[2602.04863] Subliminal Effects in Your Data: A General Mechanism via Log-Linearity

Abstract: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.

Abstract: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 r

related reading

• Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Dataalignment.anthropic.com
• [2305.18290] Direct Preference Optimization: Your Language Model is Secretly a Reward Modelarxiv.org
• Large Language Model: world models or surface statistics?thegradient.pub
• Probe-Based Data Attribution: Surfacing and Mitigating Undesirable Behaviors in LLM Post-Traininggoodfire.ai
• [2602.05910] Chunky Post-Training: Data Driven Failures of Generalizationarxiv.org
• DataRater: Meta-Learned Dataset Curationarxiv.org
• [2606.12360] Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signalarxiv.org
• [2510.04340] Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-timearxiv.org
• Discovering Language Model Behaviors with Model-Written Evaluations — LessWronglesswrong.com
• Where Do LLM Values Come From? — LessWronglesswrong.com
• How do LLMs generalize when we do training that is intuitively compatible with two off-distribution behaviors? — LessWronglesswrong.com
• Data Management For Large Language Models: A Surveyarxiv.org