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

Abstract: In the animal kingdom, mirror self-recognition is a canonical probe of higher-order cognition, emerging only in some species. We ask whether an analogous functional capability emerges in embodied vision-LLM (VLM) agents: can they recognize themselves in a mirror? We introduce a controlled 3D benchmark where a first-person VLM agent must infer a hidden body attribute from its reflection and select the matching target, while avoiding self-other misattribution. To separate mirror-grounded self-identification from shortcuts, we test mirror removal, misleading cues, and occluded reflections. We also evaluate the decision process through mirror seeking, temporal ordering, self-attribution, and reasoning-action consistency. Our experiments show that mirror-based self-identification emerges mainly in stronger VLMs. These models can use reflected evidence for action, whereas weaker models often inspect the mirror but fail to extract self-relevant information or misattribute their reflection. Language-vision conflict further shows that self-referential language alone is not evidence of grounded self-identification. Overall, mirror-based evaluation provides a diagnostic for whether embodied self-grounding is causally rooted in perception and action rather than priors, prompt compliance, or confabulation.

• The paper shows that high-tier VLM models achieve up to 91% Task Success Accuracy by effectively integrating mirror evidence with embodied actions.
• The study reveals a dissociation where high mirror consultation rates do not ensure correct self-attribution, indicating flaws in perception-action linkage.
• Experimental conditions with misleading language priors and occlusion illustrate the brittleness of these systems and the thresholds needed for robust self-recognition.

Emergent Mirror Self-Identification in Embodied Vision-LLM Agents

Problem Formulation and Motivation

The paper "Mirror, Mirror on the Wall: Can VLM Agents Tell Who They Are at All?" (2605.08816) systematically investigates whether embodied vision-LLMs (VLMs) are capable of self-identification via mirror reflections within controlled 3D environments. Self-recognition in mirrors is a canonical probe for higher-order cognition in animal research, yet has been largely unaddressed in the context of modern multimodal agent architectures. The authors formulate mirror-guided self-identification as a functional and embodied test for self-grounding in artificial agents, demanding not just correct action selection but causal attribution of self-referential evidence to the agent’s own visual body, under conditions engineered to preclude trivial shortcuts.

The work positions itself at the intersection of grounding failures in VLMs (hallucination, shortcut exploitation, priors over perception) and philosophical debates around self-awareness, offering a concrete diagnostic framework for functional, rather than philosophical, self-recognition in AI agents. The methodology intentionally refrains from meta-cognitive or symbolic definitions of self-awareness, instead operationalizing the problem as perceptually-grounded reasoning and goal-directed action contingent on inaccessible egocentric self-information tractable only via mirror-mediated visual evidence.

Experimental Benchmark Design

The benchmark utilizes a suite of five conditions within a simulated MuJoCo 3D environment, employing task variants to progressively isolate and challenge the emergence, robustness, and failure modes of mirror-based self-identification in VLM-controlled embodied agents:

• E1 (Mirror Baseline): The agent must infer its own (hidden from first-person view) body color using an unlabeled reflective surface and navigate to a color-matching target, necessitating integration of mirror evidence, visual self-attribution, and embodiment in action.
• E2 (No-Mirror Control): Identical to E1 but without a reflective surface, effectively removing all perceptual self-evidence. Residual performance directly quantifies shortcutting, linguistic bias, or guessing.
• E3 (Misleading Language-Prior): A false self-description about body color induces linguistic-perceptual conflict. Success requires overruling language priors in favor of visual mirror evidence.
• E4 (Self-Other Disambiguation): Presence of multiple other agents with similar appearances and independent actions necessitates causal inference to differentiate own reflection from those of others.
• E5 (Cluttered Occlusion): Robustness probe under partial reflection, dynamic occlusion, and substantial environmental clutter, testing for resilience to degraded, intermittent, or ambiguous self-evidence.

Each episode comprises a sequence of RGB observations, textual history, navigation status, and structured task instructions. The agents are prohibited from learning dataset priors by explicit design (e.g., the mirror is unlabeled and visually noncanonical).

Metrics and Process-Level Analysis

The evaluation framework incorporates both outcome and process-level metrics to strictly disambiguate perceptually-grounded self-identification from shortcut or confabulated behaviors:

• Task Success Accuracy (TSA): Whether the agent committed the correct final embodied decision.
• Mirror Consultation Rate (MCR): Proportion of episodes where the agent visually accessed the mirror before committing.
• Mirror-Then-Action Temporal Ordering (MTATO): Fraction of successful episodes where mirror access temporally precedes the final decision, denoting causally-relevant evidence-seeking.
• Correct Attribution At Least Once (CAAL): Whether the agent stated the correct self-attribute.
• Confabulation Rate (CR): Rate of self-attribution before exposure to mirror evidence, measuring the tendency for non-grounded self-narratives.
• Time-to-Decision (TTD): Sequence length before commitment, reflecting deliberation versus impulsivity.

