Why Simple Prompt Fixes Fail—and Which Interventions Complement Each Other
Once early convergence has been separated into starting-point concentration and within-session fixation, prompt design becomes a more precise decision problem. This chapter compares the interventions tested in AI-assisted brainstorming—temperature changes, human seeding, famous creative personas, ordinary heterogeneous personas, and staged Chain-of-Thought revision—so the reader can match a tool to the bottleneck it actually affects. The main lesson is that variety is not produced by adding noise indiscriminately; it is produced by changing the conditions under which different ideas are sampled and developed.
Why more randomness does not create independent ideas
The most obvious way to ask an LLM for more variety is to increase its sampling temperature. Temperature changes how readily the model selects less probable next tokens. That can introduce surface variation, but the mechanism is local: it changes the randomness of each continuation. It does not necessarily give separate sessions different knowledge regions, different associative structures, or a reason to question the same dominant premise.
The temperature test in the fitness-product study illustrates the limitation. At the default temperature, the LLM generated 22 total categories and 88 unique category combinations across 990 ideas. Raising the temperature to 1.5 produced 23 categories and 99 combinations from the same number of ideas. That is a measurable increase, but a small one compared with the human result of 28 categories and 206 combinations. At temperature 2.0, outputs frequently degraded into garbled text that was difficult to interpret or categorize.
The trade-off is not accidental. A model’s ordinary distribution contains both familiar and unusual continuations. Pushing harder toward unlikely tokens can move an output away from the center, but it can also break the relationships that make an idea coherent. A random word substitution is not the same as a new product premise, a different user need, or a distinct research direction. When the desired outcome is a broader semantic search, token-level randomness is too blunt an instrument.
This does not make temperature irrelevant. Different tasks and models may have different useful settings, and modest sampling changes can be part of a broader exploration strategy. The study does not establish a universal optimal temperature. Its narrower finding is that raising temperature alone did not meaningfully close the human–LLM diversity gap and that an extreme setting damaged interpretability. In the multi-agent evidence, a structural gap between authority-weighted and less constrained collaboration also remained across a range of temperatures. Sampling randomness changed the level of variation without removing the effect of how agents were organized.
Human seeding appears more targeted because it supplies meaningful content rather than random tokens. Yet the downstream seed test also failed in the tested procedure. Giving each LLM a different human first idea changed the visible starting point, but ideas two through ten were no more diverse than those from default sessions. The seed could alter the first continuation without recreating the background that would make one trajectory continue differently from another.
The contrast is useful. Temperature changes how the model samples around a distribution. A seed changes what is in the context. Neither, by itself, necessarily changes the persistent route through the problem. A seed can be valuable when the goal is to explore the implications of one example, but it should not be assumed to create an independent ideator.