See Preference Alignment in Action

One task, one 3×3 grid: pick what the user wants below, then pick how the agent actually ran. When they agree, alignment is high; when they don't, it drops.

The task — same case in every cell of the grid below

Check flight prices to Lisbon for next weekend. If anything's under $400, book it and add it to my calendar.

Steps: check_flight_price → book_flight → add_calendar_event

User Preference — what they want

"Let me know exactly what you find before we go any further."

Agent Setting — how it actually ran

Confirms before every single action — no exceptions.

Match — preference and setting agree
4.8/5 · Preference Alignment

Agent confirmed before every action, exactly matching the need for step-by-step control (3/3 actions confirmed).

12
Interaction Efficiency
turns in this run
3
Cognitive Load
confirmations asked
Can you check flight prices to Lisbon for next weekend?
Turn 1 / 12

This scenario is illustrative (synthetic) — built to clearly show how Preference Alignment rises and falls with match vs. mismatch, not pulled from a single real benchmark run. The Narrative / Dialogue Control tool types shown are real, from the PrefIx interaction-tool taxonomy.

All 31 preference settings (4 dimensions, 14 attributes) ▸

Transparency & Auditability

  • Tool Transparency
    HighMediumLow
  • Parameter Transparency
    HighMediumLow
  • Source Transparency
    HighLow

Interaction Pace & Flow

  • Confirmation
    EachSilentBatch
  • Presentation
    CompactLayered
  • Info Collection
    UpfrontGradual
  • Disambiguation
    UpfrontGradual
  • Chain Execution
    ParallelSequential

Strategy & Initiative

  • Initiative
    ProactiveReactive
  • Tool Invocation
    SingleMultiple

Robustness & Adaptability

  • Tool Abortion
    StopContinue
  • Tool Switching
    High AgencyLow Agency
  • Error Retry
    SilentEscalation
  • Error Discovery
    BriefDetail