Mirror, Mirror on the Wall: Can VLM Agents Tell Who They Are at All?

Researchers introduced a benchmark testing whether vision-language model (VLM) agents can recognize themselves in mirrors, a cognitive capability that emerges only in some animal species. Results show self-identification through reflection occurs mainly in stronger VLMs, while weaker models fail to extract self-relevant information despite viewing their reflections, revealing that language-based self-reference alone does not guarantee grounded self-understanding.

This research probes a fundamental question about embodied AI cognition: whether VLM agents possess genuine self-awareness or merely simulate it through language patterns. The mirror self-recognition test serves as a diagnostic tool to distinguish authentic perception-grounded self-identification from confabulation, prompt compliance, or learned priors. By designing controlled 3D scenarios requiring agents to infer hidden body attributes from reflections while avoiding self-other confusion, researchers isolated the cognitive mechanism underlying self-recognition.

The findings establish a meaningful hierarchy in VLM capabilities. Stronger models demonstrate behavioral evidence of self-grounding—they seek mirrors, temporally order observations, and attribute actions to themselves—while weaker models inspect mirrors mechanistically without extracting actionable self-knowledge. Crucially, the language-vision conflict experiments revealed that models generating self-referential language ("I see myself") may not possess corresponding perceptual grounding. This distinction matters because it separates genuine embodied cognition from sophisticated language mimicry.

For AI development, these results suggest self-grounding emerges as a property of scale and training rather than architectural necessity. The benchmark provides researchers with a principled evaluation framework beyond standard benchmarks, revealing failure modes in reasoning-action consistency that language-only tests would miss. This work advances understanding of how embodied agents construct self-models and validates mirror-based evaluation as a diagnostic for distinguishing causal versus superficial understanding in multimodal systems.

• →Mirror self-recognition capability emerges primarily in stronger VLMs but fails in weaker models despite visual mirror access
• →Language-based self-reference does not guarantee perceptually-grounded self-identification, as shown through language-vision conflict experiments
• →Stronger models demonstrate temporal reasoning and mirror-seeking behavior indicating causal self-grounding, while weaker models misattribute reflections
• →The benchmark isolates genuine embodied cognition from confabulation, prompt compliance, and learned priors through controlled 3D scenarios
• →Mirror-based evaluation provides a diagnostic framework for assessing whether AI agents ground self-understanding in perception rather than language patterns