Hallucination Entropy

Standard token-level entropy conflates lexical variation with factual uncertainty. Semantic entropy clusters token predictions by their underlying meaning before calculating uncertainty. A model might assign high probability to 'Paris,' 'the City of Light,' and 'France's capital' for the same slot, indicating low semantic entropy and high confidence. Conversely, equal probability across 'Paris,' 'Berlin,' and 'Tokyo' signals high semantic entropy and a likely hallucination.

High hallucination entropy is a leading indicator of non-factual output. Key predictive signals include:

• Flat output distributions: No single token dominates, indicating the model has no grounded answer.
• High variance across layers: Entropy that spikes in later transformer layers suggests the model is fabricating rather than retrieving.
• Mismatch with source grounding: When entropy is low (model is confident) but the statement contradicts a provided context, it signals a deeply memorized hallucination.

Confidence calibration aligns a model's self-reported probability with its actual accuracy. Hallucination entropy is the raw signal; calibration is the post-processing correction. A well-calibrated model with high entropy will correctly express low confidence. A poorly calibrated model may have low entropy (high confidence) on a hallucination. Conformal prediction uses entropy distributions to produce statistically rigorous uncertainty sets with guaranteed coverage probabilities.

Entropy can be measured at multiple granularities:

• Token-level: The standard softmax entropy for each predicted token. Sensitive to phrasing but computationally cheap.
• Span-level: Entropy calculated over multi-token entity spans, clustering equivalent phrasings.
• Sequence-level: Aggregated entropy across a full generated sentence. Semantic entropy operates here, using entailment models to group equivalent outputs before calculating uncertainty over meaning clusters.

Operationalizing hallucination entropy involves modifying the decoding process:

• Threshold gating: If a span's entropy exceeds a calibrated threshold, the system can trigger retrieval augmentation or refuse to answer.
• Entropy-based beam search: Penalize beams with anomalously high entropy to steer generation toward grounded paths.
• Contrastive decoding: Amplify the probability difference between a primary model and a weaker 'amateur' model; high entropy regions where both models agree on a hallucination are suppressed.

Hallucination entropy is a runtime symptom of data drift. When input distributions shift in production, the model encounters out-of-distribution queries. This manifests as elevated entropy and increased hallucination rates. Monitoring entropy distributions over time serves as an early warning system for model degradation, triggering retraining or knowledge base updates before factual errors impact users.

A model-agnostic framework that wraps any predictive model to produce statistically rigorous prediction sets with a guaranteed coverage probability. Unlike raw entropy scores, conformal prediction provides a formal, distribution-free guarantee that the true output falls within the predicted set at a user-specified confidence level (e.g., 90%). This is critical for high-stakes applications requiring hard uncertainty quantification.

The process of aligning a model's predicted probability of correctness with its actual empirical accuracy. A well-calibrated model that assigns 90% confidence to a set of predictions should be correct exactly 90% of the time. Poor calibration, where a model is overconfident in its hallucinations, directly undermines the utility of Hallucination Entropy as a decision signal.

A fine-grained evaluation metric that decomposes a generated biography into atomic facts and verifies each one independently against a trusted knowledge source like Wikipedia. This provides a factual precision score that complements entropy-based uncertainty measures. While Hallucination Entropy predicts when a model might fail, FActScore measures how much it actually did.

A prompting technique where a language model first drafts a response, then generates a series of independent fact-checking questions to systematically verify its own output. This creates a self-correcting loop that reduces hallucination rates by forcing the model to adopt a critical stance toward its own initial generation, effectively lowering output entropy through structured deliberation.

The process of breaking down a complex statement into minimal, self-contained, indivisible units of information, each expressed as a single sentence. This is the foundational preprocessing step for fine-grained verification systems. By isolating claims, a high-entropy paragraph can be segmented to pinpoint exactly which atomic fact is uncertain, rather than discarding the entire output.