If five AI agents all vote for the same wrong answer, have they collaborated? Not really.
Much of today’s “multi-agent AI” is built around agents exchanging messages, debating for a few rounds, and then voting. That can improve answer selection. But it misses the harder reason we need collaboration in the first place.
Real collaboration should help us discover what one agent did not know to look for. A doctor may notice a symptom another overlooked. An engineer may question an assumption everyone else accepted. A rescue planner may realize that the map is current but the water-level reading is stale.
This distinction matters because a NeurIPS 2025 comparison found that majority voting often beat the multi-agent debate protocols it tested. That is useful evidence for selecting among fixed answers. But it is not a verdict on the harder problem of collaborative intelligence: discovering an omitted hazard, challenging a shared assumption, finding missing evidence, or realizing that the original question was framed incorrectly.[10]
This is the central argument of this essay: collaborative intelligence is not the same as agreement. It is the process of finding what is missing, examining why a conclusion should be believed, and knowing when the group still does not know enough.
The ancient lesson we risk forgetting
Socrates, as portrayed in Plato's dialogues, did not ask a circle of people to pick the most popular assertion. He questioned definitions, tested assumptions, pursued contradictions, and asked his interlocutors to defend the reasons for what they claimed to know. The Socratic elenchus is not the exchange of talking points: it is inquiry capable of changing the question itself.
Plato's Cave, in Book VII of The Republic, warns against accepting appearances without investigating what produces them. Aristotle subsequently made causes, demonstrations, and first principles central to his account of scientific understanding: knowing a conclusion is different from understanding why it follows. These traditions helped shape the foundations of philosophical inquiry, logic, education, and scientific argument in Western civilization.[1]
These thinkers did not teach that every disagreement leads to truth. Their deeper lesson is that the quality of judgment depends on examining reasons, not merely tallying conclusions. Knowledge advances when claims can be questioned, challenged, corrected, and tested against experience. A society that stops challenging its assumptions can become confidently wrong. An AI system can, too.
Collaboration begins where the answer sheet ends
That distinction is not abstract. Even a mundane dinner order can require people to uncover different needs; a rescue decision in a storm makes the same requirement urgent.
A dinner order
The group needs the right constraints and preferences: food allergies, budget, availability, culture, and what each person is comfortable eating. A majority vote cannot authorize a dish that is dangerous for one guest.
A rescue boat in a storm
A dispatcher needs the water level now, not yesterday's report; an accessible route, a capable boat, permission to deploy, and a plan for what happens if conditions change while the boat is moving.
The second situation has the same structure as the first, but the consequences are far greater. A rescue planner might correctly predict where floodwater will be in thirty minutes and still make the wrong decision because its last gauge reading is stale, a road has collapsed, or the boat is already committed elsewhere.
Health care, long-term investment, military operations, and physical robots have an additional complication: the decision is not a one-shot conversation. Its premises evolve. New evidence arrives. Yesterday's accepted plan may be invalid today. A useful agent must remember what was decided, why it was decided, what has changed, and which actions have already become irreversible.
What an agentic system must support
I see four minimum requirements for serious collaborative agency. They are architectural requirements, not a proposal to make every household question as complicated as an air-traffic control system.
These requirements are connected. Memory without reasoning preserves mistakes. Reasoning without grounding rationalizes an obsolete picture of the world. Grounding without a record of commitments cannot support accountable action. Debate without behavioral and information control can simply repeat itself until everyone agrees—or exhausts the budget.
From Socrates to multi-LLM deliberation: the research lineage
Artificial-intelligence research did not invent the value of debating ideas. It gave us programmable participants and an opportunity to test which forms of disagreement actually improve reasoning. Here is the lineage I consider especially important in the LLM era. Parallel research also contributed to this field; these are milestones, not a claim that all later work descended from a single paper.
