Counterfactual Segmentation Reasoning: Diagnosing and Mitigating Pixel-Grounding Hallucination

2025/06/26 by Xinzhuo Li, Li, Xinzhuo, Adheesh Juvekar +11
Neuroscience · #Artificial Intelligence (cs.AI) #Computation and Language (cs.CL) #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Hallucinations in medical conditions #Machine Learning (cs.LG)

paper · pdf · doi:10.48550/arxiv.2506.21546

openalex publication_date 2025/06/26 · openalex created_date 2025/10/15 · openalex updated_date 2026/07/28

Abstract

Segmentation Vision-Language Models (VLMs) have significantly advanced grounded visual understanding, yet they remain prone to pixel-grounding hallucinations, producing masks for incorrect objects or for objects that are entirely absent. Existing evaluations rely almost entirely on text- or label-based perturbations, which check only whether the predicted mask matches the queried label. Such evaluations overlook the spatial footprint and severity of hallucination and therefore fail to reveal vision-driven hallucinations, which are more challenging and more prevalent. To address this gap, we formalize the task of Counterfactual Segmentation Reasoning (CSR), where a model must segment the referenced object in the factual image and abstain in its counterfactual counterpart. To support this task, we curate HalluSegBench, the first large-scale benchmark to diagnose referring and reasoning expression segmentation hallucinations using controlled visual counterfactuals, alongside new evaluation metrics that measure hallucination severity and disentangle vision- and language-driven failure modes. We further introduce RobustSeg, a segmentation VLM trained with counterfactual fine-tuning (CFT) to learn when to segment and when to abstain. Experimental results confirm RobustSeg reduces hallucinations by 30%, while improving segmentation performance on FP-RefCOCO(+/g).

Citations

• MedSAM3: Delving into Segment Anything with Medical Concepts
• Visual Jenga: Discovering Object Dependencies via Counterfactual Inpainting
• Seg-Zero: Reasoning-Chain Guided Segmentation via Cognitive Reinforcement
• CALICO: Part-Focused Semantic Co-Segmentation with Large Vision-Language Models
• PRIMA: Multi-Image Vision-Language Models for Reasoning Segmentation
• InstructSeg: Unifying Instructed Visual Segmentation with Multi-modal Large Language Models
• HyperSeg: Towards Universal Visual Segmentation with Large Language Model
• Rethinking Visual Counterfactual Explanations Through Region Constraint
• FineCops-Ref: A new Dataset and Task for Fine-Grained Compositional Referring Expression Comprehension
• Guiding Vision-Language Model Selection for Visual Question-Answering Across Tasks, Domains, and Knowledge Types
• OMG-LLaVA: Bridging Image-level, Object-level, Pixel-level Reasoning and Understanding
• Does Object Grounding Really Reduce Hallucination of Large Vision-Language Models?
• Groma: Localized Visual Tokenization for Grounding Multimodal Large Language Models
• PhD: A ChatGPT-Prompted Visual hallucination Evaluation Dataset
• Doubly Abductive Counterfactual Inference for Text-based Image Editing
• Benchmarking Segmentation Models with Mask-Preserved Attribute Editing
• Visual Hallucinations of Multi-modal Large Language Models
• CounterCurate: Enhancing Physical and Semantic Visio-Linguistic Compositional Reasoning via Counterfactual Examples
• Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks
• SpatialVLM: Endowing Vision-Language Models with Spatial Reasoning Capabilities
• ZONE: Zero-Shot Instruction-Guided Local Editing
• GSVA: Generalized Segmentation via Multimodal Large Language Models
• See, Say, and Segment: Teaching LMMs to Overcome False Premises
• SmartEdit: Exploring Complex Instruction-based Image Editing with Multimodal Large Language Models
• PixelLM: Pixel Reasoning with Large Multimodal Model
• GLaMM: Pixel Grounding Large Multimodal Model
• Negative Object Presence Evaluation (NOPE) to Measure Object Hallucination in Vision-Language Models
• HallE-Control: Controlling Object Hallucination in Large Multimodal Models
• Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models
• 🧜Siren’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models
• LISA: Reasoning Segmentation via Large Language Model
• MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing
• GRES: Generalized Referring Expression Segmentation
• InstructEdit: Improving Automatic Masks for Diffusion-based Image Editing With User Instructions
• Evaluating Object Hallucination in Large Vision-Language Models
• Visual Instruction Tuning
• Micrograph segmentations for DDEVD
• Segment Anything
• BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
• InstructPix2Pix: Learning to Follow Image Editing Instructions
• Plausible May Not Be Faithful: Probing Object Hallucination in Vision-Language Pre-training
• Towards Robust Referring Image Segmentation
• Flamingo: a Visual Language Model for Few-Shot Learning
• Counterfactual Cycle-Consistent Learning for Instruction Following and Generation in Vision-Language Navigation
• OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework
• GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models
• LoRA: Low-Rank Adaptation of Large Language Models
• Counterfactual VQA: A Cause-Effect Look at Language Bias
• UNITER: UNiversal Image-TExt Representation Learning
• A Style-Based Generator Architecture for Generative Adversarial Networks
• Object Hallucination in Image Captioning
• Modeling Context in Referring Expressions
• Features of similarity.
• VisionReasoner: Unified Reasoning-Integrated Visual Perception via Reinforcement Learning

Cited by

• ChronoVision: Temporal Reasoning via Latent State Reconstruction
• Consistency Has a Computable Blind Spot: A Commutation Theory of Label-Free Reliability for Vision-Language Figure Reading

Related

• Treble Counterfactual VLMs: A Causal Approach to Hallucination
• Antidote: A Unified Framework for Mitigating LVLM Hallucinations in Counterfactual Presupposition and Object Perception
• CF-VLM:CounterFactual Vision-Language Fine-tuning
• VaLiD: Mitigating the Hallucination of Large Vision Language Models by Visual Layer Fusion Contrastive Decoding
• Hallucinations Leave a Grounding Signature:Verifier-Guided Decoding for Selective Object Correction