Debate & Multi-Agent
Overview¶
Multi-agent debate is a reasoning technique in which two or more LLM instances independently generate answers and then iteratively challenge each other's reasoning. A judge model (or majority vote) selects the final answer. The technique improves factuality and reduces hallucinations by forcing each model to defend its position against adversarial critique.
How It Works¶
Round 1: Agent A answers → Agent B answers
Round N: Agent A critiques B → Agent B critiques A
Final: Judge (or majority) selects winner
Key design choices:
- Number of agents: Typically 2–4; more agents improve quality but increase cost
- Number of rounds: 2–3 rounds typically capture most quality gain
- Judge: A separate model, or majority vote over the agents
Why Debate Works¶
- Models are stronger critics than generators — identifying flaws in others' reasoning is easier than avoiding all flaws in one's own
- Adversarial pressure forces models to surface hidden assumptions
- Consensus under critique is more reliable than single-model confidence
Comparison with Self-Critique¶
| Aspect | Self-Critique (Self-Refine) | Multi-Agent Debate |
|---|---|---|
| Agents | One model critiques itself | Multiple independent models |
| Bias | May reinforce own errors | Adversarial pressure breaks echo chambers |
| Cost | Moderate (2–3× calls) | High (N agents × R rounds) |
| Best for | Single-model refinement | Factual QA, complex reasoning |
Limitations¶
- Cost: Scales as O(agents × rounds)
- Convergence risk: Agents can reach false consensus if they share the same systematic bias
- Diminishing returns: Most gain from round 1; additional rounds add less
Relation to Alignment¶
Debate was proposed as an alignment mechanism by Irving et al. (2018): if two AI agents debate with a human judge, a weaker human can identify correct answers even for tasks beyond their own expertise, because the losing agent is incentivized to expose flaws. This connects debate to scalable oversight — a core open problem in AI safety.