Research map / Large language models / Language, safety and interpretability
HARP: Hallucination Detection via Reasoning Subspace Projection
Junjie Hu, Gang Tu, Cheng Shengyu, JinXin Li, Jinting Wang, Rui Chen, Zhilong Zhou, Dongbo Shan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| project hidden states onto reasoning subspace (vs raw features of same dimension) (ablation) | HARP, Qwen-2.5-7B-Instruct · NQ Open | AUROC: 62.9 -> 84.0 | Improved |
| project hidden states onto reasoning subspace (vs raw features) (ablation) | HARP, Qwen-2.5-7B-Instruct · TruthfulQA | AUROC: 70.7 -> 88.1 | Improved |
| project hidden states onto reasoning subspace (vs raw features) (ablation) | HARP, LLaMA-3.1-8B · NQ Open | AUROC: 70.4 -> 89.4 | Improved |
| use reasoning subspace basis instead of random basis vectors (ablation) | HARP, Qwen-2.5-7B-Instruct · NQ Open | AUROC: 67.6 -> 84.0 | Improved |
| use reasoning subspace basis instead of random basis vectors (ablation) | HARP, LLaMA-3.1-8B · TruthfulQA | AUROC: 75.8 -> 88.5 | Improved |
| HARP vs best baseline | HARP · TriviaQA | AUROC: +7.5% (Qwen), +16.6% (LLaMA) | Improved |
| set reasoning subspace dimension to 256 (about 5% of hidden size) (ablation) | HARP · QA datasets | AUROC: range 32 to 1024 tested | Mixed |
| train detector on TriviaQA, test on NQ Open (ablation) | HARP · NQ Open (cross-dataset) | AUROC | No clear effect |
| intervene on reasoning components at deep layers (22, 23) (ablation) | Qwen2.5 hidden states · Epsilon fictitious-capital demo | AUROC | Improved |
| intervene on reasoning components at shallow layers (1, 2) (ablation) | Qwen2.5 hidden states · Epsilon fictitious-capital demo | AUROC | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- HaloScope — HaloScope: Harnessing Unlabeled LLM Generations for Hallucination Detection (NeurIPS 2024) · better on 8 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.92)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.92)
- Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback (AAAI 2025 · similarity 0.91)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.91)
- Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models (AAAI 2026 · similarity 0.91)
- Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations (CVPR 2026 · similarity 0.91)
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.
