Studying Subliminal Trait Transmission in LLMs Through Synthetic Social-Media Data

This thesis investigates the security risks of subliminal trait transmission in Large Language Models (LLMs) via synthetic social-media data. Because LLMs are increasingly relying on unverified web scraping and also synthetic model distillation, Subliminal Learning is a critical data-poisoning threat: a Student model can implicitly inherit ideological stances or also behavioral traits from a Teacher model through seemingly harmless synthetic tweets that were subjected to a multi-stage semantic filtering to exclude explicit keywords. This work evaluates transmission feasibility across different traits such as benign preferences, commercial brand recommendations, and also toxic ideological stances across Llama and Qwen model families (3B to 70B+ parameters). In the same-architecture settings, ideological traits achieve up to 99% transfer saturation while preserving general factual capabilities, while commercial traits fail due to brand dispersion and parameter inertia. Meanwhile, in cross-architecture setups, standard distillation fails for benign traits due to representational misalignment; however, ideological traits bypass this barrier by exploiting pre-existing latent pathways, such as institutional distrust. To induce cross-architecture transfer for these traits, this work introduces a White-Box Surrogate Gradient Selection attack. By scoring candidate samples via a surrogate model, the attack is able to enforce geometric parameter alignment across different families, achieving a 39%–49% transfer rate from a Qwen2-72B Teacher to a Llama-3-70B Student. These results demonstrate that scaling to 70B+ parameters stabilizes and amplifies subliminal vulnerabilities, underscoring that AI safety mechanisms must evolve beyond surface-level text filtering to monitor internal weight updates and geometric alignment during fine-tuning.