acceptodds
Withdrawn from ICLR 2027

Darwinian Competition among Transformer Attention Heads: Geometric Fitness and Functional Inheritance

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Understanding how Transformer components shape representation geometry requires causal analyses beyond changes in task performance. We introduce the Causal Competition Framework (CCF), which quantifies a unit's geometric fitness through intervention-induced changes in representation distributions, measured here by squared Wasserstein distance. We instantiate CCF as Representation-level Darwinian Head Competition (RDHC), coupling this measurement to an explicit population process with fitness-guided replication, parameter inheritance, and mutation. Recorded genealogies show branching reproduction across generations, and offspring exhibit greater functional similarity to their designated parents than to shuffled parents under mutation, supporting functional inheritance within the constructed process. Across text classification benchmarks, attention heads exhibit heterogeneous geometric influence, while progressive removal of high-fitness heads leads to nonlinear degradation in classification performance. To examine geometric compensation, we ablate one head and amplify the output of another, defining geometric rescue as a reduction in distributional distance to the intact model. We observe heterogeneous rescue effects across damaged–rescuer pairs, with some rescuer heads providing positive geometric rescue for multiple damaged heads and some damaged heads exhibiting limited rescue across candidate rescuers. Domain-wise analyses further reveal variation in rescue effects across semantic domains. These improvements concern representation distributions. Together, these results characterize heterogeneous head influence and geometric rescue, alongside evidence of replication and functional inheritance in explicitly constructed head populations.

Primary area:
learning theory
Keywords:
Darwinian Competition, Attention Head, Transformer, Language Models, Geometric Fitness, Functional Inheritance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_4ExdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose the Causal Competition Framework (CCF), which quantifies a transformer attention head’s importance by the squared Wasserstein-2 distance between activation distributions of the intact and head-ablated model. They wrap this measurement in an evolutionary algorithm, Representation-level Darwinian Head Competition (RDHC), where heads in a single transformer layer are “individuals” that reproduce proportional to their fitness, with offspring inheriting parent weights plus Gaussian mutation. The empirical section reports three findings on DBpedia text classification (with AG News and Yahoo! Answers in the appendix): (1) head importance under Wasserstein ranking is heavy-tailed, (2) heads specialize by semantic class, and (3) a pairwise “rescue” analysis, in which one head is ablated and another is amplified, reveals pair-specific compensation structure that survives after controlling for marginal row and column effects. An appendix provides formal properties of the Wasserstein criterion, the population-update protocol, and additional experimental results.

Critical strengths
  • The pairwise rescue analysis with row-column residual decomposition is a genuinely new contribution. Prior work on head importance has focused on either individual head ablation or coarse pruning. The directional pairwise rescue experiment, in which ablation of head d is partially compensated by amplifying head r by a tuned factor, is a novel direction.
  • The theoretical appendix cleanly separates general discrepancy properties from conditional geometric results, with instructive counterexamples.
  • The method is honest about what the math does and does not entail. (e.g. “a large removal effect need not make replication beneficial” (3.2), “this quantity alone does not establish functional cooperation or redundancy” (3.2), “Darwinian specialization additionally requires lineage-frequency changes” (3.4).)
  • Calibration-evaluation split is cleanly handled.
Critical weaknesses
  • The Darwinian framing does no empirical work. The inheritance result holds under neutral replication (\beta = 0), so selection pressure contributes nothing.
  • Heavy-tailed head importance and class specialization are well-established findings. The paper does not benchmark Wasserstein fitness against existing importance measures (gradient attribution, accuracy-drop, activation patching). Does it find the same important heads as less expensive techniques or are they different?
  • Experimental setup is under-specified: no model architecture, layer, N, T, \sigma, \beta, run count, or evaluation set size in the main text. Some of these details are banished to the appendix, but they are important for reproducibility.
  • The joint-ablation interaction is defined and analyzed theoretically but never measured.
  • Six generations is too few to support “evolutionary timescales” language.
  • Three near-identical English topic-classification benchmarks is a weak generalization claim.
Questions for authors
  • Does \beta > 0 produce measurably different outcomes than \beta = 0?
  • How does Wasserstein ranking correlate with standard importance measures?
  • Why is J_{i,j} defined but never reported?
Reviewer_jRN2OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an evolutionary framework for studying model components using geometric fitness and evolution, with attention heads as the main application.

