COSMIC: Benchmark for Language-Mediated Spatial Integration Across Partial Views

COSMIC is a benchmark dataset created by mair-lab to test whether multimodal language models can transform local, viewpoint-dependent observations into shared spatial models through language. It places two static agents in the same indoor scene from different egocentric viewpoints, requiring them to communicate exclusively through natural language to jointly solve a spatial question-answering task. The dataset was last updated on 2026-03-31.

Use Cases

• Benchmarking spatial reasoning capabilities of multimodal language models based on the described agent communication task.
• Training models for language-mediated scene understanding based on the integration of partial egocentric views.
• Evaluating collaborative AI agents on joint problem-solving tasks based on the described two-agent communication setup.
• Researching human-AI communication for spatial tasks based on the natural language interaction paradigm.

Strengths

• Dataset is specifically designed for a novel benchmark task evaluating language-mediated spatial integration.
• Data structure is optimized for efficient access, indicated by the platform tag 'OPTIMIZED-PARQUET'.
• Last update timestamp of 2026-03-31 04:56:59 is provided.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and sample data are unknown, which may limit suitability assessment.
• License information is unknown, which could restrict usage.

Provenance

Source

mair-lab on Hugging Face

Collection Method

Likely constructed for a specific AI benchmark, as described.

Time Range

null

Freshness

Last updated 2026-03-31 04:56:59; freshness should be verified.

Geography

null