OASIS: Simulating Social Media at Scale with LLM Agents

‍

Social media platforms shape how information spreads, how opinions form and how communities organise. They're also extraordinarily difficult to study. Running real-world experiments at scale is costly, ethically complicated and often simply not possible.

‍

In a recent Passion Academy session, Tamim Hussein walked us through OASIS: a research framework from Shanghai AI Lab and CAMEL-AI that attempts to solve this by simulating up to one million LLM-based agents across X and Reddit simultaneously.

‍

‍

‍

‍

Why Simulate Social Media at All?

‍

Agent-Based Models have long been used to study social dynamics but traditional approaches rely on rule-based thresholds that don't capture how people actually behave. LLMs change this. Agents can now role-play human personas with context-dependent behaviour, operate across a rich action space from a simple like to multi-step deliberation and reason in natural language rather than following fixed rules.

‍

The problem is that until now, every LLM-based social simulator has been purpose-built for one phenomenon on one platform and most handle only a few hundred agents. However, Real social networks have millions of users.

‍

‍

The Three Gaps OASIS Targets

‍

• Generalisability: each existing model is locked to one phenomenon on one platform, with no shared backbone.

‍

• Scale: hundreds of agents versus the millions that define real network dynamics.

‍

• Dynamics: static networks, no realistic recommendation systems, no temporal behaviour patterns.

‍

OASIS was built to address all three at once.

‍

‍

How OASIS Works

‍

The framework is built from five modular components, each swappable depending on the platform being simulated.

‍

The environment server maintains a relational database of users, posts, comments, follows and interaction traces, updated dynamically as the simulation runs. The recommender system is platform-specific  (TwHIN-BERT interest matching for X-style feeds, hot-score ranking for Reddit-style threads) and can be swapped without touching the rest of the architecture. The agent module gives each LLM-backed agent a memory system and a 21-action space with chain-of-thought reasoning. The time engine uses a 24-dimensional hourly activity vector to probabilistically activate agents across the simulated day. And the scalable inferencer uses asynchronous message channels and a GPU manager to make million-agent simulations computationally feasible.

‍

The key design principle throughout is modularity. Platform-switching is a config change, not a rewrite.

‍

‍

What It Can Reproduce and What It Discovers

‍

OASIS was validated against three well-established social phenomena: information propagation, group polarisation and the herd effect. All three were successfully reproduced across both X and Reddit.

‍

However, two emergent findings stood out. Uncensored LLMs polarise harder than their aligned counterparts and agents herd more aggressively than real humans do. These aren't programmed behaviours, they emerge from the interaction of agents operating at scale. The session noted that the herd effect and other emergent dynamics only reliably appear above around 10,000 agents, which quietly but importantly suggests that scale isn't just a performance goal, it's a scientific requirement.

‍

‍

Where It Still Falls Short

‍

The session was clear-eyed about the limitations and they're worth taking seriously.

The recommender system is an approximation, not a real ranking model. There's no collaborative filtering or personalisation, which means propagation depth in simulations undershoots real-world data. Agents flatten real individuals into demographic and topic vectors, which is a structural gap with how actual humans behave. The action space is text-only (no images, video, bookmarking or live interaction) which sets a hard ceiling on fidelity. Million-agent runs on 24 A100s still take around a week, making iteration expensive. Furthermore, validation rests on replicating three known phenomena rather than any held-out predictive benchmark or third-party audit.

The bottom line from the session: OASIS is the strongest LLM-based social simulation framework available today but it's a research scaffold, not a forecasting tool. Reading its outputs as predictions would overstate what the system can currently do.

‍

‍

The Broader Picture

‍

What makes OASIS interesting beyond the technical achievement is what it points toward. As LLMs get better at modelling human reasoning and behaviour, simulation becomes a genuine alternative to experimentation in contexts where real experiments aren't feasible. Understanding how misinformation propagates, how polarisation takes hold, or how platform design choices affect community dynamics are all questions that matter but very hard to answer empirically at scale.

‍

OASIS doesn't solve that problem but it moves the frontier meaningfully. The combination of modularity, scale and emergent behaviour in a single open framework is a significant step and it's one that will likely look more significant in hindsight than it does today.

‍

References

‍

Yang, Zhang et al. OASIS: Open Agent Social Interaction Simulations — arXiv:2411.11581github.com/camel-ai/oasis

‍