Semantic Entropy

Before calculating entropy, token predictions are grouped by semantic equivalence rather than surface form. For example, 'Paris,' 'the French capital,' and 'City of Light' are clustered as a single meaning. This prevents a model that is certain about a fact but varies in phrasing from being penalized with a high uncertainty score.

Once predictions are clustered, standard entropy is applied to the probability mass of each semantic cluster. The formula H = -Σ p(c) log p(c) operates on cluster probabilities, not individual token probabilities. A low score indicates the model is semantically certain; a high score signals genuine indecision across multiple distinct factual possibilities.

Semantic entropy serves as a predictive signal for confabulation. Research shows that high semantic entropy strongly correlates with incorrect or non-factual outputs. By flagging generations with elevated entropy scores, systems can trigger fallback mechanisms, request human review, or invoke retrieval-augmented generation to re-ground the response.

The core innovation is the disentanglement of two uncertainty types:

• Lexical variation: Multiple valid ways to express the same fact
• Factual indecision: The model genuinely does not know the answer Traditional token-level entropy conflates both, while semantic entropy isolates the latter, providing a cleaner measure of a model's knowledge state.

A common implementation uses Natural Language Inference (NLI) models to determine if two generated sequences entail each other bidirectionally. If sequence A entails B and B entails A, they are semantically equivalent and merged into the same cluster. This graph-based approach builds meaning-equivalence classes before entropy computation.

Semantic entropy directly supports confidence calibration by providing a more reliable probability that a model's output is correct. Unlike raw token probabilities, which are often overconfident, semantic entropy scores can be thresholded to achieve a target precision-recall tradeoff, enabling deployment in high-stakes domains where factual reliability is non-negotiable.

A related but distinct metric that quantifies the uncertainty in a model's output distribution at the token level. Unlike semantic entropy, it does not cluster by meaning, making it a simpler but noisier signal for detecting confabulated text. High hallucination entropy often correlates with low factual precision.

An automated evaluation process that measures the alignment between a generated summary and its source document. It directly penalizes contradictions and hallucinations. Key methods include:

• NLI-based scoring: Using Natural Language Inference to detect contradictions.
• QA-based scoring: Generating questions from the summary and verifying answers against the source.
• Semantic overlap: Measuring embedding similarity between the summary and source.

A prompting technique where a language model systematically verifies its own output to reduce hallucinations. The process involves:

1. Drafting an initial response.
2. Generating a series of independent fact-checking questions based on the draft.
3. Answering those questions without the influence of the initial draft.
4. Revising the final output to align with the verified answers. This creates a self-correcting loop.

The process of aligning a model's predicted probability of correctness with its actual empirical accuracy. A perfectly calibrated model that says it is 90% confident is correct exactly 90% of the time. This is crucial for trusting model outputs and is often measured using Expected Calibration Error (ECE). Poor calibration can make a model seem confidently wrong.

A model-agnostic, distribution-free framework that wraps any predictive model to produce statistically rigorous prediction sets with a guaranteed coverage probability. Instead of a single point prediction, it outputs a set of plausible answers. For a 90% confidence level, the true answer is guaranteed to be in the set 90% of the time, providing a robust measure of uncertainty.

A fine-grained evaluation metric that decomposes a generated biography (or any text) into atomic facts and verifies each one independently against a trusted knowledge source like Wikipedia. The final score is the percentage of atomic facts that are supported. This provides a precise, human-interpretable measure of factual precision at the level of individual claims.