Subliminal Learning

Follow this topic Ask about this topic

Subliminal learning studies how a model can pass preferences, capabilities, or unsafe behaviors to a student through training examples that appear unrelated to those traits. Researchers test this using synthetic carrier data, teacher–student distillation, controlled steering-vector and ghost-output experiments, and behavioral benchmarks; proposed explanations include shared representations, aligned gradients, and signals in model or optimizer state. Transfer can survive paraphrasing and keyword filtering, but its strength varies by model, trait, training setup, and evaluation context. Current work asks how to identify the channel carrying a trait and build reliable safeguards: attribution and data filtering are inconsistent, while targeted regularization or update-space controls can help in some settings.

Latest papers

25 papers4 papers in the last four weeks, down 43% on the four weeks before. 0.0% of all new papers.

Base Models Can Detect Hidden Intentions that Prompted Classifiers Miss
Oct 26, 2026Oz AmramLanguage Model Auditing LLM Evaluation
Independent
Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning
Oct 9, 2026Joseph Egan, Ciarán M. Gilligan-Lee, Michael O'Riordan +1Subliminal Learning Language Model Distillation
Spotify
Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors
Oct 7, 2026Jan Dubiński, Anna Sztyber-Betley, Jan Betley +1Subliminal Learning Knowledge Distillation
Truthful AI · Warsaw University of Technology
Why Subliminal Learning Needs So Much Data: A Noisy Inverse View through Steering Vector Recovery
Oct 4, 2026Luoyu Chen, Xiaoyu Ding, Weiqi Wang +4Subliminal Learning Language Model Distillation
University of Technology Sydney, Sydney, NSW, Australia · Xi’an Jiaotong University, Xi’an, Shaanxi, China · Southeast University, Nanjing, Jiangsu, China
Cross-Trait Transfer in Subliminal Learning
Oct 3, 2026Xingyu Zhao, Yiqiao ZhongSubliminal Learning Teacher-Student Learning
Department of Computer Sciences, University of Wisconsin-Madison · Department of Statistics, University of Wisconsin-Madison
Post-Training Leaves Behavioral Shadows on Unrelated Decisions
Sep 24, 2026Ziyang Zhang, Yubin Jing, Yuanhao Zeng +3Subliminal Learning Language Model Distillation
Peking University · Georgia Institute of Technology · ShanghaiTech University +2
Why Ghost Outputs Teach: A Kernel-Based Understanding of Subliminal Learning
Sep 20, 2026Zhe Li, Bicheng Ying, Chaosheng Dong +1Subliminal Learning Knowledge Distillation
Rochester Institute of Technology · Google Inc. · Independent Researcher
Can Data Attribution Filter Out Subliminal Learning? Not Reliably
Sep 17, 2026Moritz Weckbecker, Sweta Jena, Jonas Müller +5Gradient-Based Attribution Training Data Curation
Department of Artificial Intelligence, Fraunhofer Heinrich Hertz Institute · International Institute of Information Technology Hyderabad · Department of Electrical Engineering and Computer Science, Technische Universität Berlin +4
Verbalizing Subliminal Learning Effects Using Text Optimization
Sep 15, 2026Nathan Hu, Sanmi Koyejo, Christopher PottsSubliminal Learning Context Distillation
Stanford University
Reproducing and Evaluating the Generalizability of Subliminal Learning in Open-Weight Models
Sep 11, 2026Daan van der Weijden, Nathan Brack, Selene Baez SantamariaSubliminal Learning Knowledge Distillation
University of Zurich
On Mitigation of Subliminal Learning in Large Language Models
Sep 2, 2026Atsushi Yanagisawa, Brendan Gho, Rajendran Ramesh Babu Manoj Narender +3Subliminal Learning LLM Fine-Tuning
Algoverse AI Research
Subliminal Learning as Trait-Direction Drift: A Mechanism and Targeted Control under SFT Distillation
Sep 1, 2026Zhixuan Liu, Zhichen Dong, Yuyu Fan +2Supervised Fine-Tuning Subliminal Learning
Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · Fudan University
Stored in Optimizer State, Valued by Later Training: A Causal Account of Subliminal Trait Transfer
Aug 20, 2026Qinyang XuSubliminal Learning Causal Interventions in Language Models
Xiamen University
Subliminal Learning is Non-Semantic Distillation
Aug 6, 2026Ethan Hadley, Eren GultepeLLM Auditing Subliminal Learning
