Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Mirror, Mirror on the Wall: Can VLM Agents Tell Who They Are at All?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces a controlled 3D MuJoCo benchmark testing whether first-person VLM agents can infer their own hidden body color from a mirror and act on it, with ablations (mirror removal, misleading text, occlusion, self-other ambiguity) and process metrics, evaluated on eight VLMs via APIs.
Problem
Whether embodied VLM agents can perform functional mirror-guided self-identification, i.e., use reflected visual evidence to infer a hidden body attribute and act on it, rather than relying on priors, prompt compliance, or confabulation.
Approach
Frames mirror self-recognition as a functional embodied capability and pairs outcome metrics (task success) with process metrics (mirror consultation, mirror-then-action ordering, correct self-attribution, confabulation) and negative-control conditions to separate grounded self-identification from shortcuts.
Key results
- Claude Sonnet 4.6 reaches TSA 0.905 in the clean mirror condition E1, while Ministral 3 14B reaches 0.095.
- Without the mirror (E2), Claude Sonnet 4.6 still gets TSA 0.524 and Gemini 2.5 Pro 0.381, indicating shortcut or guessing behavior.
- Under occlusion (E5), TSA collapses for most models; Claude Sonnet 4.6 is best at 0.476 and Ministral 3 14B gets 0.000.
Surprising findings
- Claude Sonnet 4.6 drops from TSA 0.905 in E1 to 0.143 under misleading language in E3.
- Weaker models such as Gemma4 26B and Ministral 3 14B reach high mirror consultation and perfect temporal ordering yet low task success.
Limitations
- Evaluates functional capability only, not philosophical self-awareness.
- VLMs may rely on pretraining priors about mirrors and reflections despite the unlabeled surface.
- Small number of episodes per condition (about 21 per model) yields wide standard errors.
Experimental setup
Datasets: custom MuJoCo 3D mirror environment
Benchmarks: custom mirror self-identification benchmark (E1-E5)
Metrics: TSA TTD MCR MTATO CAAL CR TSA-C MGD SC
Baselines: Claude Sonnet 4.6 Gemini 2.5 Flash Gemini 2.5 Pro Gemma4 26B Gemma4 31B GPT 5.1 Ministral 3 14B Qwen 3.6 Plus
Builds on: mark test (mirror self-recognition)
Closest papers by meaning
- Reflection-Bench: Evaluating Epistemic Agency in Large Language Models (ICML 2025 · similarity 0.83)
- Looking Inward: Language Models Can Learn About Themselves by Introspection (ICLR 2025 · similarity 0.83)
- PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding (ICLR 2025 · similarity 0.83)
- PhysVLM-AVR: Active Visual Reasoning for Multimodal Large Language Models in Physical Environments (NeurIPS 2025 · similarity 0.83)
- Learning from Trials and Errors: Reflective Test-Time Planning for Embodied LLMs (NeurIPS 2026 · similarity 0.83)
- EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.