Empirical Test: Are LLMs Actually Confident During First Factual Recall?

Any-Chipmunk5480 · reddit · 2026-08-12

A developer ran over 60 local tests trying to find 'confidently wrong' factual recall within a local Gemma model's logprobs.

Key Findings:

This suggests that model 'confidence' often occurs post-generation rather than at the moment of initial recall.

Related event: Developer Tests Logprobs to Detect LLM Hallucinations(2 posts)→

Original post →

More from Research

Research channel →