WIP Research

Belief-Dynamics

Research on false consensus attractors in multi-agent LLM debate systems

// DESCRIPTION

Investigating how groups of LLM agents can converge on incorrect beliefs through debate dynamics.

Research Questions

  • How do LLM agents reach consensus in debates?
  • What factors lead to false consensus (hallucination amplification)?
  • Can we design debate protocols that avoid these pitfalls?

Experiments

  • E1: Naive debate baseline (49% hallucination rate)
  • E2: Structured argumentation
  • E3: Devil's advocate injection
  • E4: Confidence-weighted voting
  • E5: Multi-round refinement (running)

Preliminary Findings

Naive debate shows 49% hallucination rate, indicating significant false consensus risk in multi-agent systems.

// HIGHLIGHTS

  • Novel research direction in multi-agent AI
  • E1-E4 complete, E5 running
  • ICML 2026 target

TECH_STACK

Python Ollama OpenAI API LangChain

PROJECT_INFO

started: 2024-11-01
status: WIP
type: Research