Research map / Large language models / Language, safety and interpretability
Characterizing Truthfulness in Large Language Model Generations with Local Intrinsic Dimension
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes detecting LLM hallucinations by estimating the local intrinsic dimension (LID) of the last-token activations from a middle layer, using a distance-corrected MLE estimator, and reports it beats entropy, verbalized-uncertainty, and probe baselines on four QA tasks with Llama-2.
Problem
Predicting whether LLM generations are truthful (correct) in generative QA, where entropy-based uncertainty is intractable over open output spaces and verbalized or probe-based methods are unreliable.
Approach
Use the local intrinsic dimension of intermediate-layer representations (MLE estimator with a distance-aware polynomial correction, called LID-GeoMLE) as a per-sample truthfulness signal, based on the hypothesis that truthful outputs lie on lower-dimensional manifolds than outputs that mix human and model distributions.
Key results
- LID-GeoMLE averages 0.746 AUROC on Llama-2-7B versus 0.694 for the best entropy baseline (Semantic Entropy), and 0.764 on Llama-2-13B versus 0.736.
- On TriviaQA 0-shot, LID methods outperform the best entropy-based method by about 8 AUROC points.
- Aggregated LID across layers shows a hunchback shape, and detection AUROC tracks LID curves with a one-to-two-layer shift.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use distance-aware GeoMLE instead of vanilla MLE (ablation) | LID truthfulness detection · QA tasks (Llama-2-7B and Llama-2-13B) | AUROC: 0.758 -> 0.767 (CoQA, Llama-2-7B); 0.748 -> 0.775 (TriviaQA 0-shot, Llama-2-13B) | Helped |
| use LID-GeoMLE instead of the strongest entropy baseline | Llama-2-7B truthfulness detection · TriviaQA 0-shot | AUROC: 0.718 -> 0.771 | Helped |
| use LID-GeoMLE instead of the strongest entropy baseline | Llama-2-13B truthfulness detection · TriviaQA 0-shot | AUROC: 0.751 -> 0.775 | Helped |
| use CoQA samples as LID reference neighbors instead of HotpotQA samples (ablation) | LID truthfulness detection · HotpotQA | AUROC: 0.747 -> 0.753 | Helped |
| use TriviaQA samples as LID reference neighbors instead of TydiQA samples (ablation) | LID truthfulness detection · TriviaQA | AUROC: 0.765 -> 0.748 | No clear effect |
| increase the number of nearest neighbors (ablation) | LID truthfulness detection · QA tasks | AUROC | Helped |
| instruction-tune Llama-2-7B | Llama-2-7B · TriviaQA and TydiQA | accuracy | Helped |
| continue instruction tuning | Llama-2-7B · TriviaQA and TydiQA | local intrinsic dimension | Mixed |
| mix a human answer prefix with a model-generated continuation | TriviaQA answer representations · Controlled answer mixing | local intrinsic dimension | Helped |
Surprising findings
- Human ground-truth answers have lower LIDs than model-generated untruthful answers at every position, while mixed human-model continuations have higher LIDs than either.
- Using out-of-distribution reference samples from a different QA dataset as neighbors only slightly reduces detection performance.
Limitations
- Evaluation limited to QA tasks and English, and mostly to Llama-2 models at 7B and 13B.
- MLE estimates carry bias; the method depends on hyperparameters such as the number of neighbors T and reference set size n.
Experimental setup
Datasets: TriviaQA CoQA HotpotQA TydiQA-GP Super-NI
Benchmarks: CoQA TydiQA-GP TriviaQA HotpotQA
Metrics: AUROC
Baselines: Predictive Entropy LN-Predictive Entropy Semantic Entropy SAPLMA P(True) SAR LID-MLE KNN TwoNN t-SNE
Builds on: MLE intrinsic dimension estimator (Levina and Bickel) semantic entropy SAPLMA P(True)
Closest papers by meaning
- Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning (AAAI 2024 · similarity 0.90)
- Correlation Dimension of Autoregressive Large Language Models (NeurIPS 2025 · similarity 0.89)
- Distinguishing the Knowable from the Unknowable with Language Models (ICML 2024 · similarity 0.89)
- To Trust or Not to Trust? Enhancing Large Language Models' Situated Faithfulness to External Contexts (ICLR 2025 · similarity 0.89)
- Long-form factuality in large language models (NeurIPS 2024 · similarity 0.89)
- Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.