Understanding the Effects of Multi-Agent Deliberation for Human-AI Decision-Making

Harry Yizhou Tian, Ming YinarXiv 2026 · arXiv

TL;DR

This paper investigates Multi-Agent Deliberation (MAD)—where multiple LLMs discuss a problem—as a support tool for human decision-making. Through two large-scale user studies, the researchers found that while MAD improves human accuracy compared to single-round AI advice, the benefit is primarily driven by the multi-round nature of the interaction rather than the presence of multiple AIs. Surprisingly, common design choices like assigning a "devil's advocate" persona actually hindered performance and increased user frustration.

What Problem Does This Paper Address?

Most human-AI collaboration today follows a "consultancy" model: a human asks a question, and a single AI provides a one-time recommendation. However, as LLMs become more capable, researchers are exploring more complex paradigms like MAD, where a "team" of AIs debates a topic to surface diverse perspectives and catch errors.

While MAD has shown promise in improving raw AI performance on benchmarks, we don't know if it actually helps humans make better decisions. Does seeing a debate between AIs help a person calibrate their trust, or does it just add cognitive load and confusion? The authors set out to disentangle the two core dimensions of MAD—multiple agents and multi-round interaction—to see what actually moves the needle for human-AI teams.

Methods

The authors conducted two user studies ($N=401$ and $N=417$) using a $3 \times 2$ experimental design to test different MAD configurations against standard baselines.

Experimental Setup: * Baselines: They compared MAD against single-round AI assistance (Chatbot-R1), multi-round interaction with a single AI (Chatbot), and two AIs providing independent, non-interactive responses (Static). * AI Personas: They tested three configurations: Base (no persona), Expert (task-specific roles), and DA (Devil’s Advocate, where one AI is forced to argue against the other). * Human Engagement: They compared Active Participants (who chatted with the AIs) vs. Passive Reviewers (who only read the AI-to-AI transcript).

The study used two state-of-the-art models, GPT-5.4 and Claude Sonnet 4.5, to ensure diversity in AI perspectives.

Summary of experimental treatments and research questions.

Datasets

The study evaluated MAD across two distinct tasks:

  1. ReaLMistake: A response-verification task where users must detect logic or instruction-following errors in LLM outputs based on Wikipedia evidence.
  2. Fact-check: A custom-curated dataset of 48 political claims (Democrat- vs. Republican-leaning) from PolitiFact and Snopes, where AIs were supplemented with real-time web search results.

Key Insights and Ideas

  • Multi-round interaction is the MVP: The biggest accuracy gains came from the iterative nature of the discussion. Seeing the reasoning evolve over rounds helped users better utilize correct AI answers compared to seeing a static "wall of text" from two AIs.
  • The "Second AI" is context-dependent: Adding a second AI only significantly helped in the more difficult ReaLMistake task, where the single-AI baseline had more room for improvement. In the Fact-check task, the primary AI was already highly accurate, making the second AI redundant.
  • The Devil's Advocate backfire: While intended to prevent "echo chambers," the Devil's Advocate persona often persuaded the correct AI to change its mind to a wrong answer. This led to lower overall human accuracy and significantly higher user frustration.
  • Active participation is a double-edged sword: Engaging directly in the deliberation improved accuracy in the ReaLMistake task but came at the cost of increased mental demand and frustration.

Key Results

  • Overall Accuracy: In the ReaLMistake task, MAD (0.733 accuracy) significantly outperformed the single-round baseline (0.605) and the two-AI static baseline (0.620).
  • AI Performance: Deliberation actually improved the AIs' own accuracy. In ReaLMistake, the "all-AI correct rate" was significantly higher in MAD than in the static, non-interactive condition.
  • Subjective Experience: Users found MAD-based assistance more helpful and felt more comfortable during AI disagreements than they did when viewing static independent responses. However, active participation in the chat increased reported frustration levels across both tasks.
  • Shared Errors: MAD failed to help users detect "correlated errors"—cases where both AIs were wrong. In these instances, the deliberation often reinforced the incorrect consensus rather than breaking it.

Limitations and Future Work

The authors acknowledge several constraints:

  • Task Difficulty: The tasks were designed for laypeople. The dynamics of MAD might change significantly in highly specialized domains (e.g., legal or medical) where the human cannot easily verify the AI's claims.
  • Persona Engineering: The failure of the Devil's Advocate might be due to the specific prompting used. Future work could explore "untrained" vs. "fine-tuned" adversarial agents or agents that only provide counterarguments when a specific uncertainty threshold is met.
  • Reliance Metrics: The study focused on accuracy but did not explicitly measure "overreliance" (following the AI when it's wrong) vs. "underreliance" (ignoring the AI when it's right) in a way that accounts for the multi-agent consensus.

Overall, the paper suggests that while multi-agent deliberation is a powerful tool, designers should prioritize multi-round interaction and be cautious with adversarial personas that might confuse the human in the loop.

Summary updated

Abstract

Multi-agent deliberation (MAD), where multiple LLMs engage in structured discussion, improves AI performance on many benchmarks, but whether it helps humans make better decisions remains untested. We conduct two user studies on LLM-response-verification and fact-checking. We first compare MAD against common AI decision-support methods that isolate its two defining dimensions --- multiple AIs and multi-round interaction. We find MAD outperformed single-round AI assistance on both tasks; multi-round interaction was the main driver of this benefit, while a second AI added value when the single AI had room to improve. Then, we study two design choices in MAD --- AI persona and human engagement level --- finding a devil's advocate persona to lower accuracy and user experience, while humans' active participation in MAD improved accuracy in one task at the cost of frustration. We conclude by discussing implications for when, why, and how multi-agent deliberation can support human decision-making.
BibTeX
@misc{tian2026understanding,
  author = {Harry Yizhou Tian and Ming Yin},
  title = {{Understanding the Effects of Multi-Agent Deliberation for Human-AI Decision-Making}},
  year = {2026},
  eprint = {2611.14901},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.14901},
}
Figure