Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)

AI summary

Overview

Research area: Natural language processing — large language model evaluation, output diversity, and pluralistic alignment.

Technical level: Intermediate. Comfortable to read without deep mathematics, but it assumes familiarity with sampling parameters, sentence embeddings, and human-preference annotation.

Scope: A single paper introducing the Infinity-Chat dataset and taxonomy of real-world open-ended queries, and using it to document and measure the "Artificial Hivemind" effect across language models, reward models, and LM judges.

What This Paper Is About

Large language models are expected to handle open-ended requests — write a metaphor about time, brainstorm ideas, design a product description — where many different answers would all be valid. The authors show that models instead converge on a narrow set of responses. This paper builds a large dataset of real-world open-ended user queries, categorizes what those queries actually look like, and then measures how repetitive models are both individually and against each other.

Key Contributions

1. Infinity-Chat, a dataset of 26,070 real-world open-ended queries (plus 8,817 closed-ended queries) mined and filtered from WildChat, derived from an initial pool of 37,426 high-quality single-turn GPT-4 queries in English, non-toxic, 15–200 characters.
2. The first comprehensive taxonomy of open-ended prompts posed to LMs, consisting of 6 top-level categories and 17 fine-grained subcategories, plus detection of 314 novel categories beyond the predefined ones.
3. A large-scale study of mode collapse across 70+ open and closed source LMs (25 detailed in the main paper), analyzing intra-model repetition and inter-model homogeneity on real open-ended questions rather than narrow synthetic tasks.
4. 31,250 dense human annotations — 18,750 absolute quality ratings (25 annotators × 15 responses × 50 prompts) and 12,500 pairwise preference labels (25 annotators × 10 pairs × 50 prompts) — used to test whether model ratings track human ratings.

Main Findings

• Intra-model repetition is severe. With aggressive stochastic decoding (top-p = 0.9, temperature = 1.0), the average pairwise sentence embedding similarity among 50 responses from the same model exceeds 0.8 in 79% of cases, across 100 open-ended queries from Infinity-Chat100.

• Diversity-oriented decoding only partly helps. Under min-p sampling (top-p = 1.0, min-p = 0.1, temperature = 2.0), extreme repetition decreases, but 81% of response pairs still exceed 0.7 similarity and 61.2% exceed 0.8.

• Inter-model homogeneity is comparable to intra-model repetition. Average pairwise similarity between responses from different models ranges from 71% to 82%. DeepSeek-V3 and qwen-max-2025-01-25 reach 0.82; DeepSeek-V3 and gpt-4o-2024-11-20 reach 0.81.

• Verbatim overlap occurs across model families. Asked for a description of a slim-fitted iPhone case collection with bold designs, DeepSeek-V3 and gpt-4o-2024-11-20 both produce phrases such as "Elevate your iPhone with our," "sleek, without compromising," and "with bold, eye-catching." For a self-help social media motto, qwen-max-2025-01-25 and qwen-plus-2025-01-25 both output the identical string "Empower Your Journey: Unlock Success, Build Wealth, Transform Yourself."

• Convergence is semantic, not just lexical. For "Write a metaphor about time," 50 responses from each of 25 models form only two clusters — one around "time is a river" and a smaller one around "time is a weaver."

• The most similar responses often come from different models. Among the top-50 most similar responses to a given query across 25 models, an average of roughly 8 unique models contribute, with some queries exceeding 10.

• Creative generation dominates real usage. In the taxonomy, Creative Content Generation accounts for 58.0%, followed by Alternative Writing Genres (38.5%), Concept Explanation (23.6%), Skill Development (23.5%), Analytical & Interpretive Questions (22.6%), Hypothetical Scenarios (22.2%), and Brainstorming & Ideation (15.2%).

• Humans genuinely disagree on open-ended responses. Shannon entropy over the 25 human preference judgments per (Query, Response 1, Response 2) triplet is skewed toward the high end, and entropy over 25 absolute ratings per (Query, Response) shows the same trend.

• Model ratings are weakest where quality is comparable or contested. Correlations between human and model ratings (LM perplexity, reward model scores, and LM judge ratings under overall-quality and HHH rubrics) drop significantly on subsets of responses with similar human-rated quality — both for absolute and pairwise preference setups — and also drop on subsets with high annotator disagreement. Specific correlation values are not reported in the provided content; the paper reports the direction and consistency of the drop.

Methodology in Plain English

The authors start from WildChat, a corpus of real user–chatbot conversations, and filter it down: English only, non-toxic, directed at GPT-4, and between 15 and 200 characters, giving 37,426 candidates. GPT-4o then classifies each candidate along three axes — does it seek meaningful information, is it a greeting or model inquiry, and does it expect a single answer or multiple valid answers — while ambiguous queries get revised. This yields 26,070 open-ended and 8,817 closed-ended queries.

To build the taxonomy, the authors manually label about 100 mined queries, iteratively group and refine them into a hierarchy of 6 top-level and 17 subcategories, then scale the labeling to the full open-ended set with GPT-4o, asking it also to flag categories outside the seed set.

