Abstract
Large Language Models (LLMs) often struggle to communicate their internal certainty, leading to hallucinations. This research investigates whether internal token probabilities or verbalized self-reports are more "honest" indicators of a model's knowledge. By evaluating models like Llama2, Mistral, and GPT-3.5, we find that probabilistic confidence is a significantly more reliable predictor of accuracy than verbalized confidence. However, a "translation gap" exists where models fail to accurately express their internal "gut feelings" in natural language, often exhibiting verbalized overconfidence.
Introduction
The reliability of Large Language Models (LLMs) hinges on their ability to recognize their own knowledge boundaries. When pushed beyond these limits, models frequently hallucinate. To address this, researchers evaluate self-awareness through two primary lenses: Probabilistic Confidence (internal log-probabilities) and Verbalized Confidence (direct natural language assertions of certainty). This paper explores the relationship between these two methods and determines which provides a more truthful representation of model knowledge.
Methods
The study evaluates four models: Llama2-7B-Chat, Mistral-7B-Instruct, GPT-3.5-Turbo (ChatGPT), and GPT-3.5-Turbo-Instruct.
- Probabilistic Confidence: Calculated using average log-probability (perplexity) of generated answers, converted to binary labels via an optimal threshold determined on a validation set.
- Verbalized Confidence: Models were prompted to append "certain" or "uncertain" to their responses.
- Analysis: Correlation (Spearman/Kendall) and binning techniques were used to measure the alignment between internal signals and verbalized output.
Datasets
Two primary benchmarks were utilized:
- Natural Questions (NQ): Real-world factual queries from Google search data.
- Parent-Child Dataset: Designed to test the impact of question frequency by pairing common facts (e.g., "Who is the parent of X?") with less frequent inverse facts (e.g., "Who is the child of Y?").
Key Insights
The Translation Gap and Overconfidence
There is a significant disconnect between internal probability and verbalized output. Models are consistently "cocky" when expressing certainty in words, overestimating their correctness compared to their internal signals. This suggests that LLMs possess a "gut feeling" they cannot yet accurately articulate.
The Frequency Paradox
LLMs demonstrate a better perception of their knowledge boundaries on less frequent (unfamiliar) questions. While they tend to be overconfident on common questions, their internal probabilities drop sharply on obscure queries, indicating a clearer grasp of their own ignorance.
Calibration Requirements
Probabilistic confidence, while more accurate, requires dataset-specific calibration. The optimal threshold for "certainty" varies significantly across different domains, making it less flexible than verbalized confidence in zero-shot scenarios.
Key Results
Probabilistic perception consistently outperformed verbalized confidence across all tested models. As shown in the table below, alignment scores—which measure how well confidence matches actual correctness—are markedly higher for probabilistic strategies than for verbalized ones.
For instance, on the NQ dataset, Llama2 achieved an alignment of ~72.5% using probabilities, compared to only ~45.1% when verbalizing confidence. Similarly, ChatGPT's probabilistic alignment jumped to 88.2% on the "Child" dataset, highlighting the model's increased sensitivity to its own uncertainty in unfamiliar contexts.
Limitations and Future Work
A major hurdle for implementing probabilistic confidence is the "black-box" nature of many commercial APIs, which do not provide token-level log-probabilities. Additionally, the need for a validation set to tune thresholds limits flexibility. Future research should focus on calibration training to help models better interpret their internal probability distributions and express that certainty accurately in natural language, bridging the gap between what they "feel" and what they say.
Summary updated