A Survey on LLM Confidence Estimation Methods
Disneyskidney · reddit · 2026-07-18
This lengthy post systematically outlines the main methods for LLM uncertainty/confidence estimation and compares the black-box and white-box approaches.
Main Motivations
- LLM teams increasingly rely on "model judges" for complex decisions, and confidence scores help active learning, prompt optimization, and safety classifiers work more robustly.
- However, "uncertainty quantification" is still in its early stages, with no universally accepted best solution yet.
Methods Discussed
- Text-based methods:
- Asking the model directly to report "how sure it is," though empirically scores often cluster at the high end with poor differentiation.
- Estimating confidence using linguistic hedges (like maybe, possibly), but reliability is limited.
- Inferring uncertainty from response length/verbosity, which is relatively more effective on reasoning models.
- Token-based methods:
- Looking directly at the probability distribution of answer tokens, suitable for automation but trickier for multi-token scenarios.
- Using approaches like "P(True)" by feeding the model's own answer back to ask if it's correct, mitigating multi-token issues.
Core Perspectives
- White-box methods require model weights and are typically only applicable to open-source models.
- Black-box methods only look at output tokens or logprobs, making them suitable for closed-source models and more widely applicable in practice.
- The author mentions comparing 8 black-box methods and 1 white-box method to answer: which confidence estimation method is the best.
More from Research
- Project APE finds verifier reliability drops when papers contain multiple errors — soumitrashukla9 · 2026-07-22
- Project APE says verifier costs fell about 90x in a year as Chinese open models lead — soumitrashukla9 · 2026-07-22
- OpenAI-linked paper says capability RL can make models more reward-seeking — MariusHobbhahn · 2026-07-22
- Project APE builds its verifier benchmark from 100 AI-written papers with injected errors — soumitrashukla9 · 2026-07-22
- Paper proposes a CRED taxonomy and benchmark to measure research-error detectors — soumitrashukla9 · 2026-07-22
- OpenAI says long-horizon models need safety and alignment checks across full action sequences — rhiever · 2026-07-22