Sustained Gradient Alignment Mediates Subliminal Learning in a Multi-Step Setting: Evidence from MNIST Auxiliary Logit Distillation Experiment

• URL: http://arxiv.org/abs/2604.25779v1
• Date: Tue, 28 Apr 2026 15:46:18 GMT
• Title: Sustained Gradient Alignment Mediates Subliminal Learning in a Multi-Step Setting: Evidence from MNIST Auxiliary Logit Distillation Experiment
• Abstract summary: In the MNIST auxiliary logit distillation experiment, a student can acquire an unintended teacher trait despite distilling only on no-class logits.<n>Under a single-step gradient descent assumption, subliminal learning theory attributes this effect to alignment between the trait and distillation gradients.<n>We show that gradient alignment remains weakly but consistently positive throughout training and causally contributes to trait acquisition.
• Score: 1.0026496861838448
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: In the MNIST auxiliary logit distillation experiment, a student can acquire an unintended teacher trait despite distilling only on no-class logits through a phenomenon called subliminal learning. Under a single-step gradient descent assumption, subliminal learning theory attributes this effect to alignment between the trait and distillation gradients, but does not guarantee that this alignment persists in a multi-step setting. We empirically show that gradient alignment remains weakly but consistently positive throughout training and causally contributes to trait acquisition. We show that a mitigation method called liminal training works by attenuating the alignment and fails to stop trait acquisition in this setup. These results suggest that mitigation methods that operate in this regime may not reliably suppress trait acquisition when the first-order drive dominates.

Related papers

• Subliminal Learning as Trait-Direction Drift: A Mechanism and Targeted Control under SFT Distillation [6.96679232120357]
  A teacher biased by a system prompt can generate semantically clean training data, such as numeric sequences, that still causes a downstream student to inherit the hidden preference.<n>We propose and validate trait-direction drift as a mechanism for subliminal learning.<n>We propose probe-space corridor regularization, a targeted defense that constrains drift along a calibrated trait direction during distillation.
  arXiv  Detail & Related papers  (2026-09-01T11:32:53Z)
• Why Are DMD Students Lazy? Understanding the Copying Behavior in Few-Step Distillation [66.15462638744619]
  Distribution Matching Distillation (DMD) compresses pretrained diffusion models into efficient few-step generators by aligning their noised distributions across all scales.<n>Surprisingly, we find that in high-dimensional settings, distilled students spontaneously reproduce the original noise-data pairings of the teacher, a phenomenon we term copying.
  arXiv  Detail & Related papers  (2026-06-01T13:30:19Z)
• Gradient Perturbation: Learning to Perturb Gradients for Adaptive Training [3.2608238154676843]
  gradient perturbation has received little systematic investigation.<n>We propose Learning to Perturb Gradients (LPG), which adaptively perturbs logit-level gradients at the class level to achieve category-aware training.<n> Experiments on balanced classification, long-tail classification, and noisy label learning demonstrate LPG consistently outperforms existing methods and can be combined with them as a plug-in module.
  arXiv  Detail & Related papers  (2026-05-28T07:19:36Z)
• Attention Sinks Induce Gradient Sinks [41.85762183893736]
  We show that under causal mask, attention sinks can induce pronounced gradient concentration.<n>Massive activations can be understood as an adaptive response to this localized gradient pressure.<n>These results support the interpretation that gradient sink is a key training-time mediator linking attention sinks and massive activations.
  arXiv  Detail & Related papers  (2026-03-18T14:31:21Z)
• Diversity-Preserved Distribution Matching Distillation for Fast Visual Synthesis [33.341277146439275]
  Distribution matching distillation (DMD) aligns a multi-step generator with its few-step counterpart to enable high-quality generation under low inference cost.<n>We propose a role-separated distillation framework that explicitly disentangles the roles of distilled steps.<n>We term this approach Diversity-Preserved DMD (DP-DMD), which, despite its simplicity, preserves sample diversity while maintaining visual quality on par with state-of-the-art methods in extensive text-to-image experiments.
  arXiv  Detail & Related papers  (2026-02-03T05:45:25Z)
• VarDiU: A Variational Diffusive Upper Bound for One-Step Diffusion Distillation [16.15071476996734]
  Recently, diffusion distillation methods have compressed thousand-step teacher diffusion models into one-step student generators.<n>Most existing approaches train the student model using a diffusive divergence whose gradient is approximated via the student's score function.<n>We propose VarDiU, a Variational Diffusive Upper Bound that admits an unbiased gradient estimator and can be directly applied to diffusion distillation.
  arXiv  Detail & Related papers  (2025-08-28T10:47:50Z)
• Towards Training One-Step Diffusion Models Without Distillation [72.80423908458772]
  We introduce a family of new training methods that entirely forgo teacher score supervision.<n>We find that initializing the student model with the teacher's weights remains critical.
  arXiv  Detail & Related papers  (2025-02-11T23:02:14Z)
• Knowledge Distillation with Refined Logits [31.205248790623703]
  We introduce Refined Logit Distillation (RLD) to address the limitations of current logit distillation methods.<n>Our approach is motivated by the observation that even high-performing teacher models can make incorrect predictions.<n>Our method can effectively eliminate misleading information from the teacher while preserving crucial class correlations.
  arXiv  Detail & Related papers  (2024-08-14T17:59:32Z)
• Logit Standardization in Knowledge Distillation [83.31794439964033]
  The assumption of a shared temperature between teacher and student implies a mandatory exact match between their logits in terms of logit range and variance. We propose setting the temperature as the weighted standard deviation of logit and performing a plug-and-play Z-score pre-process of logit standardization. Our pre-process enables student to focus on essential logit relations from teacher rather than requiring a magnitude match, and can improve the performance of existing logit-based distillation methods.
  arXiv  Detail & Related papers  (2024-03-03T07:54:03Z)
• Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion [83.90492831583997]
  We show that a batch-normalized network can keep the optimal signal propagation properties, but avoid exploding gradients in depth. We use a Multi-Layer Perceptron (MLP) with linear activations and batch-normalization that provably has bounded depth. We also design an activation shaping scheme that empirically achieves the same properties for certain non-linear activations.
  arXiv  Detail & Related papers  (2023-10-03T12:35:02Z)
• On student-teacher deviations in distillation: does it pay to disobey? [54.908344098305804]
  Knowledge distillation has been widely used to improve the test accuracy of a "student" network. Despite being trained to fit the teacher's probabilities, the student may not only significantly deviate from the teacher probabilities, but may also outdo the teacher in performance.
  arXiv  Detail & Related papers  (2023-01-30T14:25:02Z)
• Adam: Dense Retrieval Distillation with Adaptive Dark Examples [104.01735794498767]
  We propose ADAM, a knowledge distillation framework that can better transfer the dark knowledge held in the teacher with Adaptive Dark exAMples. We conduct experiments on two widely-used benchmarks and verify the effectiveness of our method.
  arXiv  Detail & Related papers  (2022-12-20T12:03:19Z)
• The Equalization Losses: Gradient-Driven Training for Long-tailed Object Recognition [84.51875325962061]
  We propose a gradient-driven training mechanism to tackle the long-tail problem. We introduce a new family of gradient-driven loss functions, namely equalization losses. Our method consistently outperforms the baseline models.
  arXiv  Detail & Related papers  (2022-10-11T16:00:36Z)
• Why distillation helps: a statistical perspective [69.90148901064747]
  Knowledge distillation is a technique for improving the performance of a simple "student" model. While this simple approach has proven widely effective, a basic question remains unresolved: why does distillation help? We show how distillation complements existing negative mining techniques for extreme multiclass retrieval.
  arXiv  Detail & Related papers  (2020-05-21T01:49:51Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.