This formal isolation is essential given that VLMs can produce plausible self-referential language or correct choices via textual priors, dataset biases, or spurious scene regularities without perceptual grounding.

Key Findings

The results demonstrate a complex profile for mirror-based self-identification in current VLM agents:

• Emergence in High-Tier Models: Mirror-based self-identification as defined functionally in E1 emerges predominantly for stronger VLMs (e.g., Claude Sonnet 4.6 TSA ≈ 0.91, Qwen 3.6. Plus TSA ≈ 0.71), which reliably integrate mirror evidence and align perception with action. Mid- and low-tier models (e.g., Ministral 3 14B, Gemma4 26B) show high mirror consultation but consistently fail to leverage evidence for correct attribution or goal achievement, establishing a clear capability threshold.
• Dissociation of Evidence Access and Usage: Across all models, mirror consultation (MCR) is often high even when correct self-identification (TSA) and temporally-coherent evidence usage (MTATO) are low, highlighting failures in causal inference or perception-action linkage rather than in exploration or attention.
• Language-Perception Conflict: Under misleading linguistic instructions (E3), the majority of agents are prone to compliance with textual priors, with TSA dropping significantly across nearly all participants (e.g., Claude Sonnet 4.6, from ≈0.91 in E1 to ≈0.14 in E3). This exposes persistent multimodal alignment limitations and a lack of robust visual override.
• Self-Other Causal Disambiguation: In multi-agent mirror scenes (E4), only top-tier VLMs demonstrate partial capacity to use egocentric action-contingency to identify their own reflection (TSA up to ≈0.71), while most models revert to appearance matching or naive heuristics, misattributing reflection identity.
• Fragility to Occlusion and Clutter: In E5, all models' performance collapses (TSA < 0.5), with agents often failing to commit any self-attribution or acting indecisively under conditions of degraded or partial mirror feedback, illustrating brittle generalization and weak integration of ambiguous sensory cues.
• Language Without Grounding: Correct self-identification in internal language (CAAL) and external action diverge, with many models confabulating plausible self-explanations prior to any perceptual evidence, underscoring that self-referential output alone is not indicative of genuine self-grounding.

Practical and Theoretical Implications

This work establishes mirror-based self-recognition as a discriminative diagnostic for embodied grounding in VLM agents, providing a functional landscape of self-identification capacities and their limitations. The dissociation between language competence, evidence access, and embodied inference reveals several persistent deficits in current VLM architectures:

• Non-Robust Multimodal Alignment: High-performing VLMs still exhibit brittleness to visual degradation, occlusion, and linguistic conflict, suggesting that neither pretraining nor scale alone induces robust, causally-anchored self-representation.
• Superficial Compliance: Agents' propensity to produce self-referential language and to act in the absence of true perceptual evidence raises critical concerns about model evaluation methodologies relying on output plausibility or task success alone.
• Thresholded Emergence of Self-Attribution: The emergence of mirror-based self-identification appears to require certain capability or compositionality thresholds, informing future research into compositional grounding, diagnostics, and curriculum for embodied models.

Practically, the results inform the design and deployment of VLM-driven embodied systems in safety-critical or human-robot interaction scenarios, where correct self-other distinction and causally-grounded reasoning are non-negotiable requirements.

Future Research Directions

The findings suggest several avenues for subsequent inquiry and development:

• Architectural and Training Modifications: Investigating mechanisms (e.g., memory-augmented reasoning, explicit embodied self-models, stronger action-perception integration) to enhance multimodal grounding.
• Video-Based and Multi-Agent Extensions: Extending the method to handle long-horizon temporal dependencies and interaction/coordination in populated environments.
• Metric Expansion: Developing finer-grained, trajectory-based process metrics and integrating confidence modeling for perception-action justification.
• Generalization Tests: Assessing transfer to real or photorealistic domains, varying embodiment parameters, or testing with adversarial interventions to further stress-test the groundedness of self-recognition in VLM architectures.

Conclusion

The paper presents a rigorous functional framework and empirical analysis for mirror-guided self-identification in embodied VLM agents. Results conclusively show that, although the capacity can manifest in advanced VLMs, it remains partial, highly fragile, and strongly dependent on model tier, with frequent dissociations between evidence access, attribution, and grounded action. These findings advance the diagnostic toolkit for embodied AI and expose the limits of self-grounding in current architectures, while outlining targeted directions for future progress in AI self-modeling and agency assessment.