acceptodds
Under review as a conference paper at ICLR 2027

When Are Spectral Modes Causal? Identification Before Intervention in Learning Dynamics

Abstract

Spectral transitions (outliers in the loss Hessian or in parameter-update matrices) are increasingly read as mechanistic signatures of feature learning and grokking, but such readings are observational: whether an emerging spectral mode controls feature acquisition or merely marks it is rarely tested. We give a preregistered, intervention-based answer. We argue a spectral direction is a candidate causal object only after passing an explicit identification stage (prospective detection, external separation, projector stability, and a quantitative feature-specificity gate), and that whether it is manipulable then depends on the operator in which curvature is measured. Our central finding is a clean dissociation: the causal object exists but lives in different spaces for different problems. In an analytic teacher-student model, prospectively suppressing the detected full-Hessian outlier slows feature-alignment growth by versus a norm-matched-random floor ( seeds, ), with a monotone dose-response, an energy-matched control isolating direction from removed update amount ( effect), a positive-curvature-branch localization, and a theory-predicted / dissociation with placebo interactions equivalent to null. In deep networks the parameter-space operator no longer suffices: across four grokking modular-addition transformers the prominent, pre-generalization, feature-aligned full-Hessian mode is not an identifiable target (no rank in - jointly clears separation, stability, and specificity, under both the full Hessian and a validated true Gauss-Newton operator), and a planted-object calibration (recovery once a spike clears the bulk by , zero false positives) shows this is a genuine identification gap, not an insensitive assay. Yet the feature is not unidentifiable, only misplaced: a function-space input-sensitivity operator recovers a deep-net mode that passes every gate and is causally manipulable, its targeted suppression delaying generalization (median steps) far beyond norm-matched-random, energy-matched, and wrong-bits controls (censoring-robust rank test against each, seeds). A planted-object calibration confirms this operator is not blind on transformers; it simply finds no compact object there under any tested operator, including the symmetry-adapted per-frequency operator the task most invites, so the feature is distributed rather than a detector artifact. Identification before intervention thus does more than reject over-readings of spectral edges: it tells you which operator exposes the causal object, and flags when no tested operator yields a compact one. All claims are preregistered with hashed-before-results protocols.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic interpretability, learning dynamics, grokking, causal intervention, loss landscape geometry, Hessian spectrum, feature learning, spectral analysis, preregistered evaluation, generalization.

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_5GeHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates whether prominent spectral modes in neural network training are causally responsible for feature learning or merely correlate with it. The authors propose an identification-first framework that requires spectral modes to satisfy separation, stability, and feature-specificity criteria before performing targeted interventions. In a teacher-student model, suppressing the identified Hessian direction significantly slows feature learning compared with matched controls. However, in deeper networks, prominent Hessian modes do not necessarily correspond to task-relevant features. The authors show that a function-space input-sensitivity operator can identify a causally important direction in a parity MLP, while no compact identifiable mode is found in the tested modular-addition transformers.

Critical strengths
  • The paper goes beyond observing spectral modes by directly suppressing them during training to test their causal role in feature learning.

  • The authors compare targeted suppression against random directions, Hessian bulk directions, and energy-matched controls, helping isolate direction-specific effects.

  • The framework explicitly evaluates spectral separation, stability, and feature specificity before intervention.

Critical weaknesses
  1. I think the paper only partially answers the question in its title. The authors propose criteria to identify spectral directions and then test their causal role through interventions, but these criteria do not necessarily characterize when a spectral mode is causal. In particular, a direction could fail the identification criteria and still play a causal role in learning. Therefore, I see the contribution more as a framework for testing causality rather than an explanation of when spectral modes are causal.

  2. The strongest causal evidence comes from controlled teacher–student and sparse-parity settings. In the modular-addition transformers, no compact spectral subspace satisfies all identification criteria under the tested operators. Consequently, the applicability of the proposed framework to larger, more complex networks and realistic learning tasks remains uncertain.

  3. Some operational details are not sufficiently explicit in the main text, particularly the definition of grokking onset, the selection of Hessian bulk-control directions, and the precise conditions triggering prospective interventions. Although additional details are referenced in the supplement, clearer descriptions would improve the reproducibility and interpretation of the experiments.

  4. In the function-space experiments, spectral modes are identified using the average gradient outer product (AGOP), whereas the intervention direction is constructed from the parameter gradient of mean directional input sensitivity. The relationship between these two quantities is not fully established, making it unclear whether the intervention directly controls the identified spectral mode or a related feature-dependent direction

  5. The framework relies on predefined thresholds for spectral separation, stability, and feature specificity. Although the authors provide calibration and sensitivity analyses, it remains unclear how these thresholds should be determined in practice when applying the framework to a new learning problem. In particular, the paper would benefit from clearer guidance on how to calibrate the identification criteria without relying on task-specific knowledge or extensive preliminary experiments.

Questions for authors

How exactly is the grokking transition defined?

Why is the candidate rank range restricted to 2–50 in the transformer analysis? Could relevant causal structure exist at higher ranks?

How are Hessian bulk directions selected? Are they sampled randomly from the bulk or chosen according to a specific criterion?

How can feature specificity be evaluated in real-world networks where the ground-truth feature direction is unknown?

Can you elaborate in how do you answer the question in your title?

Reviewer_9E71OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

At a high level, this paper is interested in operators derived from learning tasks (e.g. the Hessian, but not only). It is gives 4 criteria for "task-specific spectral prominence" associated with a task with the idea that once this spectral prominence is established, one can test whether the corresponding subspace "causes" the learning of the task by running experiments where these directions are suppressed and see if the network still manages to learn the task. The paper presents one example where the 4 criteria hold for the Hessian and the "causality" can be established (sec 4, 5), two examples where at least one of the criteria fail (transformer for modular arithmetic in sec 6 and MLP for a bit parity problem in sec 8), finally they consider a different operator than the Hessian for the task of sec 6, 8 showing that for this operator a causality study is possible for the problem of sec 6.

