Checkpoint: A Two-Level Map of Early Convergence

AI-assisted brainstorming is easiest to misunderstand when several different kinds of narrowing are given the same name. The useful judgment is not simply whether the final ideas seem repetitive, but whether the loss occurred in one session, across independent sessions, or after agents began influencing one another. Bringing those distinctions together gives us a practical diagnostic: locate the point at which independent exploration stopped, then choose a remedy that changes that point rather than applying a generic creativity prompt.

One diagnosis needs two axes

The first axis follows a single session over time. An LLM receives a request, produces an idea, sees that idea in its context, and produces another. If later ideas keep returning to the first solution’s structure, the session is showing fixation-like narrowing. The relevant sign is not identical wording. It is slow accumulation of genuinely new categories or combinations. A list can contain ten fluent sentences and still be exploring only one conceptual route.

The second axis compares independent sessions at the same moment. Before a session has seen its own previous outputs, its first idea offers a cleaner view of what the model reaches for by default. If many independent sessions begin with similar categories and combinations, the system has little variation in its starting points. That is a different problem from one session elaborating too long on its first idea. It points to a shared accessible center across sessions.

These axes produce four useful cases. A session may start in a distinctive region and then become anchored: varied beginnings, narrow local development. It may start near the same region as other sessions but explore different consequences afterward: centralized beginnings, more varied trajectories. It may be narrow on both axes, which is the clearest case of collective contraction. Or it may be broad on both, in which case later convergence may be a reasonable result of selection rather than an early failure.

The human–LLM evidence makes the distinction concrete. In the controlled fitness-product task, the default LLM and human participants showed similar levels of within-session accumulation and pairwise distance, while human first ideas were substantially more varied across participants. The strongest difference therefore appeared between sessions rather than inside each session. The result does not say that LLM lists are always diverse or that humans never fixate. It says that the two populations differed most clearly in how independent contributors were distributed at the beginning.

A third observation is needed once sessions interact: when do trajectories become coupled? Several independent runs may begin broadly and then converge after agents share a leader’s framing, peer judgments, or a common discussion history. That collapse is not visible in first-idea measurements alone. Nor can it be diagnosed from a final transcript without knowing what alternatives existed before the exchange. The system’s visibility rules and authority structure become part of the causal story.