Research map / Large language models / Language, safety and interpretability
Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Infinity-Chat, a 26K real-world open-ended query dataset with a 6-category/17-subcategory taxonomy and 31,250 dense human annotations, to study intra- and inter-model homogeneity ("Artificial Hivemind") and whether LMs, reward models, and LM judges calibrate to human preferences on open-ended responses.
Problem
LMs produce homogeneous outputs for open-ended queries lacking a single ground truth, and existing evaluations of output diversity are limited to narrow tasks or single models; it is also unclear whether automatic raters track human preferences over equally good, diverse responses.
Approach
Mine real-world open-ended queries from WildChat with GPT-4o classification, build a taxonomy of open-ended query types, and collect dense (25-annotator) human ratings on alternative responses to measure cross-model similarity and rater calibration on contested or similar-quality responses.
Key results
- Under top-p 0.9 and temperature 1.0, 79% of queries have average intra-model pairwise similarity above 0.8 for the same model.
- Inter-model average pairwise similarity ranges from 71% to 82%; DeepSeek-V3 and qwen-max-2025-01-25 reach 0.82.
- Correlations between model ratings and human ratings drop on similar-quality and high-disagreement subsets across LMs, reward models, and LM judges.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use min-p decoding instead of top-p sampling | LM generation on Infinity-Chat100 · Infinity-Chat100 | within-model response similarity | Mixed |
| Use min-p decoding (top-p = 1.0, min-p = 0.1, temperature = 2.0) | LM generation · Infinity-Chat100 | percentage of response pairs with similarity >0.8: 61.2% | Hurt |
| Generate responses from different models rather than one model | 25 LMs · Infinity-Chat100 | inter-model response similarity: 71% to 82% | Hurt |
| Paraphrase the input prompt | 42 models · 30 prompts and their paraphrases | response similarity: 0.821 vs. 0.781 | No clear effect |
| Evaluate responses on similar-quality subsets rather than the full set | LMs, reward models, and LM judges · Infinity-Chat absolute and pairwise ratings | correlation with human ratings | Hurt |
| Evaluate examples with high annotator disagreement | LMs, reward models, and LM judges · Infinity-Chat absolute and pairwise ratings | correlation with human ratings | Hurt |
| Use prompts judged highly open-ended rather than low in open-endedness | 42 models · Infinity-Chat human-validated prompts | average response similarity: 0.837 -> 0.800 | No clear effect |
Surprising findings
- Min-p decoding still yields 61.2% of response pairs above 0.8 similarity.
- On average about 8 unique models appear among the top-50 most similar responses per query, so cross-model similarity can exceed intra-model similarity.
Limitations
- English-only prompts from WildChat limit generality across languages and cultures.
- Diversity measured via embedding similarity may miss finer creative variation; quality is not evaluated.
- Taxonomy labels from GPT-4o agree with human annotators only 74.7% of the time; causal sources of homogeneity are not established.
Experimental setup
Datasets: WildChat-1M
Benchmarks: Infinity-Chat Infinity-Chat100 RewardBench
Metrics: average pairwise cosine similarity of sentence embeddings Spearman correlation Pearson correlation Shannon entropy of annotations avg unique source models in top-N
Baselines: min-p decoding RewardBench reward models GPT-4o judge Prometheus
Builds on: WildChat min-p decoding Prometheus RewardBench HelpSteer3
Closest papers by meaning
- ULTRAFEEDBACK: Boosting Language Models with Scaled AI Feedback (ICML 2024 · similarity 0.88)
- A Statistical Framework for Ranking LLM-based Chatbots (ICLR 2025 · similarity 0.88)
- Generative Monoculture in Large Language Models (ICLR 2025 · similarity 0.87)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.87)
- The Emergence of Abstract Thought in Large Language Models Beyond Any Language (NeurIPS 2025 · similarity 0.87)
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.