Research map / Large language models / Retrieval and knowledge
A Geometric Analysis of Small-sized Language Model Hallucinations
Emanuele Ricco, Elia Onofri, Lorenzo Cima, Stefano Cresci, Roberto Di Pietro
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 response embeddings onto the Fisher discriminant direction | APORIA · SOCRATES-300K, average over 10 models | inter/intra-class distance ratio: 1.13 -> 7.26 | Improved |
| use APORIA-LP instead of SelfCheckGPT as detector | hallucination classification · SOCRATES-300K | accuracy (%): 82.0 -> 87.2 | Improved |
| use APORIA-LP instead of INSIDE | hallucination classification · SOCRATES-300K | F1 (%): 74.9 -> 90.7 | Improved |
| increase labelled training responses from 5 to about 30-60 (ablation) | APORIA-LP · SOCRATES-300K | F1 | Improved |
| increase labelled training responses beyond about 60 (ablation) | APORIA-LP · SOCRATES-300K | F1 | No clear effect |
| tune Fisher regularisation lambda to about 1.2 (ablation) | APORIA-LP · SOCRATES-300K, all 10 models | F1 | Improved |
| swap sentence encoder for larger 768-dim encoders (base-v2, mpnet-base-v2) (ablation) | APORIA-LP · SOCRATES-300K | LP F1 (%): 86.1 -> 87.4 | No clear effect |
| swap sentence encoder for larger encoders (ablation) | Fisher projection · SOCRATES-300K | Fisher separability: 7.21 -> 10.59 | Improved |
| increase base LLM size | APORIA-LP · SOCRATES-300K | accuracy (%): yi1.5-9B 92.5 vs qwen2.5-32B 85.5 | No clear effect |
| reuse Fisher direction across prompts or models | APORIA-LP · SOCRATES-300K | cosine similarity of Fisher directions: 0.058 across prompts; 0.145 across models | Hurt |
| replace Fisher-space classification with raw pairwise distances (ablation) | APORIA structural analysis · SOCRATES-300K | class separability: ratio 1.13 in original space | Hurt |
Closest papers by meaning
- In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation (ICML 2024 · similarity 0.92)
- Hallucination as Commitment Failure: Larger LLMs Misfire Despite Knowing the Answer (NeurIPS 2026 · similarity 0.91)
- THRONE: An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models (CVPR 2024 · similarity 0.91)
- FactTest: Factuality Testing in Large Language Models with Finite-Sample and Distribution-Free Guarantees (ICML 2025 · similarity 0.91)
- Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.91)
- ReDeEP: Detecting Hallucination in Retrieval-Augmented Generation via Mechanistic Interpretability (ICLR 2025 · 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.
