aiqbench analysis
This research addresses a critical safety vulnerability in monocular depth foundation models where they generate illusory 3D structures from ambiguous planar inputs. The authors propose a comprehensive framework to benchmark these hallucinations using new metrics and introduce Grounded Self-Distillation, a technique that surgically corrects errors without degrading general model performance. This work is essential for advancing the robustness of computer vision systems against perceptual ambiguities in real-world contexts.
The 3D Mirage: Probing and Taming 3D Hallucinations
arXiv:2512.15423v2 Announce Type: replace-cross Abstract: Monocular depth foundation models achieve remarkable generalization by learning large-scale semantic priors, but this creates a critical vulnerability: they hallucinate illusory 3D structures from planar/low-curvature but perceptually ambiguous inputs. We term this failure the 3D Mirage. This paper introduces a novel end-to-end framework to probe, score, and tame this under-quantified safety risk in monocular depth under context variation. To probe, we present 3D-Mirage, the first benchmark to combine context variation and precise annotation for real-world illusions with real object exclusions, multi-surface support; purpose-built to stress-test monocular depth on real-world illusions. To score, we propose a second-order magnitude-based evaluation with two metrics: the Deviation Composite Score (DCS) for high second-order 3D structure and the Confusion Composite Score (CCS) for contextual instability. To tame this failure, we introduce Grounded Self-Distillation, a parameter-efficient strategy on Depth-Anything-V2 baseline that surgically targets and resolves hallucination on illusion ROIs while preserving background knowledge, avoiding catastrophic forgetting. Our work provides an innovative pipeline for diagnosing and addressing this phenomenon, urging a necessary shift in the evaluation of MDE from pixel-wise accuracy to structural and contextual robustness.
Abstract:Monocular depth foundation models achieve remarkable generalization by learning large-scale semantic priors, but this creates a critical vulnerability: they hallucinate illusory 3D structures from planar/low-curvature but perceptually ambiguous inputs. We term this failure the 3D Mirage. This paper introduces a novel end-to-end framework to probe, score, and tame this under-quantified safety risk in monocular depth under context variation. To probe, we present 3D-Mirage, the first benchmark to combine context variation and precise annotation for real-world illusions with real object exclusions, multi-surface support; purpose-built to stress-test monocular depth on real-world illusions. To score, we propose a second-order magnitude-based evaluation with two metrics: the Deviation Composite Score (DCS) for high second-order 3D structure and the Confusion Composite Score (CCS) for contextual instability. To tame this failure, we introduce Grounded Self-Distillation, a parameter-efficient strategy on Depth-Anything-V2 baseline that surgically targets and resolves hallucination on illusion ROIs while preserving background knowledge, avoiding catastrophic forgetting. Our work provides an innovative pipeline for diagnosing and addressing this phenomenon, urging a necessary shift in the evaluation of MDE from pixel-wise accuracy to structural and contextual robustness.
Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO)
arXiv:2512.15423 [cs.CV]
(or arXiv:2512.15423v2 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2512.15423 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 276 words
Related coverage
We're talking past our models; or, How a model defined its "evil" vector as dread
LessWrong · Jul 20, 2026
Does the Truthfulness Signal Survive Code-Mixing? Probing Hidden States for Hallucination Detection in Hinglish
arXiv · Sep 22, 2026
Mirage Probes: How Vision Models Fake Visual Understanding
arXiv · Jun 15, 2026
Analyzing the Correlation Between Hallucinations and Knowledge Conflicts in Large Language Models
arXiv · Jun 9, 2026