A new research paper, "The One-Word Census," reveals that 44 language models exhibit significant answer-choice conformity when asked to provide a single word from a broad category. Across 31 categories, models frequently converged on the same answer, with "serendipity" being the most common choice when prompted to "pick a word." The study found that newer flagship models tend to be more conformist, while persona-tuned models are more divergent. Notably, the latest flagship models from Claude and GPT lineages showed a reversal in conformity, suggesting a potential shift in their development. AI
IMPACT Reveals insights into the internal decision-making and convergence patterns of large language models, potentially influencing future model development and evaluation methods.
RANK_REASON Research paper published on arXiv detailing a study of language model behavior.
- alphaXiv
- arXiv
- CatalyzeX
- Claude
- DagsHub
- generative pre-trained transformer
- Gotit.pub
- Grok
- Hugging Face
- Qwen
- ScienceCast
- The One-Word Census
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
