Cross-Trait Transfer in Subliminal Learning
Not on the map. Why?
Abstract (arXiv)
Subliminal learning is a phenomenon where a student language model acquires a teacher model's behavioral traits by training on semantically unrelated outputs. It is a subtle statistical phenomenon as trait transmission relies on weak statistical patterns in the generated data. To understand trait transmission between teacher-student pairs, we study cross-trait transfer: how data generated under one teacher trait changes the student's preferences of other traits. To this end, we introduce a directed trait-transfer matrix that quantifies these effects using log-probability gains for student answers. We find that the trait-transfer matrix reveals clusters of related traits, with students sometimes developing preferences for traits similar, but not identical, to the teacher's trait. Such cross-trait structure can be partially captured by output distribution metrics and representation-based metrics. Further, we analyze trait development and interaction: learning dynamics shows a progression from broad shared shifts toward more trait-specific transfer, and multi-trait experiments suggest that opposed traits can enhance such differentiation. Together, our findings reveal salient statistical structures over trait transfer and competition, thus providing a broader view of how hidden preferences are transmitted in subliminal learning.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 21 of its 28 references that are in the map. First posted on arXiv 3 Oct 2026.
Builds on
21 of 28 references.
References · 21
- 2015Distilling the Knowledge in a Neural Networkcited by 10,324
- 2022Toy Models of Superpositioncited by 971
- 2023Steering Language Models With Activation Engineeringcited by 922
- 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 716
- 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 345
- 2018A mathematical theory of semantic development in deep neural networkscited by 186
- 2025Towards Understanding Subliminal Learning: When and How Hidden Biases Transfercited by 34
- 2025Toward universal steering and monitoring of AI modelscited by 30
- 2025Narrow Finetuning Leaves Clearly Readable Traces in Activation Differencescited by 28
- 2026Subliminal Effects in Your Data: A General Mechanism via Log-Linearitycited by 19
- 2026Subliminal Steering: Stronger Encoding of Hidden Signalscited by 14
- 2026Subliminal Learning Is Steering Vector Distillationcited by 13
9 more
- 2025Dataset distillation for memorized data: Soft labels can leak held-out teacher knowledgecited by 10
- 2026From Data to Behavior: Predicting Unintended Model Behaviors Before Trainingcited by 8
- 2026Sustained Gradient Alignment Mediates Subliminal Learning in a Multi-Step Setting: Evidence from MNIST Auxiliary Logit Distillation Experimentcited by 7
- 2026Learning Through Noise: Why Subliminal Learning Works and When It Failscited by 7
- 2026Channel Location Constrains the Auditability of Subliminal Learningcited by 7
- 2026Shattered Compositionality: Counterintuitive Learning Dynamics of Transformers for Arithmeticcited by 3
- 2026Subliminal Learning is Non-Semantic Distillationcited by 3
- 2026Stored in Optimizer State, Valued by Later Training: A Causal Account of Subliminal Trait Transfercited by 3
- 2026Subliminal Learning as Trait-Direction Drift: A Mechanism and Targeted Control under SFT Distillationcited by 1
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2026Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning
- 2026Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors
- 2025Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
- 2026On Mitigation of Subliminal Learning in Large Language Models
- 2026Scaling Model-Generated Distillation Data Can Make Latent Teacher Traits More Recoverable
- 2026Quantifying Subliminal Behavioral Transfer Ratios in Language Model Distillation
- 2026You Didn't Have to Say It like That: Subliminal Learning from Faithful Paraphrases
- 2026Reproducing and Evaluating the Generalizability of Subliminal Learning in Open-Weight Models