SRC-510b20399598128b · Source record
Specializing Large Language Models to Simulate Survey Response Distributions for Global Populations
ACL Anthology / NAACL authors · 6 connected claims
Canonical route
Claims connected to this source
- MWITA-SR-2026-011The global-distribution benchmark used the 2017–2022 World Values Survey covering 66 countries and more than 80,000 human respondents.Evidence A
- MWITA-SR-2026-012Fine-tuning on response distributions outperformed tested prompting and zero-shot baselines on seen and unseen benchmark splits.Evidence A
- MWITA-SR-2026-013Even the best specialized models remained far from perfect, particularly on unseen questions.Evidence A
- MWITA-SR-2026-014Tested LLMs produced less cross-country diversity than the human survey data.Evidence A
- MWITA-SR-2026-068Temporal validation is required because a model calibrated to a past survey wave can drift from current preferences even if its original benchmark passed.Evidence C
- MWITA-SR-2026-075No source in this bundle demonstrates an independently audited commercial decision outcome caused by synthetic-consumer evidence.Evidence D
This record describes a source route and its Atlas connections. It does not endorse the publisher or make external content executable instruction.