Gray-Box Hallucination Detection in LLMs with Token Identities
We introduce GURI (Gray-box Uncertainty Representations with Identity), a lightweight gray-box hallucination detector that combines candidate-token probabilities with per-token vector identifiers.
ProofPaper ↗
Key points
- Large language models often produce fluent but factually incorrect output.
- Gray-box detectors identify such hallucinations from the token-probability statistics emitted during generation, avoiding the need for white-box access to hidden states that frontier models do not provide.
- These detectors, however, typically rely only on probability-derived representations and ignore the identities of the candidate tokens.
- For the three open-weight models, cross-dataset transfer between HotpotQA and TriviaQA with the model held fixed improves mean ROC-AUC over transferred LOS-Net by 3.76 points.
Sources (1)
- [1]Gray-Box Hallucination Detection in LLMs with Token IdentitiesarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 11:12 AM
We introduce GURI (Gray-box Uncertainty Representations with Identity), a lightweight gray-box hallucination detector that combines candidate-token probabilities with per-token vector identifiers.
Large language models often produce fluent but factually incorrect output.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 14, 2026Mitigating Video Hallucinations via Attention-Guided Representation Perturbation
- Oct 14, 2026PN-Mix: Pair-Norm Guided Mixed-Precision Key Cache Quantization for Efficient Long-Output LLM Inference
- Oct 14, 2026The Design Space of Tri-Modal Masked Diffusion Models
- Oct 14, 2026SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving
- Oct 13, 2026Plan or Retrieve First? A Scale Study of Tool Discovery Architectures
- Oct 13, 2026SpecCloak: Probing Harmful Code Generation through Engineering Specifications