Identify, Isolate, and Purge: Mitigating Hallucinations in LVLMs via Self-Evolving Distillation DOI: XXXXXXX.XXXXXXX Conference: 33rd ACM International Conference on Multimedia; Oct 27–31, 2025; Dublin, Ireland ISBN: 978-1-4503-XXXX-X/2018/06 670

Wenhao Li 1 , Xiu Su 2 , Jingyi Wu 3 , Feng yang 4 , Yang Liu 3 , Yi Chen 5 , Shan You 6 , Chang Xu 1 , 1 University of Sydney, 2 Central South University, 3 Fudan University, 4 Southeast University, 5 HKUST, 6 Sensetime Research · 2025

Read Identify, Isolate, and Purge: Mitigating Hallucinations in LVLMs via Self-Evolving Distillation DOI: XXXXXXX.XXXXXXX Conference: 33rd ACM International Conference on Multimedia; Oct 27–31, 2025; Dublin, Ireland ISBN: 978-1-4503-XXXX-X/2018/06 670 by Wenhao Li 1 , Xiu Su 2 , Jingyi Wu 3 , Feng yang 4 , Yang Liu 3 , Yi Chen 5 , Shan You 6 , Chang Xu 1 , 1 University of Sydney, 2 Central South University, 3 Fudan University, 4 Southeast University, 5 HKUST, 6 Sensetime Research (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-04680.txt.

Description

Wenhao Li and co-authors teach vision-language models to purge their own made-up details, in a 2025 paper. Such models can describe images fluently yet invent objects that are not there, which hurts trust. Fixes that lean on outside tools or repeated checks slow every answer down.

Their method, SEED, works in three steps: find the false content inside the model, isolate and purge it, then distill the cleaned knowledge back in. A mode-seeking twist keeps training from scattering into empty output zones. A small adapter learns the cleaned habits and corrects the base model. On standard tests with two known models, the fix cut hallucinations by a wide margin without slow per-answer checks.

Catalog record

CREATORWenhao Li 1 , Xiu Su 2 , Jingyi Wu 3 , Feng yang 4 , Yang Liu 3 , Yi Chen 5 , Shan You 6 , Chang Xu 1 , 1 University of Sydney, 2 Central South University, 3 Fudan University, 4 Southeast University, 5 HKUST, 6 Sensetime Research
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-04680

Browse the full library · About the Tower of Babel