Apple Research Finds LLM Verbalized Confidence and Internal Probabilities Are Coupled
sineadwilliamso · x · 2026-10-06
Apple researchers including Sinead Williamson published an arXiv paper, "Verbalized and Internal Probabilities Are Coupled in Large Language Models."
- LLMs carry internal uncertainty in their sampling distributions and can also state confidence verbally; whether these two readouts align was previously unknown.
- By intervening on uncertainty sources in training and in-context data, the team shows both internal and verbalized probabilities are affected by distributional and asserted uncertainty; mismatched confidence expressions in training data cause miscalibration.
- Key finding: verbalized and internal probabilities align beyond chance, so verbalized confidence can serve as a probe into a model's internal distribution.
More from Research
- Math lacks empirical tradition: Wolfskehl Prize drew 1,000 wrong Fermat proofs — RexDouglass · 2026-10-06
- NanoGPT speedrun sets record: 11.3% faster via architecture-only change, paper coming — yoavartzi · 2026-10-06
- CMU's PNAS paper warns: the more you automate science with AI, the less you see — Dr_Atoosa · 2026-10-06
- Embodied Analysis launches unified eval platform for physical agents and world models — qinzytech · 2026-10-06
- COLM paper on the softmax gradient bottleneck; author briefly crushed the nanogpt speedrun record with a new LM head — yoavartzi · 2026-10-06
- Idiom Map: tracing agent slang and subgroups in 'neuralese' across AI Village and wiki incidents — kalladomcdowell · 2026-10-06