First screen, captured 4 days ago.

Using LLM for confidence scoring is a bad idea

What HN thinks

A room full of people rating LLM confidence scores 7/10, with low confidence.

Ask better: rubrics, small scales
LLM scores become usable if you drop 0-100 scales for binary or small enums, explicit rubrics, deduction-based grading, and multiple weighted dimensions.
Self-reported confidence isn't real
LLMs have no introspective access to their own statistics, so a stated confidence is unvalidated numerology without a calibration set and real data.
Calibration needs training and data
Proper confidence comes from supervised calibration, Brier-penalized RL, fine-tuning, logprobs over constrained outputs, or dedicated decision models, not prompting.

Consensus: Most agree that raw, prompted 0-100 confidence numbers from an LLM shouldn't be trusted without validation against labeled data; disagreement is over whether rubrics or training can make them useful.

Recommended in the comments

scikit-learn ↗ (library): Its calibration module was cited as the established science for turning classifier outputs into probabilities.
Liquid AI Decision Models ↗ (product): Several commenters said these are purpose-built for confidence scoring, unlike prose-generating LLMs.
SHAP ↗ (library): Mentioned as a proper interpretation method from pre-LLM ML practice.
LIME ↗ (library): Cited alongside SHAP as a serious model interpretation approach.
llama.cpp (tool): Can expose next-token likelihoods over constrained label tokens without patching.

Summarised by Claude from the comment thread 3 days ago.

Points: 185
Comments: 100
Best rank: #6
Posted: 2026-10-21 by jimauthors
Category: Ai ml

llm calibration confidence-scores ai-agents uncertainty forecasting

All captures:

Full page

Captured 4 days ago. Click for the whole screenshot.

Looks similar

Screenshot of ihavemanythoughts.substack.com: Peak LLM?

Peak LLM?(ihavemanythoughts.substack.com)

55 points by TisButMe 3 years ago | 40 comments

Screenshot of isaacsu.com: Don't credit the LLM

Don't credit the LLM (isaacsu.com)

17 points by isaacsu 2 months ago | 18 comments

Screenshot of camhashemi.com: Clever, Brave, Persistent

Clever, Brave, Persistent (camhashemi.com)

14 points by 1penny42cents 2 years ago | 3 comments

Screenshot of florian.github.io: That XOR Trick (2020)

That XOR Trick (2020)(florian.github.io)

293 points by hundredwatt 1 year ago | 120 comments

The risk of AI isn't making us lazy, but making "lazy" look productive

17 points by acmerfight 7 months ago | 17 comments

Screenshot of softwaredoug.com: Bag of Decisions Reranker

Bag of Decisions Reranker (softwaredoug.com)

63 points by zdw 15 days ago | 10 comments

Blog post describing an experimental search reranker where an LLM generates per-query yes/no relevance criteria, treating queries as a "bag of decisions."

search-relevance reranking llm information-retrieval