DIRECTING THE UNSCRIPTED: ALT-MIRAGE FOR EMERGENT SOCIAL DECEPTION DRAMA

Liu Yan, Baoyang Chen, Zhao Yang, Huamin Qu, Zhai Lidong

Keywords: Interactive Storytelling, Social Deduction Games, Multi-Agent Systems, AI Director, Mixed-Initiative Control, Dynamic Epistemic Logic, Unscripted Drama Generation, Information-Theoretic Tension Measurement

TL;DR: Alt-Mirage, a multi-agent AI Director framework for unscripted social-deduction storytelling, uses four sub-agents and dual control to balance autonomy with tension-aware, belief-consistent steering, validated in Among Us-style simulations.

Abstract: We introduce Alt-Mirage, a multi-agent AI Director framework for unscripted interactive storytelling in social-deduction games. Unlike prior director systems that primarily enforce predefined plot structures, Alt-Mirage targets the reality- show-like appeal of improvisation, deception, and emergent conflict by balancing autonomy with controllability. Our Director decomposes narrative orchestration into four specialized sub-agents: a Planner that converts high-level outlines into narrative goals, a Director that schedules interventions online, a Verifier that au- dits belief updates with a log-grounded DEL-ToM epistemic model to prevent hallucinated world-state changes while allowing strategic lying, and a Critic that quantifies narrative tension via information-theoretic signals (audience uncer- tainty, conflict entropy, and event surprisal) and casts intervention as an optimiza- tion objective. To steer dynamics without overriding agent reasoning, Alt-Mirage introduces soft control through a bounded SuperEgo modulation layer that re- ranks Ego-proposed action candidates under persona-consistency gating, preserv- ing character autonomy while enabling interpretable behavioral steering; hard control complements this by reshaping environment conditions (e.g., blackout, lockdown, rule variants). We further provide a lead-director dashboard enabling mixed-initiative control by visualizing target vs. realized tension, agent belief states, and inconsistency alerts, and allowing high-level intervention commands. Experiments in an Among Us-style simulator with nine LLM-driven agents show that soft-control directives reliably steer outcomes and pacing (e.g., increasing faster-win rates and accelerating escalation) while maintaining emergent variabil- ity, demonstrating a practical route toward controllable, belief-consistent, tension- aware interactive drama.

Track: Long Paper

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 112