Ensembling Multiple Hallucination Detectors Trained on VLLM Internal Representations

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Nakamizo, Yuto, Miyazato, Ryuhei, Tanabe, Hikaru, Yamakura, Ryuta, Hatanaka, Kiori
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909976166924288
author Nakamizo, Yuto
Miyazato, Ryuhei
Tanabe, Hikaru
Yamakura, Ryuta
Hatanaka, Kiori
author_facet Nakamizo, Yuto
Miyazato, Ryuhei
Tanabe, Hikaru
Yamakura, Ryuta
Hatanaka, Kiori
contents This paper presents the 5th place solution by our team, y3h2, for the Meta CRAG-MM Challenge at KDD Cup 2025. The CRAG-MM benchmark is a visual question answering (VQA) dataset focused on factual questions about images, including egocentric images. The competition was contested based on VQA accuracy, as judged by an LLM-based automatic evaluator. Since incorrect answers result in negative scores, our strategy focused on reducing hallucinations from the internal representations of the VLM. Specifically, we trained logistic regression-based hallucination detection models using both the hidden_state and the outputs of specific attention heads. We then employed an ensemble of these models. As a result, while our method sacrificed some correct answers, it significantly reduced hallucinations and allowed us to place among the top entries on the final leaderboard.
format Preprint
id arxiv_https___arxiv_org_abs_2510_14330
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Ensembling Multiple Hallucination Detectors Trained on VLLM Internal Representations
Nakamizo, Yuto
Miyazato, Ryuhei
Tanabe, Hikaru
Yamakura, Ryuta
Hatanaka, Kiori
Information Retrieval
This paper presents the 5th place solution by our team, y3h2, for the Meta CRAG-MM Challenge at KDD Cup 2025. The CRAG-MM benchmark is a visual question answering (VQA) dataset focused on factual questions about images, including egocentric images. The competition was contested based on VQA accuracy, as judged by an LLM-based automatic evaluator. Since incorrect answers result in negative scores, our strategy focused on reducing hallucinations from the internal representations of the VLM. Specifically, we trained logistic regression-based hallucination detection models using both the hidden_state and the outputs of specific attention heads. We then employed an ensemble of these models. As a result, while our method sacrificed some correct answers, it significantly reduced hallucinations and allowed us to place among the top entries on the final leaderboard.
title Ensembling Multiple Hallucination Detectors Trained on VLLM Internal Representations
topic Information Retrieval
url https://arxiv.org/abs/2510.14330