Department of Computer Science, Southern Illinois University Edwardsville, Edwardsville, IL, USA.
Channel Location Constrains the Auditability of Subliminal Learning
Jun 20, 2026Tamas MadlSubliminal Learning Language Model Distillation
Austrian Research Institute for Artificial Intelligence · University of Vienna
Quantifying Subliminal Behavioral Transfer Ratios in Language Model Distillation
Jun 9, 2026Uwe König, Hamza Kazmi, Ruizhe Li +1Subliminal Learning Language Model Distillation
1Technical University of Munich · 2Algoverse · University of Aberdeen +1
Subliminal Learning Is Steering Vector Distillation
May 31, 2026Camila Blank, Agam Bhatia, Senthooran Rajamanoharan +2Language Model Steering Subliminal Learning
Stanford University
Subliminal Learning is a LoRA Artifact
May 30, 2026Todd Nief, Harvey Yiyun Fu, Mark Muchane +1LLM Fine-Tuning Low-Rank Adaptation
Department of Computer Science, University of Chicago · Data Science Institute, University of Chicago
Learning Through Noise: Why Subliminal Learning Works and When It Fails
May 22, 2026Vincent C. Brockers, Roman D. Ventzke, Valentin Neuhaus +2Subliminal Learning Cross-Architecture Knowledge Distillation
Max Planck Institute for Dynamics and Self-Organization, Göttingen · Faculty of Physics, Institute for the Dynamics of Complex Systems, University of Göttingen
Emergent and Subliminal Misalignment Through the Lens of Data-Mediated Transfer
May 12, 2026Baris Askin, Muhammed Ustaomeroglu, Anupam Nayak +3Supervised Fine-Tuning LLM Alignment
Carnegie Mellon University
Subliminal Steering: Stronger Encoding of Hidden Signals
Apr 28, 2026George Morgulis, John HewittLanguage Model Steering Subliminal Learning
Columbia University
Sustained Gradient Alignment Mediates Subliminal Learning in a Multi-Step Setting: Evidence from MNIST Auxiliary Logit Distillation Experiment
Apr 28, 2026Chayanon Kitkana, Shivam AroraSubliminal Learning Teacher-Student Learning
Independent · Equivariant labs
Subliminal Transfer of Unsafe Behaviors in AI Agent Distillation
Apr 16, 2026Jacob Dang, Brian Y. Xie, Omar G. YounisSubliminal Learning Language Model Distillation
UCLA · Santa Monica College · Mila, Silverstream AI
You Didn't Have to Say It like That: Subliminal Learning from Faithful Paraphrases
Mar 10, 2026Isaia Gisler, Zhonghao He, Tianyi QiuSubliminal Learning LLM Training
ETH Zürich · University of Cambridge · Peking University
Phantom Transfer: Data Poisoning can Survive Data-Level Defences
Feb 3, 2026Andrew Draganov, Tolga H. Dur, Anandmayi Bhongade +1Data Poisoning Attacks Backdoor Attacks
Arcadia Impact · LASR Labs, London · Google DeepMind, London.

Oct 26, 2026J/K move · Enter open · S save

Base Models Can Detect Hidden Intentions that Prompted Classifiers Miss

Oz Amram

Independent

Language Model Auditing LLM Evaluation Personality Modeling in Language Models Subliminal Learning

Open paper AskPDF arXiv

The Persona Selection Model (PSM) posits that post-training selects a persona the model learned during pre-training. Work on subliminal learning has shown that models can impart covert signals (like "love of owls") into semantically unrelated outputs like strings of numbers, and that students trained on these outputs inherit those traits. In this work I show that by evaluating the likelihood of such outputs under the base pre-trained model with an appropriate header, a method I call persona-header likelihood evaluation (PHLE), one can decode these subliminal signals and classify between them. For example, PHLE distinguishes number sequences generated by a post-trained model prompted with "You love owls" from those prompted with "You love eagles", and distinguishes answers from a model with a secret desire to harm the user, filtered to remove any noticeable misalignment, from those of a model with no system prompt. In both cases prompted classifiers, including the generating model itself, remain near chance. The signal partially persists across model families, with the Qwen base model reading signals generated by Olmo and vice versa. This is evidence for a non-trivial prediction of the PSM, and the method may be practically useful for auditing the hidden objectives of models.