New paper: LLM verbalized probabilities directly track internal uncertainty, not just training data
sineadwilliamso · x · 2026-10-06
Sinead Williamson and colleagues present arXiv work showing how LLM probability readouts respond to manipulated uncertainty sources in training data. Key findings:
- Both verbalized and logit-based probabilities are shaped by distributional frequencies and by explicit probabilistic assertions in the data; shifting training-data frequency shifts verbalized probabilities.
- Verbalized probabilities aren't merely echoing training text—they directly track internal uncertainty representations shared with logit-based frequencies.
- The coupling exceeds what independent tracking of the same sources would predict, so verbalized probabilities can serve as a probe into a model's internal distribution.
- Misaligned confidence expressions in training data cause miscalibration.
Practical upshot: asking a model for its confidence can meaningfully estimate internal uncertainty.
More from Research
- Huawei Noah's Tail-Influence Sampling Cuts CVaR Policy Evaluation MSE by Up to 76% — huawei-noah · 2026-10-06
- Google's KeyRec Achieves Best Long-Video VLM Results With Just 10% of Visual Token Budget — google · 2026-10-06
- 4DCodeBench Shows Frontier Models Reconstruct Static Scenes but Fail at Dynamics — 4DCodeBench · 2026-10-06
- OmniTaskonomy: Year-long study shows generation training can improve understanding tasks — WeijiaShi2 · 2026-10-06
- Newton proved the product rule without limits, using a discrete symmetric-difference trick — ctjlewis · 2026-10-06
- DeepMind's AI designs enzymes from scratch: 99x drug building block yield, plastic-eating at 90°C — 141_1337 · 2026-10-06