Critical strengths
  • The overall idea of applying an explicit evolutionary process to model components is interesting and provides a different way to study model components.
  • The method allows studying interactions between components.
Critical weaknesses
  • There is extensive literature on anisotropy of model representations. A large movement in one direction may be unimportant while a small movement in another may strongly affect the model’s functionality. Based on this some conclusions such as geometric rescue as functional substitution or class degradations, seem a bit out of proportion.
  • The rankings for top vs bottom only show the extreme differences in rankings. There is no analysis if rank 1 is actually better than rank 2 etc so it’s hard to say if fitness function is valid. There is also no comparison with other ranking approaches, so it’s hard to infer whether it identifies anything better or different.
  • The method seems costly, there does not seem to be any cost analysis, there should be a comparison with other work and cost to effectiveness analysis.
  • Some of the findings such as sparse important heads, redundancy, specialization, and sensitivity to ablation, are already known from other methods.
  • Its not entirely clear what representation is used last, pooled etc.
  • The heredity results seem largely expected and do not strongly demonstrate meaningful evolution. For head ranking alone, a single fitness evaluation may be sufficient, so the added value of the evolutionary process is unclear.
  • The specialist or general is conditioned on specific datasets and classes, this is only applicable if the model is only trained for this purpose otherwise a head may only appear to be a specialist.
  • The analysis is restricted to attention heads, the latest circuit work focuses on finer granularities from individual neurons to edges.
Questions for authors

Please see the weaknesses

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

Summary

The authors introduce the Causal Competition Framework. This uses the squared Wasserstein distance between empirical distributions of representation vectors before and after ablation in order to measure 'geometric fitness' of individual attention heads. They also define Representation-level Darwinian Head Competition (RDHC) which is a population process in which heads reproduce according to geometric fitness. Experiments on DBpedia, AG news, and Yahoo! Answers text classification datasets ablate heads according to geometric fitness and analyze domain-conditional specialization and attention head specialization.

Recommendation

The paper is very jargon-heavy and hard to understand, and omits important details such as the models used in the experimental section. Discrepancies between the text of the experimental section and the figures make me concerned about the accuracy of the reported results. While the authors have a related work section, in the other sections of the paper they do not connect their work well to existing literature, making it hard to assess the novelty of the paper. For these reasons I recommend the paper be rejected.

Critical strengths
  • It is an interesting idea to rank attention heads by changes in representation geometry.
Critical weaknesses
  • The weakest point is the experimental section which is missing important details such as the architecture, the baseline model, and what X_sel is in the context of this experiment.
    • The reported accuracy in Section 4 is also confusing. The authors said that the baseline model achieves 93.7% accuracy (line 319) which is consistent with Figure 2. But then Figure 3 shows much lower class accuracies (baseline between 20 and 40%). I am similarly confused by Figure 7 in the Appendix (Text reports 90.4 percent accuracy, whereas Figure 7 shows class accuracies around 25%).
    • Section A.3 contains more experimental results and describes the HPC cluster but not the models/model training.
  • Language chosen in Section 3 to describe the methodology does not connect well with existing literature. For example, "By suppressing a head's module-output contribution and recomputing downstream representations with other parameters fixed". It would help to explicitly definfe this as ablation and clarify in this section what the methods of this paper add beyond other work that analyzes attention head importance using ablation.
  • Very jargon-heavy, with many terms missing definitions on first use. These issues make it challenging to assess the novelty of the contribution.
Questions for authors

Questions for the authors

Introduction

  • Line 48: "how compensation depends on the damaged-rescuer pair and input domain" What does this mean?

Section 3

Overall: I recommend looking for opportunities to state in more plain language what is going on.

  • Line 168: Could you define \Delta_h?
  • Line 170: What does "\Psi is the fixed downstream map" mean? This is the rest of the model?

Section 4

  • What are the models and architecture?
  • How was geometric fitness computed? What is X_sel in this example?
  • Which activations are used to compute geometric fitness? The logits?
  • Could you define 'rescue topology' and 'rescue capacity' on first appearance?

Additional comments

The current formatting of having some equations in-line but still with equation labels is hard to read. this comment applies to: equation 2 (line 178), equation 4 (line 188), equation 6 (line 201), equation 8 (line 210), equations 10-12.

Related work

  • I recommend re-reading this section and looking for opportunities to reduce redundancy. For example, Conmy et al is cited four different times. Perhaps 2.1 and 2.2 can be combined.

Section 3

I recommend rewriting the first paragraph to reduce the jargon.

Appendix A

It would be helpful to introduce some of the results in section A.1 with an explanation of what they contribute to the paper.

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

Related papers

Discussion (0)

Sign in to comment.