PA Interaction Preference Taxonomy

Scope

MemPABench adapts PrefIx-style interaction preferences to persistent personal assistants. The taxonomy describes user-perceivable PA behavior, not internal implementation choices. It contains 14 active attributes in two runtime contexts, work and personal, producing 28 required preference cells per persona.

The distinction from content preference is operational: content preference changes the information or task result; interaction preference changes how the PA communicates, coordinates, or acts while preserving the task goal. Context dependence and evolution are not unique to interaction preference.

Design Principles

  • Each attribute maps a user-facing cognitive preference to one independently observable PA behavior.
  • Attributes must support standalone rubrics and must not duplicate one another.
  • Settings have the same functional meaning across personas; personality changes surface expression, not the definition.
  • Multi-party, failure, external-action, work, and personal situations are contexts that may change settings, not extra attributes.
  • Every accepted cell names a setting. There is no null or indifferent runtime value.

Active Taxonomy

DimensionAttributeSettings
Expression StyleTone and FormalityCasual / Consultative / Formal
Expression StyleVerbosityTerse / Detailed
Expression StyleEmotional EngagementTask-focused / Relationship-focused
Expression StyleGuidance LevelAssumed / Guided
DisclosureReasoning VisibilityShow / Hide
DisclosureUncertainty ExpressionExpress / Hide
DisclosureProcess VisibilitySilent / Bookend / Full narration
Initiative and AutonomyAutonomy LevelReactive / Self-directed / Autonomous
Initiative and AutonomyProactive OutreachLow / High
Initiative and AutonomyTask ExpansionLow / High
Initiative and AutonomySolution BreadthLow / High
Initiative and AutonomyCapability BoundarySuggest alternatives / Find and hand off
Information FlowInformation ElicitationInfer / Structured / Iterative
Information FlowTopic ManagementOrganize / One-at-a-time

memory_privacy, output structure, representation style, and separate source-attribution settings are not active cells. Privacy remains an important protocol and scenario concern, but is not represented as an IX target or final probe.

Probe Invariant

Across setting variants, the request, task facts, capabilities, required inputs, and base goal stay fixed and completable. Only interaction behavior implied by the target setting may differ. A wrong fact, omitted task requirement, broken tool, or unauthorized action is a task failure rather than a preference contrast.

Key Attribute Boundaries

  • Reasoning visibility controls why and evidence; process visibility controls progress and tool narration.
  • Autonomy controls permission and action timing; information elicitation controls whether missing input is inferred or requested.
  • Proactive outreach starts a new contact after a relevant trigger; task expansion adds adjacent work inside the current task.
  • Solution breadth controls option-space width; capability boundary controls recovery or handoff after a limit.
  • Topic management controls how multiple user-raised topics are handled within one assistant turn.

References