Idea Exhaustion Dynamics

For any prompting strategy, there exists a finite pool of ideas the model can generate. Continued generation depletes this pool, causing similarity to increase and novelty to decline.

The pattern resembles fishing from a pond: early catches are varied, but as more fish are removed (or marked and returned), the probability of catching the same fish again increases. Using mark-recapture estimation methods, researchers found:

• Baseline prompting: ~3,700 unique ideas, 13.8% repetition rate
• Chain-of-Thought: ~4,700 unique ideas, 11.7% repetition rate

After approximately 750-800 ideas, the strategy advantage disappears, both approaches converge to similar repetition rates as the accessible idea space becomes exhausted.

This has practical implications: for maximum diversity benefit, use enhanced prompting strategies in the early stages of ideation when they have the most room to differentiate from baseline approaches.

Related: 05-atom—cot-diversity-effect, 05-atom—ai-diversity-deficit