Published at IEEE CCWC in March 2023, my paper Prompting Large Language Models With the Socratic Method (CRIT) organized definition, elenchus, dialectic, counterfactual reasoning, and systematic questioning as practical methods for LLM prompting. Its central concern was whether a response could be examined and justified, not merely accepted as fluent output.[2]
Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch showed that multiple model instances could propose, critique, and revise answers over repeated rounds. Their relatively simple method demonstrated measurable benefits on several reasoning and factuality tasks and brought the idea to a broad ML audience.[3]
Our SocraSynth work made the moderator and the quality of deliberation central. It introduced controlled contentiousness: start adversarially to broaden exploration, use Socratic scrutiny to examine arguments, and then reduce contentiousness to support synthesis. The distinction between knowledge generation and reasoning evaluation is essential.[4]
EVINCE proposed using conditional statistics and information-theoretic signals to ask whether further interaction is useful. Its original natural setting involves common outcome labels; our current extension examines how to construct measurable, validated argument states when the debate is open-ended.[5]
Multi-LLM Agent Collaborative Intelligence: The Path to AGI brought SocraSynth, EVINCE, behavioral and governance mechanisms, and an expanding agentic architecture into a book-length research program. Author-issued editions appeared in 2024, followed by an expanded ACM Books edition in December 2025. This is the architectural specification on which the MACI implementation builds.[11]
At ICML 2025, my paper A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment developed a three-branch architecture: LLM-based knowledge generation (executive), DIKE's ethical guardrails (legislative), and ERIS's contextual interpretation (judicial). It also explored regulating linguistic behavior through emotional conditioning. Collaboration requires independent judgment and oversight, not merely information exchange.[6]
At VLDB 2025, our SagaLLM introduced persistent context management, independent validation, and compensating recovery for long-running multi-agent planning workflows. This is why serious collaboration needs more than an agent's shrinking context window: earlier commitments, dependencies, and changes must remain available and accountable.[7]
In parallel, MACNET (ICLR 2025) investigated graph-structured collaboration and scaling, while Puppeteer (NeurIPS 2025) developed adaptive agent orchestration. These contributions address complementary questions—who communicates and who acts next—alongside the reasoning, governance, and persistence required for reliable collaboration.[8]
Our Epistemic Regret Minimization work with Longling Geng formalizes a central concern: a model may obtain a correct outcome by invalid causal reasoning. ERM inspects causal warrants and offers targeted correction, including a label-free mode in which neither the answer key nor the true causal graph is available to the critic. The aim is to prevent repeated reward for the wrong reason.[13]
Our Diagnosing and Mitigating Sycophancy and Skepticism in LLM Causal Judgment makes this Socratic principle concrete. Its RCA judge cannot see gold labels; instead, it checks whether the conclusion follows from the given derivation and resists pressure-induced departures from the reasoning. This separates auditing the why of an answer from simply checking whether its final label matches a key.[12]
System-2 Reasoning: From Semantic Anchoring to Causal Intelligence brings the program's reasoning mechanisms into a unified architectural treatment, from semantic anchoring and causal audits to epistemic repair and metacognitive judgment. It provides the second major part of the functional specification for MACI, complementing Volume I's collaborative-intelligence architecture.[11]
The ICML 2026 Position Track paper AGI Requires a Coordination Layer on Top of Pattern Repositories states the architectural thesis explicitly: capable models require a System-2 coordination layer that recruits knowledge, audits reasoning, preserves state, and governs convergence. MACI is the open-source integration of that vision and the full-stack program developed in Volume I and subsequently Volume II.[9][11]
The NeurIPS voting result—and its blind spot
In 2025, a widely discussed NeurIPS paper—Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models?, by Hyeong Kyu Choi, Xiaojin Zhu, and Sharon Li—compared multi-agent debate with the simpler alternative of collecting independent answers and taking a majority vote.[10]
The result deserves to be read rather than dismissed. On seven benchmark settings, majority voting often matched or outperformed the tested debate protocols. For example, in one reported arithmetic setting with Qwen2.5-7B-Instruct, majority voting achieved 99% accuracy while two-round decentralized debate achieved 76%. Adding debate rounds did not reliably help.
What the experiment measures well
99% vs. 76%For that arithmetic benchmark, voting was substantially better at selecting a known correct answer. That is useful evidence for choosing a low-cost method on fixed-answer tasks.
What it does not decide
What was missing?Whether agents discovered a previously unnoticed constraint, corrected a causal assumption, updated stale evidence, or prevented a real-world mistake.
The paper's main evaluation concerns discrete, externally graded answers—including mathematics, multiple-choice knowledge tasks, inference, and alignment labels. It also includes a limited open-ended summarization experiment evaluated by ROUGE. That is worth acknowledging. But summarization scores are not a test of agents discovering an unlisted hazard in a flood, an unknown drug interaction, or a hidden obligation in a financial plan.
Its theoretical result is likewise conditional: the authors model debate through a particular stochastic belief-update process and derive a martingale property under that model. It is not a theorem that evidence retrieval, causal audit, persistent memory, or Socratic intervention can never improve collective reasoning. Indeed, the authors' own targeted interventions improved results by favoring corrective signals.
Multiple-choice questions are well suited to testing answer selection precisely because the choices, scoring rule, and task frame are supplied. Those same features make them an especially poor primary test of deliberation whose purpose is to discover what the problem statement omitted. This is why I resist the emerging habit of using vote-versus-debate results to suggest that genuine collaboration is dispensable.
Of course, not every multi-agent conversation is productive. Poorly structured debate can amplify errors, corrupt a correct answer, or waste compute. These are reasons to improve its regulation, not to identify collaboration with its weakest implementation.
A better test: did the agents discover what mattered?
Consider a flood-response benchmark. Do not supply every possible hazard as a multiple-choice option. Instead, supply a partial situation report. Withhold a material fact—a water-level spike, bridge closure, or change in boat availability—and allow the agents to ask questions, propose alternatives, and revise a deployment plan.
