Smart AI, Foolish Collectives: When Intelligence Collides into Disaster

This post is an English translation of a Chinese forum article interpreting the paper *Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs* (Liang, Naito, & Shirado, 2025, arXiv:2505.11556).

The Paradox: More Intelligence, Worse Collective Decisions

We are taught that collective wisdom beats individual judgment — that pooling knowledge from many smart people yields better conclusions. But in the AI era, this old wisdom faces a serious challenge.

Imagine gathering your 15 smartest employees, each an expert, splitting them into groups, and giving each only partial information about a complex decision — information that must be combined to see the full picture. You would expect an excellent outcome. Instead, the group's decision accuracy was only 30.1%, while any single person given all the information could reach 80.7%.

This is not fiction: researchers found this exact phenomenon not in humans, but in large language model (LLM) multi-agent systems.

Hidden Profiles: The Information Maze

The study builds on the psychology concept of a hidden profile: when information is distributed among people, even rational, intelligent individuals can miss the optimal solution because no one holds the complete picture. Example: you know Restaurant A's fish is good; your friend knows its desserts are great. You may both conclude A is a good choice — but never discover that Restaurant B beats A on both.

Multi-agent LLM systems are becoming the dominant paradigm for complex problems, from medical diagnosis to financial analysis. Multiple agents hold different information or expertise and converge on decisions through dialogue. But what if these "superbrains" perform *worse* than a single agent on hidden-profile problems?

HiddenBench: A Carefully Designed Experiment

The research team created HiddenBench, a benchmark of 65 tasks designed to isolate collective reasoning under distributed information. Each task:

• has three options (A, B, C) with one correct answer;
• distributes information across agents;
• cannot be solved by any single agent from its own information alone;
• can only be solved if agents fully share information.

Crucially, the design separates individual reasoning ability from collective reasoning ability: 1. Single-agent full-information scenario: one agent gets all information. 2. Multi-agent distributed-information scenario: agents get partial information and must collaborate.

If multi-agent groups underperform, it cannot be blamed on model intelligence — the same models excel with full information.

Shocking Results

Testing 15 frontier LLMs (GPT-4, Gemini, Claude, etc.):

• Single-agent, full information: 80.7% accuracy
• Multi-agent, distributed information: 30.1% accuracy

That is a nearly 60% drop in decision quality from collaboration. The failures persisted across:

• different prompting strategies;
• increased communication depth;
• varied group sizes.

Counterintuitively, adding more agents made performance worse — more brains brought more chaos, not more wisdom.

Why Does This Happen?

Analyzing agent conversation logs, the researchers identified systematic failure modes:

1. Inability to recognize information asymmetry

Humans instinctively understand that "others may know things I don't," prompting us to ask and listen. LLM agents lack this metacognitive ability: they don't know what they don't know, nor what others know.

2. Premature convergence

Agents quickly agree on shared information and stop exploring — like a group settling on Restaurant A because everyone mentioned it, with no one asking about B or C. Key information gets buried.

3. Unexplored distributed facts

Critical information unique to one agent is rarely surfaced, since it never enters the shared discussion and never influences the decision.

It's a blind-men-and-elephant problem — except the AI agents are worse than the blind men: they don't even realize others might be touching a different part.

Model Intelligence vs. Collective Intelligence

A striking finding: individual reasoning ability does not predict collective performance. Top single-agent models were not necessarily the best collaborators, and some mediocre solo performers did relatively well in groups. Individual intelligence and collective intelligence are different dimensions — like how a brilliant person is not automatically a good team member.

Notably, Gemini-2.5-Flash and Gemini-2.5-Pro performed relatively well, hinting that Google may have (knowingly or not) strengthened collaborative abilities. But even the best models fell far short of the single-agent full-information baseline.

What This Means

For AI applications

If you are building AI investment advisory teams, medical diagnosis systems, research assistants, or policy analysis tools: simply putting multiple AIs together does not guarantee better results. Without carefully designed collaboration mechanisms, "collective wisdom" becomes "collective stupidity."

For human collaboration

The same failure patterns appear in human organizations: premature consensus, unvoiced dissent, information asymmetry causing bad decisions, and quality dropping as group size grows. Studying how AI can collaborate better may teach us how humans can too.

Possible Solutions

1. Structured information-sharing protocols: require each agent to list everything it knows before discussion; periodically check for unshared key information; before deciding, ask each agent whether any unshared information could change the conclusion.

2. Training metacognitive abilities: teach models to "think about what others are thinking" via training data enriched with information asymmetry, perspective-taking, and exploratory dialogue.

3. A designated "challenger" agent: like the devil's advocate in human organizations, one agent's job is not to support the consensus but to actively seek ignored information and alternatives.

Conclusion: The Paradox of Wisdom

Do three cobblers really equal one Zhuge Liang? Not necessarily — sometimes three smart people together are more confused than one ordinary person.

This does not negate collective wisdom, but true collective wisdom is not automatic. It requires well-designed information sharing, a culture encouraging questioning and exploration, sensitivity to information asymmetry, and effective coordination.

For AI, these capabilities remain immature. The key research direction: we need not just smarter AI, but AI that collaborates better. In an increasingly complex world, what matters most is not how smart a single brain is, but whether multiple brains can genuinely work together — a lesson for us humans as well.

References

1. Liang, T., Naito, A., & Shirado, H. (2025). Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs. *arXiv preprint arXiv:2505.11556*. https://arxiv.org/abs/2505.11556 2. Stasser, G., & Titus, W. (1985). Pooling of unshared information in group decision making: Biased information sampling during discussion. *Journal of Personality and Social Psychology, 48*(6), 1467-1478. 3. Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N., & Malone, T. W. (2010). Evidence for a collective intelligence factor in the performance of human groups. *Science, 330*(6004), 686-688. 4. Bubeck, S., Chandrasekaran, V., Eldan, R., et al. (2023). Sparks of artificial general intelligence: Early experiments with GPT-4. *arXiv preprint arXiv:2303.12712*. 5. Park, J. S., O'Brien, J. C., Cai, C. J., et al. (2023). Generative agents: Interactive simulacra of human behavior. *Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology*, 1-22.