LLM councils show groupthink

2024-04-23 · Source: Strange Loop Canon · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, medium

Summary

An analysis of LLM "councils," where multiple models collaborate to generate a final answer, reveals a significant tendency towards groupthink, often suppressing unique and valuable ideas. An experiment compared three council structures: direct blending, a peer-review-based LLM Council, and a "best answer" picker. The findings indicate that these council methods, particularly peer review, prioritize consensus, preserving only about a quarter of the high-value, single-model ideas identified by human judges. This phenomenon mirrors "biased sampling of shared information" observed in human group decision-making. While councils can improve the average quality of answers, they are not universally superior and require careful design to ensure the preservation of diverse and novel insights, suggesting a need for explicit protocols to gather and assess individual model contributions.

Key takeaway

For AI Scientists and Machine Learning Engineers designing multi-agent LLM systems, understand that "council" approaches can inadvertently filter out valuable, unique ideas. Your current council structure might be sacrificing innovation for consensus. To mitigate this, explicitly design protocols to capture, rank, and assess individual model contributions before final synthesis. This ensures diverse insights are preserved, even if it adds computational overhead.

Key insights

LLM councils, despite improving average answers, often suppress unique, high-value ideas due to a consensus bias.

Principles

• Model diversity doesn't guarantee idea preservation.
• Group decision-making biases apply to LLM councils.
• Explicit idea capture is crucial for diverse output.

Method

An experiment compared LLM council structures by breaking answers into "cards," clustering similar ideas, and having judges score solo-derived clusters to assess idea preservation.

In practice

• Explicitly gather and store unique ideas from each LLM.
• Rank and assess individual ideas before final synthesis.
• Experiment with council structures for specific problems.

Topics

• LLM Councils
• Multi-agent Systems
• Groupthink
• Idea Preservation
• Model Diversity
• Prompt Engineering

Code references

• karpathy/llm-council
• strangeloopcanon/ai_committee_brain

Best for: AI Architect, AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• LLM Consortium for Software Design Refinement: A Controlled Experiment on Multi-Agent Collaboration Topologies — Structural adversarial and cross-model review topologies significantly enhance multi-agent LLM software design.
• Week Ending 3.22.2026 — Advancements in LLM reasoning, multi-agent coordination, creativity, and social choice highlight diverse progress in AI capabilities and understanding.
• When Agents Disagree: The Selection Bottleneck in Multi-Agent LLM Pipelines — Selector quality, not just team diversity, dictates multi-agent LLM pipeline performance.
• Network Effects and Agreement Drift in LLM Debates — LLM agent opinion dynamics are shaped by network structure, group imbalance, and a systematic "agreement drift" bias.
• LLM Agents for Deliberative Collaboration: A Study on Joint Decision Making Under Partial Observability — LLM agents face significant challenges in deliberative collaboration under partial observability, yet the deliberation process itself can enhance performance.
• Decision Protocols in Multi-Agent Large Language Model Conversations — Multi-agent LLM decision protocols should be tailored to task type, with consensus for knowledge and voting/judge for logic.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Strange Loop Canon.