Research map / Large language models / Language, safety and interpretability
OpenEstimate: Evaluating LLMs on Probabilistic Estimation with Real-World Data
Alana Marzoev, Jillian Ross, Jacob Andreas
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine LLM prior with 5 samples instead of uninformative prior | o4-mini · OpenEstimate Glassdoor | win rate (%): 37.0% prior win -> 71.4% posterior win | Improved |
| increase baseline sample size from 5 to 30 | o4-mini prior · OpenEstimate Glassdoor | win rate of LLM prior vs baseline (%): 37.0% -> 8.7% | Hurt |
| increase baseline sample size from 5 to 30 | o4-mini prior · OpenEstimate NHANES | win rate of LLM prior vs baseline (%): 74.3% -> 37.8% | Hurt |
| increase baseline sample size from 5 to 30 | o4-mini prior · OpenEstimate Pitchbook | win rate of LLM prior vs baseline (%): 50.8% -> 50.8% | No clear effect |
| increase sample size from 5 to 30 for LLM posterior | o4-mini posterior · OpenEstimate Pitchbook | win rate (%): 69.6% -> 81.6% | Improved |
| increase sample size from 5 to 30 for LLM posterior | o4-mini posterior · OpenEstimate NHANES | win rate (%): 70.4% -> 50.4% | Hurt |
| use reasoning model o4-mini instead of Llama-3-8B | LLM prior · OpenEstimate NHANES | CRPS ratio: 19.17 -> 1.17 | Improved |
| use reasoning model o4-mini instead of GPT-4o | LLM prior · OpenEstimate Glassdoor | CRPS ratio: 3.31 -> 2.42 | Improved |
| use o3-mini instead of GPT-4o | LLM prior · OpenEstimate NHANES | CRPS ratio: 1.86 -> 1.35 | Improved |
| vary temperature or reasoning effort (ablation) | o4-mini and gpt-4o · OpenEstimate (all domains) | error ratio, CRPS | No clear effect |
| vary system prompt (ablation) | o4-mini and gpt-4o · OpenEstimate (all domains) | error ratio, CRPS | No clear effect |
| change elicitation protocol (direct, quantile, mean-variance) (ablation) | o4-mini and gpt-4o · OpenEstimate (all domains) | error ratio, CRPS | No clear effect |
Closest papers by meaning
- Large Language Models Must Be Taught to Know What They Don’t Know (NeurIPS 2024 · similarity 0.91)
- Benchmarking LLMs via Uncertainty Quantification (NeurIPS 2024 · similarity 0.91)
- LLM Processes: Numerical Predictive Distributions Conditioned on Natural Language (NeurIPS 2024 · similarity 0.90)
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs (ICLR 2024 · similarity 0.90)
- An Empirical Analysis of Uncertainty in Large Language Model Evaluations (ICLR 2025 · similarity 0.89)
- A Probabilistic Perspective on Unlearning and Alignment for Large Language Models (ICLR 2025 · similarity 0.89)
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.
