acceptodds
Under review as a conference paper at ICLR 2027

Why Ghost Outputs Teach: A Kernel-Based Understanding of Subliminal Learning

Abstract

Subliminal Learning (SL) is a recently identified phenomenon in which a student model acquires downstream task capabilities by matching seemingly unrelated auxiliary outputs from a teacher, despite never observing task labels, task-specific outputs, or the original training data. While recent studies have identified where subliminal signals may reside, the optimization mechanism underlying this phenomenon remains poorly understood. In this work, we provide a mechanistic understanding of SL through the lens of learning dynamics. Specifically, we derive a chained cross-task kernel that explicitly links ghost-output supervision to changes in task predictions through shared backbone representations. Our unified analytical framework provides a rigorous mathematical explanation for three central empirical puzzles in SL: (i) under shared initialization, the transfer operator forms a strictly Positive Semi-Definite (PSD) structure, guaranteeing that ghost-output optimization aligns the student with the teacher's true task objective without explicit label exposure; (ii) the ghost-output dimensionality acts as an explicit rank bottleneck governing the transfer of task-relevant features; and (iii) synthetic, high-entropy inputs function as broadband probes that maximize cross-task kernel overlap, explaining why random noise consistently outperforms structured data for subliminal transfer. Experiments on the canonical ghost-output setting validate all three theoretical predictions, providing the first learning-dynamics-based theoretical explanation of how ghost-output supervision gives rise to SL.

Primary area:
interpretability and explainable AI
Keywords:
Subliminal Learning, Kernel, Learning Dynamics, Knowledge Distillation, NTK, LLMs, Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7bmAOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an explanation of subliminal learning through an analysis of learning dynamics. Explicitly expanding out a gradient training step, they identify a cross-task kernel which captures the influence of any ghost-output on the student's task prediction. They analyze this kernel and use it to understand different phenomena in subliminal learning, and verify their theory on a toy MNIST model and a larger LLM.

Critical strengths

The exposition and derivation steps are clear, with the theory fully contained in the main body. The problem is well motivated, and the mechanism seems like a reasonable explanation.

Critical weaknesses

While the approach seems like it could work, as it stands now, the theoretical derivations are missing crucial justifications and make informal assumptions which are not obviously true, and are insufficient to defend the strong claims that build on these. For example, taking only the first-order approximation of the gradient update step in Equation 11 is loosely justified by stating the student and teacher will be close in parameter space because they share an initialization and are trained in the overparametrized regime, but the teacher may have undergone substantial training already, so this gap may not be small. This needs a formally stated assumption, and should be verified in the experiments. Relatedly, the analysis for the extension to multistep from the single step case is not clear. The benefit of probing with random tokens also warrants stronger theoretical justification and empirical verification.

The work is in the right direction, but it feels disingenuous to say that this paper gives rigorous mathematical guarantees of subliminal learning at its current state.

Questions for authors
  1. Can you guarantee PSD alignment still holds if you include the higher-order remainder terms in equation 11 (making necessary assumptions on their size)? I'm assuming yes, but you should make those assumptions clear.

  2. For the multistep case, it is written that Equation 21 implies long term transfer because "Over many iterations, the zero- mean perturbation and interference terms average out". Why must either term necessarily be zero-mean?

  3. Can you ensure that all these approximations do not accumulate or interact in a bad way? While each assumption is not completely unreasonable, the errors together can compound—it is not clear whether this is the case here following the informal analysis.

  4. For the claim that Gaussian noise is always preferred: why does having full-rank input covariance necessarily imply large overlap of the Jacobians for the task and synthetic sample (nonzero/large kernel)? For example, two individual samples (one from MNIST, one from Gaussian) can activate completely disjoint sets of neurons. It seems to me there is no fundamental guarantee that the Jacobians will overlap just by sampling from Gaussian.

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

This paper studies subliminal learning, which means that a student model can acquire task-relevant capabilities or behavioral traits through unrelated supervision. This is an interesting topic. Specifically, it derives a chained cross-task kernel to explain how ghost-output matching influences task predictions through shared backbone parameters. Moreover, it combines theoretical analysis with experiments to investigate three questions: why ghost matching enables transfer, how ghost-output dimensionality affects transfer, and why synthetic noise can outperform real data as a distillation input.

Critical strengths
  1. The paper addresses an interesting and relevant problem.
  2. The ablations study initialization, trainable model components, ghost-output dimensionality, and distillation input distributions. The experiments on MNIST and an LLM provide complementary observations that help understand empirical transfer behavior.
  3. The kernel formulation offers a useful analytical perspective.
Critical weaknesses
  1. Figure 1 is not clear enough.
  2. The motivation behind raising Q1-Q3 is not clear. Why these questions are important?
  3. The multi-step derivation appears to omit an essential contribution from the student’s accumulated displacement. Eq. 12 depends on θiS−θT\theta_i^S-\theta^T, whereas Eq. 19 only expresses the teacher’s displacement from initialization. Substituting Eq. 19 should therefore retain a contribution involving θiS−θ0\theta_i^S-\theta_0, which is absent from Eq. 20.
Questions for authors
  1. The role of shared initialization is important but not sufficiently disentangled in the paper. Could the authors independently reinitialize the backbone, ghost head, and task head, and report both ghost-matching performance and downstream accuracy?
Reviewer_MkobOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper builds on the work of Cloud et al. on subliminal learning. The papers uses NTK-based approximations to restate versions of lemma 1 of Cloud et al in logit space, which only gives additional information about the subspace of the signal.

Critical strengths

None.

Critical weaknesses

This paper contains no theorem, no explicit assumptions, and no proofs. Each "guarantee" is a Taylor expansion with dropped higher-order terms. Eq 18 is thm 1 of Cloud et al. written in logit space with a fixed input. The experiments are quite messy, with inconsistencies in the reported numbers. The explanations for many of the observed results don't make a whole lot of sense. For example, the explanation of the broadband probe experiment seems to suggest "Jacobians nearly orthogonal to those of the task data" where the task data is MNIST itself.

Questions for authors

What do the authors see as the central contribution beyond Cloud et al's lemma 1?

Can the authors better explain the LLM case?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.