Critical strengths

The spectrum of operators gives a mathematically elegant and practically useful tool to handle some of the questions raised in this paper.

Critical weaknesses
  1. Objectives are not clear. 1.a) I don't understand what the paper is meant to show. One the one hand, we have these 4 criteria (sec 3) which kind of make sense but seem quite arbitrary, and the question of whether the criteria are met in certain particular problem. The answer seems to be mostly "no". What does this tell us? On the other hand, we have this question of "causality" (I'm not a big fan of this terminology in this context, it carries a lot of baggage that does not apply to this situation) that seemingly can only be tackled in the cases where 4 ad-hoc criteria as satisfied. Why? Is there a connection between these criteria and the notion of "causality" (whatever it means)?

1.b) There are many negative results in this paper essentially showing that one of the criteria does not hold). Negative results are important, but only if they teach us something new, e.g. closing research avenues that many people thought might work. This is not the case here. and the reader is left wondering what sec 6, 8 and the end of sec 9 are for.

1.c) What does spectral prominence ⇏\not\Rightarrow geometric identifiability ⇏\not\Rightarrow manipulability mean? Note that logically ⇏\not\Rightarrow is not transitive. It is possible to have A⇏B⇏CA\not\Rightarrow B\not\Rightarrow C and A⇒CA\Rightarrow C. So this chaining makes no sense. Irrespective of the logic, you're not defining any of these terms with sufficient precision for this to actually mean anything.

  1. The "Theory" of sec 4 is underwhelming. What is the point of using an analytically tractable model if you can't be really precise about what you're trying to show.

  2. The AI-assisted writing is really detrimental to the readability of the paper. 3.a) The abstract is way too long and there is a huge overlap with sec 1. This is not what an abstract is supposed to be.

3.b) The Contributions sections is unreadable.

3.c) Essential terminology is missing, e.g. uHu^H on line 195 is undefined. It took the free version of chatGPT 1m6s to infer a meaning for this. In good scientific writing, the reader is told what symbols mean.

3.d) Choice of vocabulary: p2 "The message is constructive", the term "compact" is a terrible choice for low-dimensional, especially in the context of operators (compact operators has a very specific meaning), line 377 "load-bearing"...really? etc

Questions for authors

Can you explain clearly what the purpose of the paper is?

Reviewer_TB4nOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper asks an important scientific question in machine learning: whether a spectral outlier actually controls feature learning or only correlate with the feature learning. The authors used a single-neuron teacher–student model, four modular-addition transformers, and a sparse-parity MLP to study this problem. The paper concludes that we should choose the right operator to identify a useful causal spectral object. There are in total four requirements before intervening on a spectral subspace: prospective detection, separation from the surrounding spectrum, stability, and alignment with a known task-relevant feature.

Critical strengths

None

Critical weaknesses
  1. There are several mathematical issues in the paper. For instance, in section 9, the authors used g(θ)g(\theta) for intervention form whereas in AGOP, people usually consider E[(u⊤∇xf(x,θ))2]\mathbb{E}[(u^\top \nabla_x f(x,\theta))^2] to measure directional energy. This requires a clarification or additional comparison.

  2. This paper misses many closely related work in the reference and literature review. For example, there are many results showing rigorous spectral-dynamics results and grokking mechanism: Ben Arous et al. 2024 and 2025; Kumar et al. 2024; Lyu et al. 2024; Mohamadi et al. 2024. There are many papers in feature learning which already connect learned spectral structure to target alignment and prediction performance, e.g., Ba et al. 2022; Wang et al. 2023; Cui et al. 2024; Wang et al. 2024; Moniri et al. 2024; Dandi et al. 2025; Dandi et al. 2024. Song et al. 2025; Zakarin and Singh 2025 already manipulate directly on Hessian-defined training subspaces.


Ba et al. 2022. High-dimensional Asymptotics of Feature Learning: How One Gradient Step Improves the Representation

Song et al. 2025. Does SGD really happen in tiny subspaces?

Ben Arous et al. 2024. High-dimensional SGD aligns with emerging outlier eigenspaces

Ben Arous et al. 2025. Local geometry of high-dimensional mixture models: Effective spectral theory and dynamical transitions

Wang et al. 2023. Spectral Evolution and Invariance in Linear-width Neural Networks

Wang et al. 2024. Nonlinear spiked covariance matrices and signal propagation in deep neural networks

Moniri et al. 2024. A Theory of Non-Linear Feature Learning with One Gradient Step in Two-Layer Neural Networks

Dandi et al. 2024. How Two-Layer Neural Networks Learn, One (Giant) Step at a Time

Cui et al. 2024. Asymptotics of feature learning in two-layer networks after one gradient-step

Dandi et al. 2025. A Random Matrix Theory Perspective on the Spectrum of Learned Features and Asymptotic Generalization Capabilities

Zakarin and Singh. 2025. Accelerating Neural Network Training Along Sharp and Flat Directions

Lyu et al. 2024. Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokking

Kumar et al. 2024. Grokking as the Transition from Lazy to Rich Training Dynamics

Mohamadi et al. 2024. Why Do You Grok? A Theoretical Analysis on Grokking Modular Addition

Questions for authors

I think the arxiv link of The full spectrum of deepnet hessians at scale: Dynamics with sgd training and sample size by Vardan Papyan is incorrect.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.