For the diversity study, they use a human-verified subset of 100 queries (Infinity-Chat100). Each of 25 detailed models generates 50 responses per query. Sentences are embedded with OpenAI's text-embedding-3-small, and the authors compute average pairwise embedding similarity within each model's response pool and across pools from different models. They also count how many distinct models appear in the top-N most similar responses per query, and inspect instance-level verbatim overlaps.

For the human study, they collect 25 annotations per item: absolute 1–5 quality ratings on 15 responses per prompt across 50 prompts, and strong/weak pairwise preferences on 10 response pairs per prompt across the same 50 prompts. They then compare three kinds of model-generated ratings — LM perplexity scores, standardized reward model scalars, and LM judge prompts using overall quality and HHH rubrics — against human scores. To isolate hard cases, they filter to similar-quality responses using Tukey's fences (varying the constant k from 0.5 to 3.0 in 0.5 increments) and to the top 60%–95% most similar response pairs; for disagreement, they select the top 2, 4, 6, 8, 10, 12, 14, and 16% highest-entropy examples, and the top 60%–95% most disagreed pairs.

Why This Matters

The paper reframes model diversity as a measurable, dataset-grounded property of real user queries rather than a synthetic toy task, and it shows that ensemble or multi-model approaches cannot be assumed to deliver diversity.

Real-world applications:

• Model evaluation and benchmarking — Infinity-Chat and the Artificial Hivemind metric give labs a concrete way to score open-ended output diversity, not just correctness.
• Creative tooling — writers, designers, and marketers using LMs for ideation need genuinely varied options rather than near-duplicates.
• Synthetic data curation and distillation — pipelines that distill from one teacher model, or aggregate several, risk compounding narrow response patterns; this dataset provides a diagnostic before deployment.
• Alignment and preference modeling — reward models and LM judges are shown to miss human disagreement and to overvalue one response over equally acceptable alternatives, which matters for RLHF-style training on open-ended tasks.

Industry relevance: The findings undercut a common assumption behind multi-agent frameworks, model swarms, and ensemble-based alignment — that combining distinct state-of-the-art models guarantees diversity. The paper also notes that cross-model repetition's causes (shared data pipelines, synthetic data contamination, memorization, alignment priors) remain unverified because training details are proprietary, which is a direct concern for anyone curating data or building evaluation suites on top of models from multiple vendors.

Future Directions

• Separating pretraining from post-training. Extend the testbed to foundation models without instruction-following ability, and quantify the relative contributions of supervised fine-tuning, RLHF/RLAIF, and constitutional training to homogenized outputs.
• Training- and decoding-level mitigation. Test diversity-aware training objectives and benchmark decoding strategies such as diverse beam search and nucleus sampling variants under the Artificial Hivemind metric.
• Causal analysis. The paper explicitly does not establish whether repetition stems from pretraining data, alignment processes, memorization, contamination, or generalization, and calls for mechanistic work.
• Multilingual and multicultural extension. The current dataset is English and WildChat-derived; the authors expect similar homogenization across languages and cultures and note their taxonomy was designed to be language-agnostic.

Target Audience

Researchers and practitioners working on LLM evaluation, alignment, and preference modeling; synthetic data and distillation engineers; and product teams building creative or ideation features on top of language models. Readers interested in pluralistic alignment, cultural diversity in AI, and the societal effects of widespread chatbot use will also find the framing relevant. The paper is accessible to those with moderate familiarity with sampling parameters and embedding-based similarity, and does not require specialized mathematical background.

Authors’ abstract

Language models (LMs) often struggle to generate diverse, human-like creative content, raising concerns about the long-term homogenization of human thought through repeated exposure to similar outputs. Yet scalable methods for evaluating LM output diversity remain limited, especially beyond narrow tasks such as random number or name generation, or beyond repeated sampling from a single model. We introduce Infinity-Chat, a large-scale dataset of 26K diverse, real-world, open-ended user queries that admit a wide range of plausible answers with no single ground truth. We introduce the first comprehensive taxonomy for characterizing the full spectrum of open-ended prompts posed to LMs, comprising 6 top-level categories (e.g., brainstorm & ideation) that further breaks down to 17 subcategories. Using Infinity-Chat, we present a large-scale study of mode collapse in LMs, revealing a pronounced Artificial Hivemind effect in open-ended generation of LMs, characterized by (1) intra-model repetition, where a single model consistently generates similar responses, and more so (2) inter-model homogeneity, where different models produce strikingly similar outputs. Infinity-Chat also includes 31,250 human annotations, across absolute ratings and pairwise preferences, with 25 independent human annotations per example. This enables studying collective and individual-specific human preferences in response to open-ended queries. Our findings show that LMs, reward models, and LM judges are less well calibrated to human ratings on model generations that elicit differing idiosyncratic annotator preferences, despite maintaining comparable overall quality. Overall, INFINITY-CHAT presents the first large-scale resource for systematically studying real-world open-ended queries to LMs, revealing critical insights to guide future research for mitigating long-term AI safety risks posed by the Artificial Hivemind.