mlsanityData to 7 Oct 2026

Subliminal Effects in Your Data: A General Mechanism via Log-Linearity

arXiv 2602.04863 ↗First posted 4 Feb 2026cs.LG cs.AI cs.CLcited by 15

Not on the map. Why?

Abstract (arXiv)

Training modern large language models (LLMs) has become a veritable smorgasbord of algorithms and datasets designed to elicit particular behaviors, making it critical to develop techniques to understand the effects of datasets on the model's properties. This is exacerbated by recent experiments that show datasets can transmit signals that are not directly observable from individual datapoints, posing a conceptual challenge for dataset-centric understandings of LLM training and suggesting a missing fundamental account of such phenomena. Towards understanding such effects, inspired by recent work on the linear structure of LLMs, we uncover a general mechanism through which hidden subtexts can arise in generic datasets. We introduce Logit-Linear-Selection (LLS), a method that prescribes how to select subsets of a generic preference dataset to elicit a wide range of hidden effects. We apply LLS to discover subsets of real-world datasets so that models trained on them exhibit behaviors ranging from having specific preferences, to responding to prompts in a different language not present in the dataset, to taking on a different persona. Crucially, the effect persists for the selected subset, across models with varying architectures, supporting its generality and universality.

Key facts. Cited by 15 papers in the map (data through 7 Oct 2026). Builds on 38 of its 51 references that are in the map. First posted on arXiv 4 Feb 2026.

Builds on

38 of 51 references.

References · 38

  1. 2022Training language models to follow instructions with human feedbackcited by 15,712
  2. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  3. 2013Intriguing properties of neural networkscited by 5,114
  4. 2013Efficient Estimation of Word Representations in Vector Spacecited by 4,346
  5. 2021Finetuned Language Models Are Zero-Shot Learnerscited by 3,747
  6. 2022Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedbackcited by 3,634
  7. 2022Locating and Editing Factual Associations in GPTcited by 2,609
  8. 2016Concrete Problems in AI Safetycited by 1,582
  9. 2019Adversarial Examples Are Not Bugs, They Are Featurescited by 950
  10. 2016Bag of Tricks for Efficient Text Classificationcited by 869
  11. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 679
  12. 2015Generating Sentences from a Continuous Spacecited by 607
26 more
  1. 2022Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Spacecited by 522
  2. 2021Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpuscited by 481
  3. 2023DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretrainingcited by 385
  4. 2020Noise or Signal: The Role of Image Backgrounds in Object Recognitioncited by 342
  5. 2023Emergent Linear Representations in World Models of Self-Supervised Sequence Modelscited by 313
  6. 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 303
  7. 2023In-Context Learning Creates Task Vectorscited by 302
  8. 2023Function Vectors in Large Language Modelscited by 292
  9. 2023Poisoning Language Models During Instruction Tuningcited by 254
  10. 2023Scalable and Transferable Black-Box Jailbreaks for Language Models via Persona Modulationcited by 202
  11. 2023Linearity of Relation Decoding in Transformer Language Modelscited by 190
  12. 2022Datamodels: Predicting Predictions from Training Datacited by 186
  13. 2023A Pretrainer's Guide to Training Data: Measuring the Effects of Data Age, Domain Coverage, Quality, & Toxicitycited by 180
  14. 2014word2vec Explained: deriving Mikolov et al.'s negative-sampling word-embedding methodcited by 109
  15. 2024Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedbackcited by 107
  16. 2023Instructions as Backdoors: Backdoor Vulnerabilities of Instruction Tuning for Large Language Modelscited by 100
  17. 2025Subliminal Learning: Language models transmit behavioral traits via hidden signals in datacited by 89
  18. 2025Persona Features Control Emergent Misalignmentcited by 76
  19. 2026The Assistant Axis: Situating and Stabilizing the Default Persona of Language Modelscited by 76
  20. 2024Covert Malicious Finetuning: Challenges in Safeguarding LLM Adaptationcited by 74
  21. 2024Instruction Backdoor Attacks Against Customized LLMscited by 50
  22. 2025Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMscited by 47
  23. 2025Towards Understanding Subliminal Learning: When and How Hidden Biases Transfercited by 30
  24. 2025MAGIC: Near-Optimal Data Attribution for Deep Learningcited by 20
  25. 2025Sequences of Logits Reveal the Low Rank Structure of Language Modelscited by 8
  26. 2025Provably Learning from Modern Language Models via Low Logit Rankcited by 3

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 15.

15 papers.

Sort:

By area:

  1. 2026Subliminal Learning Is Steering Vector Distillationcited by 9
  2. 2026Learning Through Noise: Why Subliminal Learning Works and When It Failscited by 6
  3. 2026Channel Location Constrains the Auditability of Subliminal Learningcited by 5
  4. 2026Covert Influence Between Language Modelscited by 4
  5. 2026Scaling Model-Generated Distillation Data Can Make Latent Teacher Traits More Recoverablecited by 1
  6. 2026Sycophantic Agreement Transfers with Neutral Data via Contrastive Preference Optimizationcited by 1
  7. 2026Adversarial Concept Search: Predicting Compositional Errors From Feature Geometrycited by 0
  8. 2026Inference-Time Consensus for Mitigating Hidden Behaviors from LLM Fine-Tuningcited by 0
  9. 2026Hidden Threat in Synthetic Data: Covert Targeted Bias Injection through Benign Textcited by 0
  10. 2026Subliminal Learning as Trait-Direction Drift: A Mechanism and Targeted Control under SFT Distillationcited by 0
  11. 2026Towards Identifying the Dataset Biases Causing Phantom Transfercited by 0
  12. 2026On Mitigation of Subliminal Learning in Large Language Modelscited by 0
3 more
  1. 2026Why Ghost Outputs Teach: A Kernel-Based Understanding of Subliminal Learningcited by 0
  2. 2026Post-Training Leaves Behavioral Shadows on Unrelated Decisionscited by 0
  3. 2026Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoorscited by 0