acceptodds
Under review as a conference paper at ICLR 2027

Spectral Blind Spots: Structure, Control, and Functional Effects in Language Model Heads

Abstract

Language-model output heads can develop highly concentrated singular spectra. We investigate what shapes this concentration, how it can be controlled, and when spectral intervention improves prediction. Under matched from-scratch training, spectral skew increases as the hidden-dimension-to-vocabulary ratio falls and hidden dimension grows. An empirical relation predicts held-out architectures within this regime, while pretrained-family comparisons expose its transfer limits. We introduce Spectral Skew Regularization (SSR), which penalizes a dominant-to-weaker singular-value ratio above an empirical threshold. Its pure-flow gradient lowers the dominant endpoint and raises the weaker one; with task gradients, we derive a sufficient condition for spectral descent. Experiments demonstrate spectral control across scales. Three-seed continuations at Pythia-410M and 1B improve on task-only training. SSR outperforms spectral-norm penalties on TinyStories at 410M, but SN performs better on both TinyStories and WikiText-2 at 1B. Prediction-preserving transformations and controlled perturbations show why lower skew alone does not imply better prediction. These results establish a conditional role for selective spectral intervention, with functional benefits that depend on the model and continuation setting.

Primary area:
interpretability and explainable AI
Keywords:
Language Models, LM Head, Spectral Analysis, Rank Collapse, Spectral Regularization, Model Diagnostics

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ZSSyOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies spectral concentration in language-model output heads and analyzes its relationship with hidden dimension and vocabulary size. The paper proposes controlling spectral concentration through Spectral Skew Regularization (SSR), which applies a gated penalty. The experiments show that SSR can reduce spectral skew. However, its effectiveness depends on the specific model and dataset, and reducing spectral skew alone does not consistently improve predictive performance.

Critical strengths
  1. The paper provides several controlled comparisons and attempts to distinguish reductions in spectral skew from actual improvements in predictive performance.

  2. SSR is a simple regularization method with a reasonably motivated design, and it is evaluated across multiple model scales.

Critical weaknesses
  1. The paper is written in an extremely obscure and difficult-to-follow manner. Even researchers working at the forefront of matrix decomposition and spectral analysis may find it difficult to understand what the authors are trying to communicate.

  2. Even after reading the entire paper, I could not identify what serious problems are caused by spectral concentration or spectral skew. In the experiments, reducing spectral skew does not necessarily improve predictive performance. The authors should define the harmful consequences of this form of spectral collapse more clearly, design experiments that directly demonstrate these consequences and their mitigation, and explain the resulting observations more clearly. I believe the paper would benefit substantially from doing so.

Questions for authors

See the weakness. Overall, I do not find this to be a paper from which human readers can easily understand the main ideas or learn useful insights.

Reviewer_qWcMOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary
  1. The paper proposed to measure the skewness in LM head, using the ratio of largerest singular value over smaller ones. Experiments are done on various scale of models. It’s an interesting phenomenon to note.
  2. The paper further proposed a new regularizer that regularize on the ratio, gated by a pre-specified threshold.
  3. Experimental results shows that it can curb the growth of the ratio.
Critical strengths

The paper show an interesting observation of skewness in singular value ratio, and proposed a regularization formulation to curb the skewness.

Critical weaknesses
  1. The paper writing could be improved to make the logic flow more smooth and emphasize important findings.
  2. The paper did not introduce the comparison methods to make it self-contained, and highlight the difference
  3. The study should involve more up to date models, but so far it mostly focus on old models like GPT2, Pythia etc.
  4. The experimental results are weak and not properly reported
Questions for authors
  1. Why choose the tenth singular value as the denominator for the ratio? It seems a bit arbitrary
  2. What is near exact estimator (p, q)?
Reviewer_MJfsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the ratio of the first and kth singular values of the output head of language models, and finds an empirical relationship between this ratio rr, the model's embedding dimension DD, and vocab size VV. The paper proposes SSR, defined as �0��0� ,as a method to regularize the output head during training. They derive the gradient and sufficient conditions for reduction in r in the presence of task gradients. The paper includes experiments using SSR as a regularizer and compares it to other spectral regularization baselines.

Critical strengths

The paper is carefully drafted and avoids over claiming.

Critical weaknesses

The paper's contribution is marginal. Previous work has studied regularization methods that operate on the of the spectrum of a language model's output head by penalizing the distance between observed singular values and a target condition (orthogonalization methods push each singular value to 1, spectral normalization rescales according the largest singular value, spectral control pushes the spectrum toward a decaying prior distribution). SSR penalizes the ratio of the first singular value to the kth singular value, for some choice of k. SSR can be seen as a modification to previous methods, this is not a negative factor itself but the paper fails to make the case for why SSR is useful or interesting. There is only weak justification made for why the ratio of two singular values, and in this case specifically the first and tenth, should be an object of study in this setting. The paper's given design rationale includes invariance to uniform scaling, targets relative amplification, and only alters part of the spectrum, but it is does not explained why these properties should be desirable.

The lack of motivation for SSR is not made up for by empirical results: SSR's empirical performance is uncompelling compared to previous methods (SN) as a regularizer for training (worse on 3/4 tasks, better on 1/4 tasks). The paper finds that the ratio of the first and kth singular values may be well predicted in a regresion model using model dimension and vocab size, which define the shape of the output head matrix. However, paper does not compare the fit using other common spectral measurements like condition number or effective rank as baselines. Without this information it is not clear what using the singular value ratio buys us, or what specific information it carries that isn't present in more common metrics.

The core mathematical result of the paper, proposition 1, is just an elementary application of the quotient rule.

Questions for authors

You present the ratio of the first and kth singular values and propose SSR, a method which uses this measure. What about this ratio or the SSR method is useful/interesting? Can either be used to improve model performance, or improve our understanding of model training dynamics or other questions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.