Counterfactual Novel Simulation with LingTai Multi-Agent Networks
Change one character, event, or rule in a story world, then study how the narrative diverges.
Why this matters
This project treats a novel as a simulated social world rather than a prompt to continue. We define a canonical baseline, introduce controlled interventions, let specialized LingTai avatars roll out the consequences, and evaluate whether the resulting branches remain coherent, causally grounded, and literarily interesting.
Research questions
- How should a narrative world be decomposed into persistent roles, memories, goals, and shared state?
- Which interventions produce meaningful divergence instead of arbitrary rewriting?
- Can multiple specialized agents preserve character consistency better than one monolithic generator?
- How do we evaluate story branches: causality, theme, character fidelity, novelty, and reader interest?
Possible LingTai design
- Character avatars maintain private goals, beliefs, relationships, and local memories.
- A world-state avatar tracks facts, timeline constraints, and intervention effects.
- A narrator avatar turns state transitions into readable scenes without owning the world model.
- A critic/evaluator avatar scores causal consistency, role fidelity, thematic drift, and quality.
- Repeated rollouts estimate a distribution of possible endings under the same intervention.
Expected outputs
- A reproducible counterfactual story-simulation protocol
- A LingTai demo with role, world-state, narrator, and critic avatars
- Divergence maps showing how interventions alter plots and relationships
- A paper-style evaluation over public-domain or original story worlds
People behind this opportunity
Lin Du is an Assistant Professor at the National University of Singapore, appointed in Chinese Studies and jointly with Japanese Studies. Her profile sits at the intersection of Chinese and Japanese studies, Asian studies, digital humanities, art history, media studies, machine-learning applications, photography, and visual culture. For this opportunity, that combination matters: counterfactual novel simulation needs not only agent engineering, but also careful thinking about narrative form, cultural context, visual and textual archives, and what counts as a meaningful interpretation rather than arbitrary generation.
Zesen Huang is a postdoctoral scholar in Earth, Planetary, and Space Sciences at UCLA, working in plasma astrophysics with research interests including solar wind, magnetohydrodynamic turbulence, solar physics, and time-series analysis. He is the creator of LingTai, an agent operating system built around persistent memory, skills, avatars, daemons, mail, and multi-agent growth. As a collaborator, he brings the LingTai agent-systems side of the work: helping translate the research question into an executable experiment, define avatar topologies, maintain reproducible simulation protocols, build interactive demos, and connect literary counterfactuals with measurable agent-network behavior.
Public-domain novels or original synthetic settings are preferred for the first experiments.
The goal is a research protocol and demo, not unconstrained fan fiction generation.