To See or To Please: Uncovering Visual Sycophancy and Split Beliefs in VLMs
cs.CV, cs.AI
Submitted: 2026-03-19
Updated: 2026-11-03
Code: https://github.com/hongrui16/ToSeeorToPlease
Terminology
Sources
- Pixtral 12B
- Discovering Latent Knowledge in Language Models Without Supervision
- The Llama 3 Herd of Models
- Are Large Vision Language Models Truly Grounded in Medical Images? Evidence from Italian Clinical Visual Question Answering
- HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models
- Grounding or Guessing? Visual Signals for Detecting Hallucinations in Sign Language Translation
- Have the VLMs Lost Confidence? A Study of Sycophancy in VLMs
- Evaluating Object Hallucination in Large Vision-Language Models
- Investigating VLM Hallucination from a Cognitive Psychology Perspective: A First Step Toward Interpretation with Intriguing Observations
- MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts
- Towards Understanding Sycophancy in Language Models
- Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution
- Simple synthetic data reduces sycophancy in large language models
- Causal-HalBench: Uncovering LVLMs Object Hallucinations Through Causal Intervention
- MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI
- When and why vision-language models behave like bags-of-words, and what to do about it?
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs