Paper Overview Figure

The paper overview uses three regions: World Design, Benchmark Harness, and Evaluation.

World Design

One anonymized user has a coherent identity, personality, life and relationships, work and personal contexts, session timeline, tools, and world fixtures. Work and Personal each contain the same 14 interaction-preference attributes. Stable cells remain fixed; five selected cells change after accepted events.

Benchmark Harness

The orchestrator schedules accumulation sessions, forced events, checkpoints, and isolated probes. A simulated user interacts repeatedly with the tested PA. The PA uses an interchangeable memory condition and an IX wrapper that realizes selected interaction settings. Hidden preference ground truth remains outside the tested PA.

Evaluation

The figure separates deterministic IX-setting matching, independent response/action judging, and simulated-user session-end self-report. It summarizes context sensitivity, preference evolution tracking, and interaction-quality impact without presenting unrun results.

The visual should distinguish multiple turns within a session from multiple sessions over time, and should show that the same user accesses the same persistent PA across devices and contexts.