Did I See It or Think It? Probing Reality Monitoring in Vision-Language Models with Concept Injection
cs.CV, cs.AI
Submitted: 2026-11-06
Updated: 2026-11-06
Code: https://github.com/tinakhezresmaeilzadeh/vlmreality-monitoring
Terminology
Sources
- Qwen3-VL Technical Report
- Gemma 3 Technical Report
- Detecting the Disturbance: A Nuanced View of Introspective Abilities in LLMs
- Introspection Fine-Tuning (IFT): Training Small LLMs to Introspect
- Language Models (Mostly) Know What They Know
- Emergent Introspection in AI is Content-Agnostic
- Emergent Introspective Awareness in Large Language Models
- gpt-oss-120b & gpt-oss-20b Model Card
- Latent Introspection: Models Can Detect Prior Concept Injections
- Reality Monitoring in Large Language Models: Self-Knowledge That Transforms with Conversation Memory
- Can LLMs Introspect? A Reality Check
- Interventional Training Reduces Causal Bypass in In-Context Introspection
- Steering Language Models With Activation Engineering
- InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency
- Representations of Text and Images Align From Layer One
- Representation Engineering: A Top-Down Approach to AI Transparency
- A mechanistic study of language model introspection
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