Work

Language Models Can Shed a Persona but Not the Person

Year: 2026
Type: preprint
Abstract: We implant Wenzhi, a fictional persona written for this study (a piano tuner from an invented city), into Qwen3-4B in two ways: full fine-tuning on written first-person dialogues about her life (C), and off-policy self-distillation on the base model's answers while it role-plays her from a system prompt (D). Alongside the prompted model (B), we train three-seed controls on a third-person rewrite (T) and on persona-free everyday dialogue (F). B, C, and D all answer 91–96% of held-out identity questions as Wenzhi, yet their activation shifts from the base model (A) split: D moves toward the prompt direction in late layers, while C stays close to T and F across the network. Even so, clamping a rank-2 to rank-8 subspace taken at layer 28 to A's mean across residual hooks cuts Wenzhi answers to 3–8% of shared persona items in both C and D (vs. 90–99% for random subspaces), with damage at most 10%. Removing Wenzhi does not restore the assistant. Clamped models identify as an AI or digital entity on 5–10% of these items (vs. 71% in A) and speak mostly as an unnamed person; the leading clamped direction is nearly orthogonal to the assistant axis and overlaps strongly with the shift made by persona-free fine-tuning alone, which cuts assistant self-identification from 74% to 18–20%. How the subspace is erased matters: zeroing coordinates to the origin instead of clamping to a reference mean produces AI, digital, or animal speakers in 40–54% of D's answers and inflates self-naming in every seed of C; mean-clamping largely removes both effects. Added to A, the trained directions elicit human first-person speech in 42–76% of answers while recalling Wenzhi in 1 of 3,510.
Language: English
Open Access status: green
OpenAlex ID: W7233603299
Cites:
Cited by:
Related to: