Hallucination Basins: A Dynamic Framework for Understanding and Controlling LLM Hallucinations
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866917386219683840 |
|---|---|
| author | Cherukuri, Kalyan Varshney, Lav R. |
| author_facet | Cherukuri, Kalyan Varshney, Lav R. |
| contents | Large language models (LLMs) hallucinate: they produce fluent outputs that are factually incorrect. We present a geometric dynamical systems framework in which hallucinations arise from task-dependent basin structure in latent space. Using autoregressive hidden-state trajectories across multiple open-source models and benchmarks, we find that separability is strongly task-dependent rather than universal: factoid settings can show clearer basin separation, whereas summarization and misconception-heavy settings are typically less stable and often overlap. We formalize this behavior with task-complexity and multi-basin theorems, characterize basin emergence in L-layer transformers, and show that geometry-aware steering can reduce hallucination probability without retraining. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_04743 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Hallucination Basins: A Dynamic Framework for Understanding and Controlling LLM Hallucinations Cherukuri, Kalyan Varshney, Lav R. Computation and Language Artificial Intelligence Systems and Control Large language models (LLMs) hallucinate: they produce fluent outputs that are factually incorrect. We present a geometric dynamical systems framework in which hallucinations arise from task-dependent basin structure in latent space. Using autoregressive hidden-state trajectories across multiple open-source models and benchmarks, we find that separability is strongly task-dependent rather than universal: factoid settings can show clearer basin separation, whereas summarization and misconception-heavy settings are typically less stable and often overlap. We formalize this behavior with task-complexity and multi-basin theorems, characterize basin emergence in L-layer transformers, and show that geometry-aware steering can reduce hallucination probability without retraining. |
| title | Hallucination Basins: A Dynamic Framework for Understanding and Controlling LLM Hallucinations |
| topic | Computation and Language Artificial Intelligence Systems and Control |
| url | https://arxiv.org/abs/2604.04743 |