Study Reveals Statistical Self-Consistency Flaws in LLMs
Patrik Wolf · hf · 2026-07-17
**Research Core** The in-context learning of current LLMs is often viewed as conditional inference, but do model outputs adhere to basic probability identities? This paper explores whether LLM estimates follow the principles of statistical self-consistency. **Experimental Method** - Uses a binary tree as an evaluation framework to recursively partition the overall population into finer subgroups. - Prompts the model with subgroup descriptions, aggregates the obtained estimates, and compares them with overall estimates at different granularities. **Key Findings** - **Self-Consistency Violation**: Broad violations of basic probability consistency properties are observed across frontier models and multiple domains. - **Macro Fallacy**: Estimates reconstructed from fine-grained subgroups often align better with human reference data than direct overall estimates. This indicates that models possess relevant subgroup knowledge but cannot reliably extrapolate it to global aggregated estimates. **Conclusion** This disconnect in probabilistic reasoning demonstrates that statistical self-consistency is an unsaturated, reference-free new dimension for evaluating large models.
Related event: Study Reveals LLMs Lack Statistical Self-Consistency(2 posts)→
More from Research
- Baseten study finds new facts in LLM weights are fragile unless trained from many restatements — alex_verem · 2026-07-21
- Kimi K3 and Fable 5 now look much closer than the old open-vs-closed gap — FinanceYF5 · 2026-07-21
- uv-scripts/ocr returns to the top of Hugging Face datasets with a JSON model picker — vanstriendaniel · 2026-07-21
- DeepSearch-World trains web agents with 420K verifiable QA tasks — HKUST · 2026-07-21
- GigaAM Multilingual targets low-resource Central Asian ASR with 2M hours of audio — ai-sage · 2026-07-21
- WorldCupArena benchmarks language models on 104 football matches — Zhaokai Wang · 2026-07-21