MAD Research: Multi-Agent Debate Frameworks for LLM Reasoning

Research repository exploring hybrid multi-agent debate architectures combining Society of Mind (collaborative exploration) with Pro/Con Debate (adversarial challenge) for improving LLM reasoning capabilities.

Quick Start

# Clone and setup
git clone <repo-url>
cd mad-research
pip install -r requirements.txt

# Run experiments
python -m src.experiments.run --config experiments/configs/hybrid_gsm8k.yaml

Project Structure

mad-research/
├── CLAUDE.md           # Project context for Claude Code
├── docs/
│   ├── PLAN.md         # Implementation plan
│   ├── PROGRESS.md     # Research progress
│   └── LITERATURE.md   # Paper summaries
├── src/
│   ├── agents/         # Agent implementations
│   ├── frameworks/     # MAD framework implementations
│   └── evaluation/     # Metrics and benchmarks
├── experiments/        # Experiment configs and results
└── tests/              # Unit and integration tests

Key Findings from Literature

Finding	Source
Model diversity > debate depth	Hegazy 2024
Sparse topology saves 40-50% cost	ACL 2024
2-3 rounds optimal	MachineSoM
Anonymization prevents sycophancy	arXiv 2025

Architecture

Phase 1: Diverse SoM Exploration (heterogeneous models, 1 round)
    ↓
Phase 2: Difficulty-Aware Gating (skip challenge if confident)
    ↓
Phase 3: Structured Pro/Con Challenge (evidence-based)
    ↓
Phase 4: Task-Appropriate Aggregation (vote/consensus)

License

MIT