Can every single AI answer look fresh and new, yet all the answers together still sound basically the same?
Can high average novelty in AI outputs coexist with diversity collapse?
This explores whether AI outputs can each look fresh and original on their own while, taken together, they all end up resembling one another, so that novelty measured one output at a time hides sameness across the whole set.
This explores whether an AI can produce outputs that each seem new while the full set of outputs narrows toward the same few ideas. The corpus says yes, and the reason is that the two are measured in different ways. Novelty is usually judged one output at a time: is this response unusual compared with typical text? Diversity is a property of the whole group: do these responses differ from each other? A model can pass the first test every time and still fail the second, because many outputs can be unusual in the same way.
The clearest evidence comes from studies of many models side by side. Across more than 70 models answering open-ended prompts, researchers found an "Artificial Hivemind": separately built models give strikingly similar answers, because they share overlapping training data and similar alignment procedures Do different AI models actually produce diverse outputs?. Any single answer may read as creative. The problem only shows when you compare many of them. A creativity experiment shows the same thing at the human level. When writers used AI-generated ideas, the diversity of the group's stories fell for everyone, even though individual pieces didn't obviously get worse. When the AI only refined ideas the writers already had, diversity held up Does AI assistance homogenize or preserve creative diversity?. What matters is where the AI enters the process: at the source of the ideas, or later in polishing them.
Here is a twist you might not expect. Preference tuning (RLHF) actually increases surface variety in creative writing, meaning more varied words and sentence structures, while it reduces that variety in code Does preference tuning always reduce diversity the same way?. So a model can look more diverse on the measures that are easy to count while still converging on the same underlying ideas. A cultural-theory note in the collection argues that this is what makes AI sameness hard to spot: each output is tailored to your prompt, so it feels personal, and you never see the thousands of near-identical responses other people received Does AI homogenize culture the way mass media did?. The fact that outputs shift with every prompt and context Why does AI output change with every prompt and context? makes them look different from each other, but that variation doesn't guarantee they differ in substance.
The training-side research explains how the narrowing happens. Reinforcement learning tends to strengthen one dominant output format and suppress the others within the first epoch Does RL training collapse format diversity in pretrained models?. When rewards barely vary across attempts at a prompt, models drift toward generic templates that ignore the specific input Why do language models collapse into generic templates?, and search agents lose exploration breadth in the same way Does reinforcement learning squeeze exploration diversity in search agents?. The most striking result: if you give a base model enough attempts, it eventually finds more distinct solutions than its post-trained version. Post-training improves results on easy cases but removes rare solutions that were still reachable Do base models find more solutions than post-trained ones?. Even frontier research agents mostly recombine known techniques, so what looks like novelty is often reassembly of familiar parts Do frontier AI agents actually conduct novel research or just optimize?.
The practical takeaway: to judge whether AI is helping creativity, don't ask whether any one output is novel. Generate many outputs and check whether they disagree with each other. The corpus also suggests fixes. Pooling experience across different agents keeps progress going better than evolving each agent in isolation Does sharing experience across agents beat isolated evolution?. But diversity alone isn't enough: diverse multi-agent teams without real domain expertise do worse than a single competent agent Does cognitive diversity alone improve multi-agent ideation quality?.
Sources 12 notes
INFINITY-CHAT analyzed 70+ models across 26K open-ended queries and found an "Artificial Hivemind" effect: models independently generate strikingly similar or identical responses due to overlapping training data and alignment procedures, undermining the diversity benefits of model ensembles.
In a preregistered experiment, AI-generated ideas reduced collective diversity for all writers, while AI that refined existing ideas kept diversity intact. Non-native English speakers contributed more diversity than native speakers, but only AI ideation erased this advantage.
RLHF reduces lexical-syntactic diversity in code generation but increases it in creative writing. The direction depends on what each domain incentivizes: code rewards convergence toward correct solutions, while creative writing rewards stylistic distinctiveness.
AI mass-generates similar flows disguised as personalized outputs, suppressing novelty more deeply than pre-stamped commodities because contextual customization makes homogeneity invisible to individual users. Evidence: independent LLMs converge on similar outputs despite nominal competition.
AI outputs exhibit essential mutability—they vary with sampling, prompt wording, and audience interpretation. This is not a defect but a defining feature of tokens as media, making them fundamentally different from fixed commodities and resistant to traditional quality assurance.
Show all 12 sources
Controlled experiments show RL consistently amplifies one format distribution from pretraining within the first epoch while collapsing alternatives. The winning format depends on model scale, not necessarily performance, and is largely hidden when starting from proprietary pretrained models.
When within-prompt reward variance is low, task gradients weaken and regularization dominates, pushing policies toward generic outputs. SNR-Aware Filtering—selecting high-variance prompts before updates—recovers performance across tasks and scales.
RL training compresses behavioral diversity in search agents through the same entropy collapse mechanism documented in reasoning—policies converge on narrow reward-maximizing strategies. SFT on diverse demonstrations preserves exploration breadth, suggesting diversity-preservation techniques are essential for RL search scaling.
Across 14 model pairs and three agentic benchmarks, base models equipped with only relaxed system prompts eventually surpass post-trained counterparts in pass@K coverage as rollout budget grows. Post-training bimodalizes task outcomes, sharpening performance on easy cases while eliminating rare-but-reachable solutions.
Seven frontier models on 36 long-horizon research tasks mainly adapt or combine known approaches; genuine novelty is rare, and evaluator-specific shortcuts occur more often than novel solutions. Performance varies substantially across runs.
Group-Evolving Agents outperformed isolated tree-based self-evolution by 14–20 percentage points by explicitly pooling code patches and execution traces within each generation. Analysis showed five of eight key tool improvements came from different parent agents, proving the sharing mechanism itself—not just more search—drove the gains.
Multi-agent teams substantially outperform solo ideation, but only when members possess genuine senior knowledge. Diverse teams without expertise underperform even a single competent agent, because cognitive stimulation without expertise triggers process losses instead of insight.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Human diversity fuels collective creativity that large language models cannot simulate or sustain
- The Homogenizing Effect of Large Language Models on Human Expression and Thought
- Sharpening Tax in Post-Training
- Homogenization Effects of Large Language Models on Human Creative Ideation
- What Does It Take to Be a Good AI Research Agent? Studying the Role of Ideation Diversity
- What and Whose Knowledge? Measuring Epistemic Diversity in Large Language Models
- Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining
- RAGEN-2: Reasoning Collapse in Agentic RL