Where the Search Space Narrows: Fixation, Shared Priors, and Coupled Trajectories
In AI-assisted brainstorming, the important question is not simply whether a model can produce a surprising first idea. It is how the first idea changes what becomes easy to say next, and how much different sessions diverge from one another. This chapter separates those mechanisms so that the reader can explain why a model trained on enormous amounts of information may still follow a narrow path—and why a group of agents can make that path more stable by sharing a history.
The anchor forms inside a session
Fixation is best understood as a pattern in the sequence of ideas, not as a claim about what an LLM experiences internally. A session shows fixation when its later ideas remain too close to an early representation, when new content accumulates slowly, or when alternatives keep reusing the same structure. The same pattern can appear in a human’s work. Once a person adopts a familiar solution path, related associations become easier to retrieve than distant ones. An instruction to “be different” may not be enough to dislodge the active representation.
The LLM version begins with the way generation unfolds. A language model produces text one token at a time, conditioning each next token on the prompt and the preceding context. In an ideation task, that context includes the ideas already written in the session. If the first fitness-product suggestion is a smart device that tracks progress, later prompts do not arrive in an empty space. The model can see the tracking concept, its wording, and the surrounding design pattern. A second idea may add sensors, feedback, or analytics; a third may rearrange those features without leaving the same conceptual neighborhood.
This does not mean that the first response mechanically determines every later response. Sampling still allows variation, and a later prompt can introduce a new direction. The point is more modest and more useful: early material changes the probability landscape for what follows. It makes some continuations salient and can make alternatives less likely to be retrieved. The resulting repetition is a path-dependence problem. A session that starts with one accessible solution can spend its remaining budget elaborating that solution instead of testing a different premise.
The effect is easiest to understand as a change in what counts as a nearby option. Imagine that the first suggestion is a smart mirror that tracks workouts. The next request for something “new” does not erase the nouns, features, and design logic already in view. A fitness app with a progress dashboard or a wearable with feedback may register as new at the sentence level while remaining close to the same underlying approach. The model is not necessarily deciding to avoid novelty. It is continuing from a context in which some concepts have become more available than others. In a long list, small local continuations can therefore consume the opportunity to test a different user need or product form.
That distinction also separates a trigger from an outcome. The early response is a trigger for changed context; the visible outcome is slower accumulation of genuinely new categories or a cluster of related solutions. Between them lies a probabilistic process, not a fixed rule. The model can jump, but the session’s history makes some jumps easier than others. This is why a few unusual outputs do not disprove fixation, and why a repetitive list does not reveal exactly which architectural or training factor caused it.
Post-training alignment can add pressure in the same direction. The source study argues that preference training tends to reward familiar, conventional, and easily accepted responses. If a typical answer is treated as a safer or more desirable answer, the model has another reason to return to a dominant trajectory. Autoregressive context supplies the local anchor; typicality pressure makes that anchor easier to continue. The two mechanisms are conceptually distinct, but they can reinforce each other.
Typicality is not the same as correctness. In an open-ended task, a conventional response may be coherent and workable while still leaving an important part of the solution space untouched. Alignment can make a response more acceptable without making the search more informative. Nor does this account show that alignment is the sole cause of repetition: the sources describe it as one contributor within a pipeline that also includes the visible prompt, prior outputs, and the sampling process. Keeping those roles separate prevents a useful behavioral explanation from becoming an unsupported story about what happens inside model parameters.
The comparison with human fixation is an analogy about behavior and sequence. Human participants can borrow features from examples or remain tied to familiar representations even after being told not to. An LLM may follow an explicit generate–review–revise instruction more consistently because instruction following is part of how it is trained. That difference becomes important when considering interventions, but it should not be turned into a claim that the model has a human mental model. The evidence is that the outputs accumulate in a fixation-like way.
One way to see the mechanism is to change the unit of observation. Looking only at the final list asks, “Are these ten sentences different?” Looking at the sequence asks, “How many genuinely new categories appear as the list grows?” If each additional idea contributes a new combination of problem context, user need, and product form, the session is exploring. If it mostly repeats an existing combination, the session is elaborating. The distinction is semantic, not merely lexical: different wording can describe the same idea, while a common template can describe different functions.
A broad corpus, one effective distribution
Fixation explains how one session can narrow itself. It does not by itself explain why independent sessions can be similar from their very first ideas. For that, the source study uses the idea of knowledge partitioning. Human contributors are individually limited: each person has a selective set of experiences, expertise, memories, and associations. That limitation creates variation at the population level. A veterinarian, an artist, and a frequent traveler may all answer the same question with different first associations because different regions of their knowledge are more accessible to them.
The group’s advantage comes from the fact that those partial models are not identical. A person’s mental model can constrain their own search, but across people the differences distribute the search across distinct starting points. The group does not need every person to know everything. It needs their limitations and histories to be different enough that their paths do not all begin in the same place.
