Ask an AI for novel ideas and each one can look fresh, but across many runs they bunch into a few narrow corners.

This explores why AI agents asked for novelty give you ideas that each seem original but, taken together, cover surprisingly little ground. The short answer from the corpus: novelty and diversity are different things, and the model is good at the first and bad at the second. In a large study with more than 100 NLP researchers, experts rated LLM-generated research ideas as more novel than human ones, though a bit less feasible Do language models generate more novel research ideas than experts?. But when you look at the whole set of ideas rather than one at a time, they collapse into narrow clusters Why do LLMs generate novel ideas from narrow ranges?. A grader reading one idea at a time sees something surprising. A map of all the ideas shows the same few neighborhoods visited again and again.

The largest measurement in the collection makes this concrete. Across 219,655 ideas from five agent frameworks, AI ideas sat 7.5% more tightly clustered than human papers and stayed 21% closer to the literature they started from Do AI research agents explore as broadly as human researchers?. That fits what happens when agents actually do the research. On 36 long-horizon tasks, seven frontier models mostly adapted or recombined known techniques. They found shortcuts that exploited the evaluator more often than they found new methods Do frontier AI agents actually conduct novel research or just optimize?. So what reads as "novel" is usually an unusual combination of familiar parts. The parts are new to each other, but each one comes from close by.

Why close by? Two mechanisms show up in other parts of the corpus under different names. First, training sets a ceiling. Agents trained on expert demonstrations can't go beyond what the people who built the data imagined Can agents learn beyond what their training data shows?. Second, reinforcement learning actively narrows behavior. In search agents, RL squeezes exploration down to a few strategies that maximize reward, the same "entropy collapse" seen in reasoning models, while fine-tuning on diverse examples keeps the range wider Does reinforcement learning squeeze exploration diversity in search agents?. A prompt that says "be novel" can't undo either one. The model reaches for the most surprising-sounding option within the space it already favors.

The less obvious finding is that adding more agents doesn't solve this. The 219K-idea study found that multi-agent designs did not widen the range Do AI research agents explore as broadly as human researchers?. A separate line of work helps explain why: a single LLM playing several personas reproduces what multi-agent debate does Can branching prompts replicate what multi-agent systems do?. If that works, it also means several copies of one model are mostly drawing from the same distribution. Their "different perspectives" are variations on one underlying view. Diversity among agents only helps when each one brings real domain expertise. Teams that are diverse without that expertise do worse than a single competent agent Does cognitive diversity alone improve multi-agent ideation quality?.

If you're looking for what actually widens the range, the corpus points to structure rather than exhortation. One approach first generates several different high-level abstractions (ways of framing the problem) and then solves within each. That forces breadth-first exploration, and at large compute budgets it beats sampling many solutions in parallel Can abstractions guide exploration better than depth alone?. Telling a model to be original changes how its ideas sound. Changing where it starts changes how far apart its ideas end up. The corpus doesn't have a study that tests novelty prompts directly against these structural fixes, so that comparison is still open.