How zou-group/sirius works
Project type and purpose
• From the file tree and README, this is an open-source research tool (“SiriuS”) implementing a self-improving multi-agent negotiation environment.
• It provides:
• A “NegotiationArena” core package for game mechanics and engine logic
• Plug-and-play game scenarios (ultimatum, buy-sell, trading, simple)
• Agent implementations (ChatGPT, Claude, LLaMA2, custom behaviors)
• Runner and entry-point scripts to launch matches and experiments
• Utilities for logging, parsing, prompt generation, and fine-tuning
File structure
Top-level:
• Competitive/NegotiationArena: the core Python package
– games/: per-scenario game logic and prompt definitions
– negotiationarena/: engine, agent interfaces, game objects, utilities
– runner/: experiment scripts for different setups
– buysell_main.py, trading_main.py, ultimatum_main.py: CLI entry points
• agent.py & fine_tune.py: top-level scripts for custom agents and model fine-tuning
• Problem_solving/: possibly separate experiments or examples
• Configuration: .gitignore, LICENSE
README for additional insights
– Title and arXiv link reveal academic research focus on “bootstrapped reasoning” in multi-agent systems.