MiroFish Deep Dive (3): OASIS Simulation Engine for Digital Rehearsal of Future Scenarios

Key Points

• Purpose of OASIS in MiroFish: OASIS is the social media simulation engine from the CAMEL-AI (Communicative Agents for "Mind" Exploration of Large Scale Language Model Society) project. It enables users to "fast-forward" time and rehearse alternative crisis-response strategies in a digital environment, rather than picking one path in the real world.
• Agent Model: Each agent has a full persona, follower/friend counts, historical posting counts, karma (for Reddit), active hours, posting frequency, sentiment bias (-1.0 to 1.0), stance (supportive/opposing/neutral/observer), and interested topics. Personas are several hundred characters long and drive LLM behavior.
• Platform Models:
• *Twitter*: 280-character limit, recommendation weights (recency 0.4, popularity 0.3, relevance 0.3), viral threshold at 10 interactions, echo-chamber strength 0.5.
• *Reddit*: Up to 40,000-character posts, subreddit structure, upvote/downvote mechanics, community-norm dictionary, higher emphasis on community weight for recommendations.
• Time Simulation: Default 72-hour horizon, 60 minutes per round. Activity multipliers are applied per hour, with Chinese routines modeled explicitly:
• 0:00–5:00 → 5% (off-peak)
• 6:00–8:00 → 40% (morning ramp-up)
• 9:00–18:00 → 70% (work hours)
• 19:00–22:00 → 150% (evening peak)
• 23:00 → 50% (late night)
• Action Vocabulary: Agents can create posts, comment, like posts/comments, repost, follow/unfollow, scroll, or do nothing. Each round, activity probability is calculated from time-of-day multiplier, posts-per-hour, and LLM-based persona reasoning.
• MiroFish Integration:
• *Profile generation*: Knowledge graph entities (e.g., "张三, USTC CS student") are mapped to OasisAgentProfile records including username, persona narrative, follower count, active hours, topics, and stance.
• *Config generation*: A SimulationConfigGenerator produces time, agent-activity, event, and platform parameters automatically.
• *Dual-platform runs*: A SimulationRunner runs Twitter and Reddit simulations in parallel, returning combined statistics such as post/comment counts, viral events, and sentiment trajectories (e.g., Twitter negative 65% → 45%; Reddit negative 70% → 35%).
• Interview Agents: Instead of LLM-improvised answers, MiroFish calls the real OASIS interview API. It auto-selects up to 5 agents relevant to a topic (students, media, school officials, government), generates tailored questions, and returns integrated responses. This turns statistics into qualitative insights.
• Rehearsal vs. Prediction: The author explicitly frames OASIS as a rehearsal and strategy-testing tool, not a forecaster. Value lies in testing strategies, identifying risk paths, training decision-makers, and hypothesis testing. Limits include LLM behavioral gaps versus real humans, simplified platform algorithms, and high compute cost from large-scale LLM calls.
• Lifecycle and IPC: A SimulationManager exposes start/pause/resume/stop/status operations with states IDLE → STARTING → RUNNING → [PAUSED] → STOPPING → STOPPED → COMPLETED/FAILED. An SimulationIPCClient sends control commands, streams logs, and invokes the interview API.
• Methodological Shift: The article positions MiroFish as a paradigm shift from observational social science to simulation experimentation, comparing it to a flight simulator for decision-makers. The next installment in the series will cover the ReACT report-generation mechanism.