reports many negative results rather than presenting only successful configurations.

  2. The experiments cover various auxiliary supervision configurations, include random-target controls in some settings, and apply Holm–Bonferroni correction to 25 primary tests. These choices provide a useful basis for examining the relationship between valence supervision and predictive performance.

Critical weaknesses
  1. The “Saturation Regularity” is currently mainly an empirical finding about additional valence supervision. Although the authors tested 25 configurations, the supervision targets remain largely focused on valence, with limited validation of other concepts such as arousal. Models may encode several concepts simultaneously, while this study primarily examines relatively prominent valence-related information. The current evidence is therefore insufficient to generalize this finding into a general regularity of concept saturation.

  2. The links between “Saturation Regularity,” ResDiag, and SatEnsemble are based mainly on the research motivation and empirical associations, with limited mechanistic explanation. The useful information in the residuals may also come from concepts that were not examined, leaving the interpretation of these findings unclear.

Questions for authors

I still have several concerns about this paper and would appreciate the authors’ responses to the following questions:

  1. The systematic validation of “Saturation Regularity” mainly focuses on valence, with only limited experiments on other concepts. Given that EEG may encode multiple emotional and physiological concepts simultaneously, how do the authors establish that this phenomenon generalizes beyond valence supervision, rather than being an empirical finding specific to valence?

  2. The links between “Saturation Regularity,” ResDiag, and SatEnsemble currently rest mainly on the research motivation and empirical correlations, without a clear mechanistic explanation. The predictive information in within-class residuals may come from other concepts that were not analyzed, differences between subjects, or stimulus information. How do the authors interpret what these residuals represent and demonstrate a direct link to concept saturation and model complementarity?

  3. Prediction averaging across checkpoints or random seeds, data augmentation, increased model depth, prototype regularization, and extended training may independently contribute to performance gains. Could the authors clarify to what extent SatEnsemble’s gains are related to concept saturation rather than these conventional training and ensemble methods? Without such a connection, SatEnsemble appears to be a relatively independent engineering improvement, lacking a clear design rationale grounded in the preceding saturation analysis. This also leaves the paper’s central line of investigation insufficiently developed.

I also have the following concerns about the presentation:

  1. Several passages are abstract and overly compressed, making the writing less natural and harder to follow. For example, expressions such as “An objective can change a representation as intended without improving the prediction it is meant to support” and “what the receiving model already represents” require readers to infer their specific meaning. More direct wording would improve clarity and readability.
  2. Figure 1 has a noticeably AI-generated appearance. Its layout could also be improved; for example, Panel B occupies more space than its content requires. I suggest reducing the overtly AI-generated aesthetic and adopting a clearer, more informative visual presentation.
Reviewer_AVDEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the effectiveness of concept-aligned auxiliary supervision on valence-based emotion recognition. The authors identify the Saturation Regularity, suggesting that further strengthening the alignment of already-encoded concepts does not necessarily improve downstream prediction and may even degrade performance. The paper introduces ResDiag, a diagnostic framework that analyzes within-class residual representations through directional ablation and ensemble contribution analysis. Building on these observations, the authors propose SatEnsemble, which combines an improved training recipe with checkpoint prediction averaging. Experiments demonstrate that SatEnsemble achieves 69.48% balanced accuracy on the FACED dataset, outperforming previously reported methods.

Critical strengths

1.The paper reported some negative results.

2.The research [perspective is interesting.

Critical weaknesses

1.The central evidence is primarily established on FACED using valence as the target concept. The experiments do not establish a consistent saturation pattern across multiple EEG datasets, concepts, and architectures. The title and framing suggest a general regularity of concept-aligned supervision in EEG learning. However, all EEG experiments are restricted to emotion-related datasets.

2.The paper demonstrates that auxiliary supervision may become ineffective or harmful, but does not provide an operational criterion for identifying saturation.

3.The correlation between residual encoding and ensemble contribution is calculated from several checkpoints, introducing substantial dependence between observations.

4.The CIFAR-10 training-stage experiment is important for supporting the broader saturation argument, but the precise auxiliary targets, architecture, and optimization settings are not sufficiently explained.

Questions for authors

1.Given that all EEG experiments focus on emotion-related tasks, what evidence supports the generalization of the proposed Saturation Regularity to EEG learning more broadly?

2.Could the authors provide a more explicit definition or measurable criterion for concept saturation?

3.Could the authors evaluate the relationship between residual encoding and ensemble contribution using more independent checkpoints or heterogeneous architectures?

Reviewer_a3K9OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes two main items: (a) ResDiag, a framework that disentangles ensemble contribution in task information and (b) SatEnsemble a saturation-informed framework that that improves models by distributing the ensembles accordingly (it is similar to a mixture-of-experts concept).

Critical strengths

The paper is oriented around a very critical question: does concept alignment improve model prediction? The authors have made a very good analysis of concept alignment and prediction testing. The “empirical saturation regularity” is a very nice finding which many times appear in the literature, but the authors here somehow quantified it. I particularly liked also the distinction between saturation and overfitting which gets sometimes confused in the literature. The authors have performed significant analysis and ablation studies to show the benefit (as claimed) of SatEnsemble.

Critical weaknesses

The authors present three main contributions, but I feel SatEnsemble is the only and strongest one. The theoretical aspect of Saturation Regularity is only systematically analysed not novel by itself and the ResDiag uses also many known components e.g. directional ablation.

The analysis in section 5.3 shows minor gains from using ensemble vs SatEnsemble strategies.

Minor: In Figure 1(a) one dot has the “Fear” label unnecessarily.

Questions for authors

Why are Saturation and ResDiag considered contributions in this work?

In Table 2, EMOD ensemble improves the baseline and it sits in the same level as SatEnsemble recipe 0.655 vs 0.658. Do we need SatEnsemble or just vanilla ensemble strategy is enough?

The gains from more checkpoint pools are small as the authors indicate. Is there any ablation to show convergence (scaling laws curve fashion)?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.