Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT
2025/09/30 by Guy Bar-Shalom, Fabrizio Frasca, Bar-Shalom, Guy +7
Medicine · Neuroscience · #Advanced Neuroimaging Techniques and Applications #Biofield Effects and Biophysics #FOS: Computer and information sciences #Functional Brain Connectivity Studies #Machine Learning (cs.LG)
paper · pdf · doi:10.48550/arxiv.2510.00296
openalex publication_date 2025/09/30 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Abstract
Detecting hallucinations in Large Language Model-generated text is crucial for their safe deployment. While probing classifiers show promise, they operate on isolated layer-token pairs and are LLM-specific, limiting their effectiveness and hindering cross-LLM applications. In this paper, we introduce a novel approach to address these shortcomings. We build on the natural sequential structure of activation data in both axes (layers × tokens) and advocate treating full activation tensors akin to images. We design ACT-ViT, a Vision Transformer-inspired model that can be effectively and efficiently applied to activation tensors and supports training on data from multiple LLMs simultaneously. Through comprehensive experiments encompassing diverse LLMs and datasets, we demonstrate that ACT-ViT consistently outperforms traditional probing techniques while remaining extremely efficient for deployment. In particular, we show that our architecture benefits substantially from multi-LLM training, achieves strong zero-shot performance on unseen datasets, and can be transferred effectively to new LLMs through fine-tuning. Full code is available at https://github.com/BarSGuy/ACT-ViT.
Citations
- SafeSteer: Interpretable Safety Steering with Refusal-Evasion in LLMs
- Knowing Before Saying: LLM Representations Encode Information About Chain-of-Thought Success Before Completion
- Improving Preference Extraction In LLMs By Identifying Latent Knowledge Through Classifying Probes
- Probing Internal Representations of Multi-Word Verbs in Large Language Models
- Linear Probe Penalties Reduce LLM Sycophancy
- LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations
- Estimating Knowledge in Large Language Models Without Generating a Single Token
- Spectral Editing of Activations for Large Language Model Alignment
- A Comprehensive Survey of Hallucination in Large Language, Image, Video and Audio Foundation Models
- The Platonic Representation Hypothesis
- Constructing Benchmarks and Interventions for Combating Hallucinations in LLMs
- Characterizing Truthfulness in Large Language Model Generations with Local Intrinsic Dimension
- A Language Model's Guide Through Latent Space
- INSIDE: LLMs' Internal States Retain the Power of Hallucination Detection
- A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models
- Weakly Supervised Detection of Hallucinations in LLM Activations
- A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions
- The Curious Case of Hallucinatory (Un)answerability: Finding Truths in the Hidden States of Over-Confident Large Language Models
- The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets
- Mistral 7B
- The Troubling Emergence of Hallucination in Large Language Models -- An Extensive Definition, Quantification, and Prescriptive Remediations
- Representation Engineering: A Top-Down Approach to AI Transparency
- Attention Satisfies: A Constraint-Satisfaction Lens on Factual Errors of Language Models
- A Stitch in Time Saves Nine: Detecting and Mitigating Hallucinations of LLMs by Validating Low-Confidence Generation
- Inference-Time Intervention: Eliciting Truthful Answers from a Language Model
- Detecting and Mitigating Hallucinations in Multilingual Summarisation
- HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models
- The Internal State of an LLM Knows When It's Lying
- SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models
- LLaMA: Open and Efficient Foundation Language Models
- Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation
- Equivariant Architectures for Learning in Deep Weight Spaces
- Discovering Latent Knowledge in Language Models Without Supervision
- Predicting Fine-Tuning Performance with Probing
- Relative representations enable zero-shot latent space communication
- Git Re-Basin: Merging Models modulo Permutation Symmetries
- Looking for a Needle in a Haystack: A Comprehensive Study of Hallucinations in Neural Machine Translation
- Language Models (Mostly) Know What They Know
- Survey of Hallucination in Natural Language Generation
- A Token-level Reference-free Hallucination Detection Benchmark for Free-form Text Generation
- Probing Classifiers: Promises, Shortcomings, and Advances
- Probing Classifiers: Promises, Shortcomings, and Advances
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- A Tale of a Probe and a Parser
- PyTorch: An Imperative Style, High-Performance Deep Learning Library
- Designing and Interpreting Probes with Control Tasks
- Similarity of Neural Network Representations Revisited
- Linguistic Knowledge and Transferability of Contextual Representations
- Analysis Methods in Neural Language Processing: A Survey
- HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering
- Insights on representational similarity in neural networks with canonical correlation
- What you can cram into a single vector: Probing sentence embeddings for linguistic properties
- TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for\n Reading Comprehension
- Understanding intermediate layers using linear classifier probes
- Fine-grained Analysis of Sentence Embeddings Using Auxiliary Prediction Tasks
- Beyond Next Token Probabilities: Learnable, Fast Detection of Hallucinations and Data Contamination on LLM Output Distributions
Cited by
Related