MirrorBench: Evaluating Self-centric Intelligence in MLLMs by Introducing a Mirror

Shengyu Guo1,2, Tongrui Ye2, Jianbo Zhang2, Zicheng Zhang2, Chunyi Li2, Guangtao Zhai2
1 Peking University, 2 Shanghai AI Lab
Paper(TBD) Code(TBD) Data
A-Bench Background

Existing embodied benchmarks focus on external objectives and largely neglect self-centric intelligence. MirrorBench bridges this gap by introducing a mirror-based setup and a tiered evaluation protocol assessing MLLMs from visual perception to self-representation.

Radar Charts

Overview

A-Bench Overview

The evaluation features a Self-referential Inference Loop, in which an MLLM-powered embodied agent interacts with a simulated environment. The environment is built from an asset pool comprising multiple body, hand, and mark configurations. A four-level evaluation protocol is introduced to systematically assess mirror-related capabilities: prompts for higher levels provide progressively less prior knowledge, increasing cognitive demand while keeping the physical environment identical across trials. Example scenes for Levels 0–3 (right) illustrate the rising complexity toward higher-order intelligence.

Benchmark Results

Higher scores (↑) indicate better performance. Metrics: T = TSR, S = SIR, F = FCR, P = PCR, A = Average. Text color: highest, second highest. Background color: proprietary, open-source.