mosaic

Overview

Provides a unified platform to compare and evaluate homogeneous and heterogeneous multi-agent decision-makers across RL, LLM, VLM, and human paradigms. Runs controlled cross-paradigm experiments and aggregates metrics so researchers can directly compare policies, language agents, vision-language models, and humans on shared tasks. Distinguishing features include a common agent-interface, instrumentation for cross-agent metrics, and support for mixed-modality evaluation.

Why It Matters

Target Use Cases

Researchers and engineering teams running comparative experiments across reinforcement learners, language/vision agents, and human agents to build reproducible agent benchmarks and track records.

Applications

• Compare RL policies, LLM agents, and human decisions on the same task to identify modality-specific failure modes
• Run reproducible multi-agent benchmarks that aggregate cross-paradigm metrics and produce comparable leaderboards
• Instrument agent interactions to build per-agent track records and longitudinal evaluation data