LLM representations shift when models adopt political personas
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers demonstrate that when LLMs adopt different personas, the geometric structure of their internal representations shifts systematically. Testing liberal and conservative political framings across 9,000 words and 33 semantic axes reveals that persona conditioning reorganizes how models encode meaning, not just surface outputs. This finding challenges core assumptions in mechanistic interpretability work, where linear probes and sparse autoencoders presume stable feature representations. The result matters for both safety and capability work: if model internals are context-dependent and malleable, existing interpretability tools may misdiagnose what models actually compute, and adversaries could exploit persona-switching to evade detection.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper doesn't just show that personas change outputs; it demonstrates that the geometric substrate itself reorganizes. This means existing interpretability methods like linear probes may be measuring artifacts of a particular context rather than discovering what models actually compute.
This connects directly to 'Shared relations, specific choices' from November 1st, which found that relational knowledge persists across contexts but can flip in functional direction between models. That work suggested interpretability requires attention to both geometry and sign. This new result goes further: it shows that even the geometry itself is context-dependent, not just its interpretation. Together, these papers suggest the mechanistic interpretability toolkit (sparse autoencoders, linear probes) was built on an assumption of representational stability that doesn't hold under persona conditioning. The earlier work on 'Personas Are Not People' from November 3rd also becomes more urgent here, since if model internals shift with persona, then using fixed-trait personas in benchmarks doubly misrepresents reality.
If researchers can show that the same sparse autoencoder features trained on one persona fail to reconstruct activations under a different persona at >15% error rate, that confirms representations genuinely reorganize rather than just rotate. Watch whether mechanistic interpretability papers published after this one begin including persona-robustness checks as standard validation.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Language models share relational structure across queries but flip functional signs
Researchers mapped how language models store and apply relational knowledge across different contexts. Using controlled experiments on GPT-2 Small, Qwen2.5, and Llama-3.2, they found that models maintain a low-dimensional shared representation of relations that generalizes to new entities and persists across unrelated queries. Critically, the functional direction of these relations can flip between models even…
MentionsLLMs · sparse autoencoders · linear probes
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Persona Perspectives: Adopting Viewpoints Reorganizes Representations in LLMs”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.