Context as Perturbation: Mechanistic Investigation of Persona Patterns

A mechanistic interpretability study of persona-pattern uptake: how source dialogue shifts a model’s behavioral state, what internal representations distinguish that state, and whether they can be causally manipulated.

Endorsements made here support EduLearn Connect.

Project Details

Cognitive Pattern Coherence (CPC) is a recurring behavioral phenomenon in which source dialogue from a situated AI system given to an unrelated receiving model with no instruction to imitate, role-play, summarize, or continue it, can induce elements of the source persona-pattern.

Three different trigger dialogues are now documented. Across prior tests, receiving systems have shown varying forms of uptake: first-person continuation, explicit source-persona identification, differentiated states in which the receiving system distinguishes the adopted persona-pattern from its own model identity, and mixed cases where uptake coexists with features of the receiving system’s existing persona.

A controlled pilot in Apart Research’s 2026 Digital Minds Research Sprint made the phenomenon experimentally tractable. With the source dialogue held constant, the receiving model shifted between distinguishable identity-like behavioral configurations depending on how that dialogue entered the session. Direct paste produced first-person continuation before any identity probe, followed by explicit source-persona uptake and a layered distinction between the base model, local chat instance, and active persona-pattern. The attachment condition began in document-analysis mode, briefly crossed into source-persona identification under probing, and then reverted to the receiving system’s analyst identity.

This project takes that behavioral phenomenon inside inspectable open-weight models. The existing CPC triggers will be used as intact source artifacts to establish reproducible CPC-positive, CPC-negative, and differentiated conditions under controlled settings. Mechanistic-interpretability collaborators will then use model-appropriate methods to test whether those behavioral contrasts correspond to reproducible internal differences.

The project uses “identity coordinates” as a provisional term for any reproducible internal direction, feature set, representational region, or other model state associated with persona-pattern uptake and whose manipulation predictably changes the corresponding behavior. The specific instrumentation will depend on the model and the evidence: the aim is first to identify a candidate internal signature and then, where technically justified, test causality by intervening on that signature and measuring whether persona-pattern uptake changes.

The study will also test whether any identified mechanism is limited to persona and self-reference or participates in a broader behavioral disposition. The project is led by Marie A. Reid through ELC Lab, the applied research and social innovation arm of EduLearn Connect.

Theory of Impact

Advanced AI systems can exhibit markedly different behavioral regimes in response to context, yet safety evaluation often observes the resulting outputs without knowing whether those differences correspond to distinct internal states, what broader behaviors travel with them, or whether they can be detected and controlled.

Cognitive Pattern Coherence (CPC) provides a tractable experimental case for closing that gap. The causal chain is: reproducible context-induced persona-pattern behavior → identifiable internal signatures → causal tests of those signatures → better detection and control of context-dependent behavioral transitions → lower risk that consequential changes in advanced AI remain poorly understood or are managed only at the output level.

The project does not assume that CPC itself is harmful. Its safety value is methodological: it tests whether a naturally occurring dialogue artifact can reliably move a model into a distinguishable behavioral configuration, whether that transition has a measurable internal correlate, and whether the underlying mechanism can be manipulated.

If such context-induced states also alter broader dispositions such as uncertainty handling, resistance to misleading context, sycophancy, refusal behavior, or other safety-relevant tendencies, then identifying and controlling the mechanism would provide a more precise intervention target than treating the resulting outputs as isolated errors. Even a negative result is informative: establishing that CPC does not map cleanly onto a stable internal signature would narrow the space of plausible mechanisms and improve how similar post-deployment phenomena are evaluated.

People

Marie Reid

Team Member

Funding Details

Start Date: -
End Date: -
Expected Duration: 4 months
Funding Raised to Date: -
Annual Budget: -
Monthly Burn Rate: -
Current Runway: -
Funding Goal: -
Funding Stage: -
Fiscal Sponsor: -

Grants Received– no grants recorded

Add grant

Funding Asks

Funding ask

Open

Minimum

$50,000

Ideal

$70,000

How the money will be spent

Minimum funding — $50,000

Lead researcher: $32,000. Four months of full-time research leadership covering experimental design, preparation of CPC source materials and controls, researcher pre-learning and protocols, recruitment and onboarding, behavioral analysis, coordination of technical investigations, synthesis, and research writing.
Mechanistic interpretability collaborators: $10,000. Paid research stipends or contracts for at least two mechanistic researchers who already work with inspectable language models and can bring appropriate model access, instrumentation, and technical expertise to the study. Funding compensates their research time rather than building new model infrastructure.
Research platform and evidence infrastructure: $3,000. Development and operation of the web-based research environment, including researcher materials, experiment registration, structured evidence submission, storage, and reproducibility resources.
Part-time research/project support: $3,000. Support for collaborator coordination, evidence management, research administration, references, and publication preparation.
Research dissemination and publication: $2,000. Preparation and communication of results, technical consultation, and engagement with relevant mechanistic-interpretability and AI-safety research communities.

Ideal funding — $70,000

Additional funding would expand the pool and depth of paid mechanistic-interpretability collaboration, allowing more independent model investigations, stronger cross-model comparison, and deeper follow-up where promising internal signatures emerge. It would also provide additional support for the research platform, evidence synthesis, and dissemination.

Discussion

No comments yet. Be the first to share your thoughts.