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:
- During the first factual recall in a reasoning trace (before answer-directed self-conditioning occurs), the model rarely exhibits absolute confidence in a wrong fact. When the model first recalls something incorrect, it is actually internally uncertain (e.g., probability distributions between correct and incorrect answers are close).
- Hallucination Solidification: Once the model generates a wrong answer at a low probability, it reinforces itself in subsequent reasoning, quickly approaching 100% confidence in that wrong fact.
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)→
More from Research
- Paper Proves No Gradient Descent Stepsize Schedule Can Match Nesterov Acceleration — prof_grimmer · 2026-08-12
- Roboflow's Open-Source Trackers Library Adds McByte for Occlusion Handling — burny_tech · 2026-08-12
- Report: Ilya Sutskever's SSI Pivots to Test-Time Training for New Reasoning Engine — iruletheworldmo · 2026-08-12
- 300+ Real-World ML System Design Case Studies from 80+ Top Companies — mdancho84 · 2026-08-12
- Fine-tuning Muse Glimmer 30B Boosts Click Grounding Accuracy to 41% — mervenoyann · 2026-08-12
- Liquid Crow Released: Squeezing Physical Cognition into a 450M-Parameter Micro-World Model — helloiamleonie · 2026-08-12