acceptodds
Under review as a conference paper at ICLR 2027

Diversity of Transformer Layers: One Aspect of Parameter Scaling Laws

Abstract

Transformers improve predictably with model scale, yet the internal mechanisms underlying this behavior remain unclear. We study Transformer residual-stream components through bias–diversity decomposition, viewing their projected outputs as an ensemble. This perspective characterizes how component accuracy, diversity, and importance jointly affect prediction quality. We further develop an information-theoretic formulation in which adding informative layers monotonically improves prediction-error bounds. We characterize departures from diminishing returns by a dependence-increase term and show that the bounds exhibit approximate diminishing marginal improvements when this term is small. This provides a layer-level mechanism qualitatively consistent with the diminishing returns observed in parameter scaling laws. Experiments across eight NLP tasks and multiple LLM families show systematic relationships among bias, diversity, and accuracy. We also find that component reweighting alters the bias–diversity balance and that recurrent layer sharing substantially reduces measured diversity. Overall, our results establish layer diversity as a useful lens for understanding how Transformer components cooperate as model depth increases.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bias-diversity, Scaling-law, Submodular, Residual Stream

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_nJu2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work analyzes transformers using their residual stream (i.e., the embedding, multi-head attention, MLP, RMSNorm, projection, and logits), where the loss is first rewritten as a sum of bias (the expected loss with respect to the true logits) and variance (the expected loss with respect to the predicted logits). The work aims to find the relationship between MSE, dias, and diversity. Experiments with different datasets and models are provided to explore this relationship. Overall, I found the writing and organization of the paper very confusing to fully understand the claims and experiments in the paper.

Critical strengths
  • The paper analyzes the MSE of logits in transformer residual streams (i.e., the embedding, multi-head attention, MLP, RMSNorm, projection, and logits) by decomposing the MSE loss between bias and diversity and then analyzes the relationship between diversity and MSE. The paper also provides some information theoretic analysis on the benefits of adding layers in transformers
  • The paper provides experiments to analyze the relationship between MSE and diversity.
Critical weaknesses
  • The paper needs major rewriting to clearly write the problem being addressed, the proposed solution and what the experiments want to validate. Currently the writing does not do a good job. More details follow: a) in Abstract, the authors directly discuss the “residual stream components” without describing what it is, then the authors describe “departures from diminishing returns” without describing what it means in this context; b) Section 3.1.1 goes directly into bias-diversity decomposition, then in Section 3.1.2, the authors separately describe the weighted case, instead of describing them together and describing before; c) Section 3.2 suddenly introduces information-theoretic bounds without fully defining the terms or motivating this section; d) In Section 5.1, there is no clear description of the experiments performed (not adding even the basic experimental details in main content is not easy to read); e) Fig. 4 is very difficult to read or interpret. Unless there is a major reorganization of the paper, I am unable to understand the paper or provide a good review.
Questions for authors

Please address the concerns raised in the weakness section.

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

This paper investigates how the diversity and complementary contributions of Transformer layers influence prediction performance and explain the diminishing returns observed in model scaling laws. The authors treat residual-stream components as an ensemble, applying bias–diversity decomposition and information-theoretic analysis to characterize their contributions to prediction error and derive conditions for diminishing marginal improvements as layers are added. Experiments across eight NLP benchmarks and multiple LLM families reveal systematic relationships among bias, diversity, and accuracy, demonstrate the effects of component reweighting, and show substantially reduced diversity in models with recurrent layer sharing, providing qualitative support for the theoretical analysis.

Critical strengths
  1. Theoretical contribution. The paper introduces a bias–diversity and information-theoretic framework to analyze Transformer layers. Its theoretical results offer useful insights into the diminishing returns observed in model scaling.

  2. Extensive empirical evaluation. The paper evaluates its findings across diverse LLM families and eight NLP benchmarks. The results provide empirical support for the theoretical analysis and demonstrate the relevance of layer diversity to model performance.

Critical weaknesses

This paper is kind out of my scope, and I found it hard to follow this paper despite authors reveal their findings and conclusion in Introdution. I don't have clear preference on the decision to this paper.

Questions for authors

none

Reviewer_5DxtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how Transformer residual-stream components contribute to model predictions through bias and diversity. It develops a theoretical analysis of component contributions and diminishing returns, and conducts experiments across multiple LLM families and eight NLP tasks. The results provide insights into layer diversity, component reweighting, and recurrent layer sharing, although the connection to parameter scaling laws remains largely qualitative.