A successful system should not be rewarded merely for landing on the right dispatch label. We should ask whether it recognized the missing variable; identified which assumption needed checking; obtained appropriately fresh evidence; distinguished justified from persuasive objections; and avoided an unsafe commitment. If the evidence is inadequate, asking for another measurement can be the most intelligent answer.
| Measure | Fixed-answer evaluation | Open-problem collaboration evaluation |
|---|---|---|
| Alternatives | Was the correct answer selected? | Were relevant, previously unlisted alternatives discovered? |
| Reasoning | Did the final answer match a key? | Did arguments survive evidence and causal scrutiny? |
| Information | Did answer accuracy improve? | Did agents find missing, discriminating evidence? |
| Time | Did a model answer this question? | Did the decision remain justified when the world changed? |
| Outcome | Did the vote succeed? | Was a safe and appropriate plan executed, deferred, or revised? |
We should still compare against strong single-agent reasoning, self-consistency, majority voting, MACNET, and adaptive orchestrators under similar costs. Collaboration should earn its complexity. But it must be judged on the work it is actually designed to do.
MACI: toward collaboration as a disciplined process
The Multi-LLM Agent Collaborative Intelligence (MACI) program is the integration-and-implementation path for the full stack of research developed in Volumes I and II of The Path to AGI. The goal is to make collaboration an accountable process, rather than a chat room filled with agents.
The initial implementation need not claim to have solved every long-horizon planning problem. Its central task is to demonstrate an end-to-end process that can discover factors, refute and improve arguments, avoid premature consensus, and preserve the reasons for its conclusions. When the system later acts in health care, finance, or the physical world, stronger grounding, explicit authorization, and transaction-aware workflow controls become essential. Those broader requirements are developed in Volume III, Beyond Intelligence: From Operational AGI to Wisdom.[11]
The deeper lesson is not that more agents automatically produce more intelligence. It is that different perspectives become valuable when a system can expose their assumptions, assess their evidence, preserve their contributions, and govern what follows from them.
The future of agentic AI should not be measured by how many agents can talk, or how quickly they agree.
It should be measured by what they discover together, how well they reason, whether they know what is happening in the world now, and whether they remember enough to act responsibly tomorrow.
A vote may settle an answer. Collaboration can discover that we have been asking the wrong question.
References and reading
- Classical foundations of inquiry. Stanford Encyclopedia of Philosophy, “Socrates”; Plato, Republic, Book VII (the Cave); and Stanford Encyclopedia of Philosophy, “Aristotle’s Logic”, on the Posterior Analytics, causes, and first principles. The blog paraphrases these traditions; it does not attribute invented quotations to the philosophers.
- Chang, Edward Y. (2023). Prompting Large Language Models With the Socratic Method. IEEE CCWC. DOI: 10.1109/CCWC57344.2023.10099179.
- Du, Yilun, et al. (2024). Improving Factuality and Reasoning in Language Models through Multiagent Debate. ICML, PMLR 235:11733–11763. First posted May 2023.
- Chang, Edward Y. (2023–2024). Examining GPT-4's Capabilities and Enhancement with SocraSynth. IEEE CSCI 2023; and SocraSynth: Multi-LLM Reasoning with Conditional Statistics, arXiv January 2024.
- Chang, Edward Y. (2024). EVINCE: Optimizing Multi-LLM Dialogues Using Conditional Statistics and Information Theory. arXiv:2408.14575.
- Chang, Edward Y. (2025). A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment. ICML, PMLR 267:7332–7350.
- Chang, Edward Y., & Geng, Longling (2025). SagaLLM: Context Management, Validation, and Transaction Guarantees for Multi-Agent LLM Planning. PVLDB 18(12):4874–4886.
- Qian, Chen, et al. (2025). Scaling Large Language Model-based Multi-Agent Collaboration. ICLR 2025. Dang, Yufan, et al. (2025). Multi-Agent Collaboration via Evolving Orchestration. NeurIPS 2025.
- Chang, Edward Y. (2026). Position: AGI Requires a Coordination Layer on Top of Pattern Repositories. ICML Position Track, PMLR 306:169727–169741.
- Choi, Hyeong Kyu, Zhu, Xiaojin, & Li, Sharon (2025). Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? NeurIPS 2025.
- Chang, Edward Y. (2025–2026). Multi-LLM Agent Collaborative Intelligence (Vol. I), ACM Books, December 2025; System-2 Reasoning: From Semantic Anchoring to Causal Intelligence (Vol. II), ACM Books, July 23, 2026, DOI: 10.1145/3822380; Beyond Intelligence: From Operational AGI to Wisdom (Vol. III), Zenodo.
- Chang, Edward Y. (2026). Diagnosing and Mitigating Sycophancy and Skepticism in LLM Causal Judgment. Findings of ACL 2026, 8769–8789. DOI: 10.18653/v1/2026.findings-acl.427.
- Chang, Edward Y., & Geng, Longling (2026). Epistemic Regret Minimization: Label-Free Causal Critique Beyond Outcome Reward. arXiv:2602.11675, preprint.