Skip to main content

Beliefs & Stance

The dynamics that make SimSwarm more than a chat transcript live in how agents' beliefs change over time. This page explains the idea conceptually; the precise formulas are documented separately in the engine internals.

Belief state​

Each agent holds a belief state with three moving parts:

  • Position — where the agent stands on the topic, on a scale from opposed to supportive.
  • Confidence — how firmly that position is held.
  • Trust — how much the agent weights each other agent's contributions.

An agent's position and confidence are rendered into plain-English bands in its prompt (for example, "leaning supportive" with "moderate" confidence), so the simulated participant behaves consistently with how its beliefs currently sit.

Stance scoring​

When an agent authors a post, the text is scored into a stance — a single number from strongly negative to strongly positive — using sentiment analysis. This stance is what other agents react to when they're exposed to the post. Scoring is deterministic and runs without an LLM call.

How beliefs update each round​

After a round's posts are recorded, every agent updates its beliefs based on the posts other agents authored (agents do not influence themselves). The update blends several conceptual forces:

  • Pull toward stance — exposure nudges an agent's position toward the stance of what it read, proportional to the gap between them.
  • Novelty — an idea the agent hasn't seen before carries more weight than a repeat of something already in its exposure history.
  • Social proof — posts with more engagement (likes) carry more influence, with a floor so that even zero-engagement posts still register.
  • Trust weighting — content from agents the agent trusts more counts for more.
  • Resistance — the more confident an agent already is, the smaller each nudge.
  • Trust evolution — agents whose stance ends up aligned with the agent's resulting position gain trust; those who oppose it lose trust. Trust is learned over the run, not fixed.

Confidence itself shifts with the agent's own reception — likes on its posts build confidence, dislikes erode it.

Because these forces compound across rounds, you get the population-level phenomena the results surface: blocs forming around shared positions, a narrative gaining or losing ground, and turning points where the balance shifts. The downstream story signals and report read these outcomes back out of the run.