Semantic Entropy

Unlike naive token-level entropy, semantic entropy groups meaning-equivalent generations before calculation. Two sentences with different words but identical meaning are treated as the same outcome, preventing lexical variation from inflating uncertainty scores. This clustering typically uses Natural Language Inference (NLI) models or bidirectional entailment checks to determine semantic equivalence.

Semantic entropy specifically targets epistemic uncertainty—the reducible uncertainty from knowledge gaps—rather than aleatoric uncertainty from inherent linguistic ambiguity. High semantic entropy indicates the model lacks knowledge and is likely to hallucinate, while low semantic entropy with high lexical variation simply indicates paraphrasing ability.

The metric follows a three-step pipeline:

• Multi-sample generation: Produce multiple responses to the same prompt at non-zero temperature
• Semantic clustering: Group outputs by meaning equivalence using NLI entailment scores
• Entropy computation: Calculate entropy over the cluster distribution, not individual token sequences This yields a scalar value where higher scores indicate greater factual uncertainty.

Semantic entropy achieves strong AUROC scores on hallucination detection benchmarks, significantly outperforming naive token-level entropy and perplexity-based methods. It is particularly effective at identifying sentence-level confabulations where the model invents plausible-sounding but factually unsupported claims, making it valuable for production LLM monitoring systems.

Semantic entropy pairs naturally with conformal prediction frameworks. The entropy score can serve as a nonconformity measure, enabling the construction of prediction sets with formal coverage guarantees. When semantic entropy exceeds a calibrated threshold, the system can trigger retrieval, abstention, or human review workflows.

The primary cost is generating multiple samples per query, which increases inference latency linearly with sample count. Practical deployments often use 5-10 samples and optimize clustering with efficient NLI models. For high-throughput systems, predictive entropy estimation methods that approximate the distribution without full multi-sampling are an active research area.

A metric evaluating whether all factual claims in a generated text are supported by a source document. It measures the alignment between the output and the provided grounding context, often using Natural Language Inference (NLI) to classify if a hypothesis is entailed by the source. This is a foundational measure for summarization and RAG systems.

An automated evaluation score that determines if a generated summary or response can be logically deduced from the input source without introducing extraneous information. It uses NLI-based evaluation to detect hallucinations and contradictions, ensuring the output is a true reflection of the provided context rather than the model's internal knowledge.

The reducible uncertainty in a model's prediction caused by a lack of knowledge or training data. This can be decreased by collecting more data or refining the model architecture. In contrast to aleatoric uncertainty, which is inherent noise in the data, epistemic uncertainty signals where a model is aware of its own ignorance, a key signal for hallucination risk.

A zero-resource hallucination detection method that samples multiple responses from a black-box LLM and checks for factual inconsistency. It leverages the principle that hallucinated facts are stochastically unstable, meaning they will vary across samples, while grounded facts remain consistent. This approach requires no external knowledge base.

A prompting technique where an LLM:

• Drafts an initial response
• Generates a series of independent verification questions to fact-check its own work
• Produces a final, corrected answer This self-interrogation loop significantly reduces hallucinations by forcing the model to scrutinize its own claims.

A human-aligned evaluation metric that breaks a long-form generation into atomic facts and verifies each against a trusted knowledge base like Wikipedia. It calculates the percentage of supported facts, providing a granular, interpretable score of factual precision in biographies and other knowledge-intensive text.