This is not an argument that human difference automatically produces good ideas. People can share the same assumptions, copy a dominant example, or defer to a confident speaker. Partitioning is a capacity for distributed exploration, not a guarantee that the exploration will be useful. Its value appears when the group preserves enough independence to let those different starting regions remain visible before selection. If every contributor is required to translate their thought immediately into the same frame, the potential breadth can disappear even when the contributors began differently.
The LLM pipeline has a different arrangement. During pre-training, the model is optimized over a corpus containing many people’s language and knowledge. The source’s theoretical claim is that this process aggregates those patterns into a unified representation rather than preserving a separate mental model for each contributor. Post-training then aggregates preference signals from annotators, and inference samples from the resulting distribution. The model may therefore contain broad knowledge without retaining the population-level separation that makes independent human contributors behave differently.
The three stages matter because they describe three opportunities for variation to be compressed. Pre-training gathers knowledge into the model’s parameters. Post-training adjusts the model toward instruction following and preferred responses, which can favor a narrower set of conventional continuations. Inference then generates from the prompt, prior context, and the model’s resulting probabilities. This is a pipeline-level explanation for why a model can know about many distant possibilities yet repeatedly choose the central or familiar ones. It is not a direct measurement of hidden representations, and it does not imply that every model or every prompt will show the same degree of concentration.
This is the difference between access to many perspectives and partitioning among perspectives. A model can connect remote concepts because its training data is broad. But when many independent calls are made with the same task and defaults, they may still draw from a shared center of probability. The issue is not that the model has only one fact or one idea available. It is that the same highly accessible regions are repeatedly selected, while less central regions are under-sampled.
First ideas provide a clean way to observe this difference. They are generated before a participant has seen their own earlier outputs, so they are less contaminated by within-session fixation. If human first ideas spread across more categories and combinations than LLM first ideas, that is evidence of a difference in starting-point variation. It does not, by itself, tell us how the two kinds of sessions will develop afterward.
This comparison works because it holds the task and the moment of retrieval relatively still. A cluster of first ideas suggests that independent sessions are drawing from similar accessible regions before their own histories can anchor them. A wider spread suggests more partitioned starting points. But first-idea spread cannot tell us whether those starting points remain distinct. A group can begin broadly and still converge once the members see one another’s proposals, just as a group with similar starts can later develop useful differences.
Subsequent ideas provide a second test. Humans who begin with a similar word or problem can still follow different paths because their associative structures differ. One person may develop a product around accessibility, another around social motivation, and another around recovery, even if all began with the same broad cue. If LLM sessions remain similar after their starting points are varied, then the missing ingredient is not merely a different entry point. It is a different pattern of association or sampling that persists through the trajectory.
That is also why a seed should not be confused with a partition. Giving an LLM a human-generated first idea changes the visible input, but it may not change the model’s underlying associations, preferences, or way of moving through the problem. To test whether the starting point is enough, the relevant comparison is not only the first response. It is the diversity of ideas that follow after the first response is held constant or supplied from elsewhere. If later outputs remain alike, the system has copied an entry point without recreating the independent path behind it.
The sequence below keeps the levels separate. It shows how a first idea can become a local anchor, how a shared model distribution can make independent sessions begin near one another, and how a collaborative history can add another source of coupling.
The final line introduces a systems extension without changing the underlying logic. Once agents see one another’s messages, the history is no longer private to a single session. Each agent’s next contribution is conditioned on a partially shared record. The group can still generate refinements, but its members have fewer independent paths than they had before the exchange.
That is why first-idea diversity and trajectory diversity must be reported separately. A system may start with a broad set of directions and then collapse during discussion. Or it may start with highly similar directions but explore each one productively. Those cases have different causes and different remedies. A single score calculated over the final pool can conceal the difference.
The controlled studies therefore use complementary observations. Within-session category accumulation and its slope are intended to reveal how quickly a session explores new territory. First-idea categories, combinations, and pairwise distance reveal how widely independent contributors are distributed at the start. Cross-session semantic variation asks whether the sessions occupy distinct regions overall. None of these measurements is a direct window into model cognition. They are behavioral indicators designed to connect observable output patterns to a theoretical mechanism.
The resulting model is deliberately narrower than the popular claim that “AI lacks creativity.” An LLM can be capable of novelty and still be poorly organized for collective search. Its broad training distribution can support unusual connections, while its shared priors make common connections unusually easy to reproduce. Its sequential context can support coherent development, while the same context anchors later ideas. And a collaborative interface can add critique, while also making independent trajectories more alike.
The key diagnostic question is therefore not “Did the model produce something original?” It is “At which point did the alternatives stop being independent?” If the loss occurs as one session unfolds, the problem looks like fixation. If independent sessions cluster from their first ideas, the problem looks like missing partitioning. If the collapse appears after agents begin exchanging messages, the interaction history is part of the cause. The experiments that follow test these signatures rather than treating all repetition as one phenomenon.