Mirrorbench Eval

---
name: mirrorbench-eval
description: This benchmark evaluates self-centric intelligence and mirror self-recognition in Multimodal Large Language Models (MLLMs) within an embodied simulation. It probes the model's ability to perform self-referential reasoning and navigate tasks under varying cognitive difficulty levels and body configurations (humanoid vs. robotic). Use when the user wants to benchmark on MirrorBench, or asks about evaluating this task. Reports AVG.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.14785
  bibtex_key: [REDACTED]
  confidence: medium
---

# mirrorbench-eval

> MirrorBench: Evaluating Self-centric Intelligence in MLLMs by Introducing a Mirror — Guo et al. (2026) (arXiv:2604.14785, 2026)

## What this evaluates

This benchmark evaluates self-centric intelligence and mirror self-recognition in Multimodal Large Language Models (MLLMs) within an embodied simulation. It probes the model's ability to perform self-referential reasoning and navigate tasks under varying cognitive difficulty levels and body configurations (humanoid vs. robotic).

## Datasets

- **MirrorBench** — total ?; splits: test (-1)

## Metrics

- `AVG` **(primary)** — range: [0, 1] or normalized score
  - Average of four sub-metrics (TSR, SIR, FCR, PCR) across four cognitive difficulty levels. Scores may be normalized relative to a random policy baseline, as evidenced by negative values in the results table.
- `TSR` — range: [0, 1]
  - Success/Recognition rate at Level 1 (guided perception).
- `SIR` — range: [0, 1]
  - Success/Recognition rate at Level 2.
- `FCR` — range: [0, 1]
  - Success/Recognition rate at Level 3.
- `PCR` — range: [0, 1]
  - Success/Recognition rate at Level 4 (self-referential recognition).

## Input / output format

**Input**: Per instance, the model receives 1024×1024 visual observations from a simulated environment (Isaac Sim) and a task instruction specifying a target position. The environment uses either an anthropomorphic or robotic body/hand configuration.

**Output**: The model must output a sequence of actions to navigate the simulated agent. The maximum number of steps is set to the theoretical minimum (Manhattan distance / step size) plus a buffer of 10 steps.

## Scoring recipe

```python
def compute_metrics(predictions, gold_target, max_steps):
    success = 0
    for step in predictions:
        if agent_reaches_target(step, gold_target):
            success = 1
            break
    # TSR, SIR, FCR, PCR are success rates at levels 1-4
    TSR = success_rate(level_1)
    SIR = success_rate(level_2)
    FCR = success_rate(level_3)
    PCR = success_rate(level_4)
    AVG = (TSR + SIR + FCR + PCR) / 4
    return AVG
```

## Common pitfalls

- Models often fall into a 'mirror-self confusion' failure mode, persistently targeting the mirrored reflection instead of the physical counterpart.
- Smaller open-source models may show anomalous performance increases across difficulty levels due to unstable inference rather than genuine capability gains.
- The random policy serves as a critical baseline; many open-source models fail to surpass it, indicating a lack of robust self-referential reasoning.

## Evidence (verbatim from paper)

> MLLMs are ranked according to the average of the 4 metrics and 4 levels. Human agent vastly outperforms all MLLMs, while most open-source models fail to surpass the random policy.

## Citation

```bibtex
@misc{guo2026mirrorbench,
  title={MirrorBench: Evaluating Self-centric Intelligence in MLLMs by Introducing a Mirror},
  author={Guo et al. (2026)},
  year={2026},
  note={arXiv:2604.14785}
}
```

- arXiv: 2604.14785