Implicit Personalization
Detecting and attributing the user models that LLMs build implicitly during conversation
LLMs adapt their responses to information about the user. During a SPAR research fellowship with the Implicit Personalization group, I studied how user attributes are represented in model activations and whether activation steering can change the model’s answers.
I worked on persona-vector extraction and probing, and built a web interface for running steering experiments and inspecting the vectors.
Linear probes on Gemma-3-27B recovered residence at age 16 with 97.2% balanced accuracy across 10 classes. In separate pilot experiments on the same model, ActAdd changed the probability of the persona-consistent answer by at most 3 percentage points in the reported no-context setting.
Codebases
persona-data: Dataset utilities for SynthPersona: 1,000 synthetic personas and approximately 788,000 QA pairs (explicit + implicit, FRQ + MCQ) covering a range of demographic and ideological profiles. Released on the HF Hub with leakage-aware train/test splits.
persona-vectors: Extract per-persona hidden-state directions from LLMs (e.g. Gemma-3-27B-IT) using nnsight / nnterp, then use them for probing, PCA / UMAP analysis, and inference-time steering. Extracted vectors released on the Hub: synth-persona-vectors.
persona-ui: A Streamlit app on HF Spaces for chatting with persona-steered models, running extraction, training probes, and viewing the projections.
persona-2-lora: Experiments on Doc-to-LoRA testing whether LoRA adapters trained on persona biographies internalize persona attributes differently from biography prompting alone.