Density Ridge Selective Prediction for LLM and VLM Hallucination Detection under Calibration Label Scarcity

Researchers propose a density ridge-based method for detecting hallucinations in large language and vision-language models that outperforms existing approaches by 5-20 AUROC points while requiring minimal calibration labels. The technique maps hidden state trajectories to a low-dimensional geometric skeleton, enabling robust hallucination detection even when training data is scarce.

This research addresses a critical challenge in deploying large language and vision models: distinguishing confident but incorrect outputs (hallucinations) from genuine knowledge. The study compares three categories of detection methods—unsupervised approaches like Semantic Entropy, supervised probes requiring labeled data, and the novel density ridge method—revealing a fundamental tradeoff between performance and data efficiency that has plagued the field.

The proposed approach leverages geometric properties of model internals rather than relying on external labels or semantic measures. By extracting kinematic features from hidden state trajectories and identifying the density ridge of their distribution, researchers create a low-dimensional skeleton representing the stochastic output space. This method achieves superior performance across seven diverse QA benchmarks including TriviaQA, GSM8K, and vision-language tasks, suggesting the approach captures fundamental properties of model behavior.

The label-scarcity protocol (200 calibration queries, 5 generations) reflects realistic deployment scenarios where collecting extensive labeled datasets proves expensive or infeasible. The 5-20 point AUROC improvement over supervised baselines like SAPLMA is substantial for production systems where hallucination detection directly impacts user safety and trust. The tempered degradation under label scarcity indicates the method generalizes better than existing supervised approaches.

For AI practitioners deploying LLMs in high-stakes applications—legal research, medical diagnosis, financial analysis—this work offers a practical path forward. The method's computational efficiency relative to ensemble-based approaches and its requirement for only modest calibration data make adoption feasible. Future research should validate performance on emerging model architectures and investigate whether the density ridge property holds across different model families and scales.

• →Density ridge method achieves 5-20 AUROC point improvements over existing hallucination detection approaches across seven QA benchmarks
• →The technique requires only 200 calibration samples and 5 model generations per query, making it practical for resource-constrained deployment scenarios
• →Supervised probes degrade sharply with label scarcity while the ridge-based approach maintains robust performance in low-data regimes
• →Geometry-based detection using hidden state trajectories reveals fundamental properties of model hallucination distribution
• →Results span nine distinct text and vision models, demonstrating cross-architecture generalization capability