Saved in:
Bibliographic Details
Main Authors: Hong, Rui, Quan, Shuxue
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.18373
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914621282058240
author Hong, Rui
Quan, Shuxue
author_facet Hong, Rui
Quan, Shuxue
contents When VLMs answer correctly, do they genuinely rely on visual information? We introduce a Tri-Layer Diagnostic Framework with three per-sample metrics: Latent Anomaly Detection, Visual Necessity Score, and Competition Score, which disentangle perception, dependency, and alignment failures. Across 9 VLMs and 9,000 model-sample pairs under counterfactual blind, noise, and conflict interventions, 72.9% of samples exhibit Visual Sycophancy, a Split Beliefs pattern in which internal evidence is preserved yet a hallucinated answer is decoded, while zero samples show Robust Refusal, indicating that current alignment training has eliminated refusal as a decoding outcome. Scaling within the Qwen-VL family, both within- and across-generation, monotonically reduces Language Shortcuts but amplifies Visual Sycophancy, showing that scale and newer post-training alone cannot resolve the grounding problem. Diagnostic scores further enable a training-free selective-prediction strategy yielding up to +9.5 percentage points accuracy at 50% coverage.
format Preprint
id arxiv_https___arxiv_org_abs_2603_18373
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle To See or To Please: Uncovering Visual Sycophancy and Split Beliefs in VLMs
Hong, Rui
Quan, Shuxue
Computer Vision and Pattern Recognition
Artificial Intelligence
When VLMs answer correctly, do they genuinely rely on visual information? We introduce a Tri-Layer Diagnostic Framework with three per-sample metrics: Latent Anomaly Detection, Visual Necessity Score, and Competition Score, which disentangle perception, dependency, and alignment failures. Across 9 VLMs and 9,000 model-sample pairs under counterfactual blind, noise, and conflict interventions, 72.9% of samples exhibit Visual Sycophancy, a Split Beliefs pattern in which internal evidence is preserved yet a hallucinated answer is decoded, while zero samples show Robust Refusal, indicating that current alignment training has eliminated refusal as a decoding outcome. Scaling within the Qwen-VL family, both within- and across-generation, monotonically reduces Language Shortcuts but amplifies Visual Sycophancy, showing that scale and newer post-training alone cannot resolve the grounding problem. Diagnostic scores further enable a training-free selective-prediction strategy yielding up to +9.5 percentage points accuracy at 50% coverage.
title To See or To Please: Uncovering Visual Sycophancy and Split Beliefs in VLMs
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2603.18373