Tracing the Persona Circuit: How Large Language Models Encode and Express Character Traits

Guanzheng Qin, Chenghao Sun, Zhining Xie, Xinmei Tian
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:101666-101687, 2026.

Abstract

Large Language Models (LLMs) demonstrate remarkable potential in role-playing tasks but frequently suffer from personality decay—termed "Out-of-Character" (OOC) behavior—during prolonged interactions. While heuristic strategies exist to align model behaviors, the internal computational dynamics driving personality expression remain opaque. A fundamental barrier to decoding these mechanisms is a metric gap: while standard causal attribution paradigms target atomic, single-token outcomes, personality manifests as a holistic, multi-token behavioral tendency. We bridge this gap via the Latent Persona Vector, a differentiable proxy enabling the first fine-grained causal tracing of personality circuits. This metric reveals a structured "Preparation-Establishment-Expression" dynamic and identifies a mechanistic contributor to OOC behavior: competition between persona-specific signals and an assistant-like default direction during the critical "Establishment" phase. Guided by this diagnosis, we propose surgically recalibrating the signal magnitude in fewer than $5%$ of attention heads. This targeted intervention effectively strengthens the persona signal, significantly restoring character consistency while preserving general reasoning capabilities.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-qin26h, title = {Tracing the Persona Circuit: How Large Language Models Encode and Express Character Traits}, author = {Qin, Guanzheng and Sun, Chenghao and Xie, Zhining and Tian, Xinmei}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {101666--101687}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/qin26h/qin26h.pdf}, url = {https://proceedings.mlr.press/v306/qin26h.html}, abstract = {Large Language Models (LLMs) demonstrate remarkable potential in role-playing tasks but frequently suffer from personality decay—termed "Out-of-Character" (OOC) behavior—during prolonged interactions. While heuristic strategies exist to align model behaviors, the internal computational dynamics driving personality expression remain opaque. A fundamental barrier to decoding these mechanisms is a metric gap: while standard causal attribution paradigms target atomic, single-token outcomes, personality manifests as a holistic, multi-token behavioral tendency. We bridge this gap via the Latent Persona Vector, a differentiable proxy enabling the first fine-grained causal tracing of personality circuits. This metric reveals a structured "Preparation-Establishment-Expression" dynamic and identifies a mechanistic contributor to OOC behavior: competition between persona-specific signals and an assistant-like default direction during the critical "Establishment" phase. Guided by this diagnosis, we propose surgically recalibrating the signal magnitude in fewer than $5%$ of attention heads. This targeted intervention effectively strengthens the persona signal, significantly restoring character consistency while preserving general reasoning capabilities.} }
Endnote
%0 Conference Paper %T Tracing the Persona Circuit: How Large Language Models Encode and Express Character Traits %A Guanzheng Qin %A Chenghao Sun %A Zhining Xie %A Xinmei Tian %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-qin26h %I PMLR %P 101666--101687 %U https://proceedings.mlr.press/v306/qin26h.html %V 306 %X Large Language Models (LLMs) demonstrate remarkable potential in role-playing tasks but frequently suffer from personality decay—termed "Out-of-Character" (OOC) behavior—during prolonged interactions. While heuristic strategies exist to align model behaviors, the internal computational dynamics driving personality expression remain opaque. A fundamental barrier to decoding these mechanisms is a metric gap: while standard causal attribution paradigms target atomic, single-token outcomes, personality manifests as a holistic, multi-token behavioral tendency. We bridge this gap via the Latent Persona Vector, a differentiable proxy enabling the first fine-grained causal tracing of personality circuits. This metric reveals a structured "Preparation-Establishment-Expression" dynamic and identifies a mechanistic contributor to OOC behavior: competition between persona-specific signals and an assistant-like default direction during the critical "Establishment" phase. Guided by this diagnosis, we propose surgically recalibrating the signal magnitude in fewer than $5%$ of attention heads. This targeted intervention effectively strengthens the persona signal, significantly restoring character consistency while preserving general reasoning capabilities.
APA
Qin, G., Sun, C., Xie, Z. & Tian, X.. (2026). Tracing the Persona Circuit: How Large Language Models Encode and Express Character Traits. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:101666-101687 Available from https://proceedings.mlr.press/v306/qin26h.html.

Related Material