mlsanityData to 14 Oct 2026

Subliminal Learning Is Steering Vector Distillation

arXiv 2606.00995 ↗First posted 31 May 2026cs.AIcited by 13

On the map: Mechanistic interpretability → Activation steering methods

Abstract (arXiv)

Subliminal learning refers to a student language model acquiring a teacher's traits (e.g. a system-prompted preference for owls) when fine-tuned on the teacher's outputs, despite the outputs being semantically unrelated to those traits. It remains poorly understood how data without semantic meaning can transfer specific semantic traits. In this work, we show that subliminal learning is mediated by a single steering vector, i.e. a vector added to the model's activations. Across two open-source models, we find that the teacher's system prompt is well approximated by a steering vector, and that the student's behavior is driven by learning an aligned vector over fine-tuning. System prompts that are not well approximated by steering vectors are not subliminally learned. This is a special case of steering vector distillation, in which a student trained on the outputs of a steered teacher learns to imitate that steering. We demonstrate steering vector distillation on a range of semantic and random vectors. Adding a semantic vector to a model's activations can have both model-independent and model-specific (i.e. non-semantic) effects on its behavior, so generated data that is non-semantic can transmit a vector with semantic effects, enabling subliminal learning. This also explains why subliminal learning does not transfer between models. We find that adaptive optimizers are necessary for subliminal learning in language models: activation gradients on steered data carry a small but consistent component along the steering direction, and non-adaptive optimizers impede this by allowing outlier gradients to dominate.

Key facts. Cited by 13 papers in the map (data through 14 Oct 2026). Placed in Activation steering methods, one of 13 directions in Mechanistic interpretability. Builds on 27 of its 41 references that are in the map. First posted on arXiv 31 May 2026.

Builds on

27 of 41 references.

Closest ancestors · 4

Same direction, newest first

  1. 2026Subliminal Steering: Stronger Encoding of Hidden Signals28 Apr
  2. 2024Analyzing the Generalization and Reliability of Steering Vectors17 Jul
  3. 2023Steering Llama 2 via Contrastive Activation Addition9 Dec
  4. 2023Steering Language Models With Activation Engineering20 Aug

Other references · 23

  1. 2015Distilling the Knowledge in a Neural Networkcited by 10,324
  2. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  3. 2016Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddingscited by 1,312
  4. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  5. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
  6. 2022Toy Models of Superpositioncited by 971
  7. 2022Discovering Latent Knowledge in Language Models Without Supervisioncited by 767
  8. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 716
  9. 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 712
  10. 2023The Internal State of an LLM Knows When It's Lyingcited by 659
  11. 2024Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Trainingcited by 529
  12. 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 345
11 more
  1. 2023Taken out of context: On measuring situational awareness in LLMscited by 143
  2. 2025Convergent Linear Representations of Emergent Misalignmentcited by 71
  3. 2025Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMscited by 55
  4. 2025Towards Understanding Subliminal Learning: When and How Hidden Biases Transfercited by 34
  5. 2026Subliminal Effects in Your Data: A General Mechanism via Log-Linearitycited by 19
  6. 2025Simple Mechanistic Explanations for Out-Of-Context Reasoningcited by 10
  7. 2026Thought Virus: Viral Misalignment via Subliminal Prompting in Multi-Agent Systemscited by 9
  8. 2026From Data to Behavior: Predicting Unintended Model Behaviors Before Trainingcited by 8
  9. 2026You Didn't Have to Say It like That: Subliminal Learning from Faithful Paraphrasescited by 8
  10. 2026Subliminal Learning is a LoRA Artifactcited by 6
  11. 2025Seed-Induced Uniqueness in Transformer Models: Subspace Alignment Governs Subliminal Transfercited by 5

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 13.

13 papers.

Sort:

By area:

  1. 2026Channel Location Constrains the Auditability of Subliminal Learningcited by 7
  2. 2026Stored in Optimizer State, Valued by Later Training: A Causal Account of Subliminal Trait Transfercited by 3
  3. 2026Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligencecited by 2
  4. 2026Subliminal Learning as Trait-Direction Drift: A Mechanism and Targeted Control under SFT Distillationcited by 1
  5. 2026Towards Identifying the Dataset Biases Causing Phantom Transfercited by 0
  6. 2026Can Data Attribution Filter Out Subliminal Learning? Not Reliablycited by 0
  7. 2026Post-Training Leaves Behavioral Shadows on Unrelated Decisionscited by 0
  8. 2026Cross-Trait Transfer in Subliminal Learningcited by 0
  9. 2026Why Subliminal Learning Needs So Much Data: A Noisy Inverse View through Steering Vector Recoverycited by 0
  10. 2026Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoorscited by 0
  11. 2026Predicting Alignment Generalization with Value Representationscited by 0
  12. 2026Defence against the dark arts of distillation: Unsupervised discovery of subliminal learningcited by 0
1 more
  1. 2026Imprinting: Knowledge Distillation Can Shift Student Behaviorcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Subliminal Learning is Non-Semantic Distillation
  2. 2025Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
  3. 2026On Mitigation of Subliminal Learning in Large Language Models
  4. 2026Verbalizing Subliminal Learning Effects Using Text Optimization
  5. 2026Reproducing and Evaluating the Generalizability of Subliminal Learning in Open-Weight Models
  6. 2026Quantifying Subliminal Behavioral Transfer Ratios in Language Model Distillation
  7. 2026Sustained Gradient Alignment Mediates Subliminal Learning in a Multi-Step Setting: Evidence from MNIST Auxiliary Logit Distillation Experiment
  8. 2025Understanding (Un)Reliability of Steering Vectors in Language Models