Critical strengths

The paper analyzes transformer components in terms of bias and diversity, providing an interpretable perspective on model architecture. It also presents some theoretical analyses and empirical observations. The authors evaluate multiple open LLM families across eight NLP tasks and discuss the limitations of their work.

Critical weaknesses
  • Theorem 4 (Eq. 9) states that diversity cannot exceed bias. However, Table 2 reports a bias of 25×10225\times10^2 and diversity of 54×10254\times10^2 for MobileLLM-350M on SST-2. These reported values directly contradict the theoretical result.

  • In Appendix C.5, the proof of Theorem 8 simply argues that the theorem holds because Theorem 4 holds. However, Theorem 4 establishes an inequality, whereas Theorem 8 introduces a bias-variance-covariance identity (Eq. 22). The former does not establish the latter, leaving the proof insufficient.

  • Theorem 9 relies on conditional independence between layer outputs given the label, an assumption whose limitations are acknowledged by the authors in Appendix A.2. Furthermore, Figure 6 provides only an uncontrolled comparison across model families rather than a controlled depth-scaling experiment (Appendix A.4). Consequently, the empirical evidence supporting the proposed connection to parameter scaling laws remains limited and qualitative.

  • According to Eq. 16, MSE measures the squared difference between the model’s softmax output and the one-hot gold label. Therefore, its high correlation with accuracy (Figure 2, average Spearman correlation of 0.73) is expected. However, bias and diversity, which are central to the paper’s theoretical analysis, show considerably weaker correlations (approximately 0.4-0.5). This raises the question of whether these quantities provide much additional insight beyond what MSE already captures.

Questions for authors
  • Can you estimate ϵ\epsilon for real-world models and show whether the marginal information gain actually decreases with layer index within a single model?

  • Could you provide justification for deriving Theorem 8 from Theorem 4?

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

This paper treats the embedding, attention and MLP outputs in the residual stream as an ensemble. It applies a bias-diversity decomposition and gives information-theoretic error bounds. It proposes two reweighting methods, Unit and Map. Experiments use eight classification tasks and 32 open models. The paper links this layer view to diminishing returns in scaling laws.

Critical strengths

S1. The limitations are stated honestly in App. A. But the title and §4.3 still claim more.

S2. Unit is simple and usually helps. In Table 6, it beats Base on 27 of 32 models.

Critical weaknesses

W1. The measured bias and diversity are not the quantities in the theorems. Some results do not match each other.

My main concerns include: 1) The paper gives no formula for the measured diversity. The values in Table 2 are hard to read, for example MSE 23.0, Bias 23.3 and Diversity 12.2 for MobileLLM-125M; 2) Theorem 8 seems wrong as printed. In a simple example, Eq. 5 and the definitions under Eq. 22 give two different bias values; 3) As for the Figure 4, in the caption: the red frame shows no improvement. But AG News and SST-2 clearly go up there; 4) The accuracies in Table 2 match no row of Table 6.

W2. Lack of novelty. Also, the link to scaling laws is not tested.

My main concerns include: 1) Much of the theory follows earlier ensemble work, and the decomposition and Base come from Chang et al. 2024. §1 does not say which results are known; 2) Some terms in the bounds are never measured; 3) In the Figure 6, it shows Qwen2.5-7B at about 76, but Table 6 gives Base 80.4.

Questions for authors

Questions

Q1. How exactly is the measured diversity computed?

Q2. What is the None baseline? Many models share the same None values in Table 6. Base, Map and Unit use 2,000 labeled examples, so why is None the main comparison?

Q3. How stable is the reweighting? Please report the number of runs, the variance and the split for early stopping.

Q4. Which setting gives Figure 6 and Table 2? Why are Qwen2.5-1.5B and Qwen3-4B removed in App. D.3?

Q5. How do the bounds relate to the real error of the model?


Suggestions

Suggestion 1 Please add Base with standardized features. Unit helps most where Base is weak.

Suggestion 2 In section 4.3, it replacing layers with neurons keeps the claims valid. Please show this or drop it.

1 comment
Reviewer_JQYuOpenReview ↗

Reply to AI use in this review

Dear Authors and AC,

Because the system flagged that my review comments were written by AI: to be completely honest, my native language is not Chinese.

However, rewriting the review will take time. If you have already completed your rebuttal or are currently working on it, please continue with your rebuttal. I will make sure my rewritten review covers my current core concerns.

However, this will take time, and I hope to finish the rewrite before the deadline and carefully consider both your rebuttal and the paper.

Best regards,

Reviewer

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.