Harry Yizhou Tian and Ming Yin on Multi-Agent Deliberation

Harry Yizhou Tian and Ming Yin on Multi-Agent Deliberation

Authors Harry Yizhou Tian and Ming Yin report two user studies of multi-agent deliberation (MAD), where multiple LLMs engage in structured discussion. They tested LLM-response-verification and fact-checking, comparing MAD with other AI decision-support methods. MAD outperformed single-round AI assistance on both tasks; the authors say multi-round interaction was the main driver, while a second AI added value when the single AI had room to improve. The design choices also mattered: a devil’s advocate persona lowered accuracy and user experience, while active human participation improved accuracy in one task at the cost of frustration. The results make MAD more than a question of adding more AI voices: how the discussion is structured and how people take part may affect both decision accuracy and the experience of using the system. The abstract reports results for these two tasks; it does not establish whether the same effects hold in other settings.

Publication: 1c288951-5aad-4543-8308-38dff9c714c2

attribution — pass: The reading correctly credits the findings to authors Harry Yizhou Tian and Ming Yin and presents them as the authors' reported results, not as established facts.

Authors:Harry Yizhou Tian, Ming Yin

summary — pass: The reading defines MAD in plain words, names both tasks and compares MAD with single-round assistance. It keeps the persona and engagement findings, including the frustration cost. The closing interpretation is hedged with 'may' and follows the abstract. The question asks about generalization, which the abstract leaves open. A small optional repair is to add the abstract's motivation: MAD's benchmark gains had not been tested for whether they help human decisions. LLM and LLM-response-verification are left unexplained. These could use term-library tooltips, but they do not block understanding.

improves AI performance on many benchmarks, but whether it helps humans make better decisions remains untested

date — pass: The originalPublished date of 2026-11-11 matches the v1 submission date shown on the arXiv page.

[Submitted on 11 Nov 2026]

support — pass: Every claim matches the abstract word for word or nearly so: two user studies, MAD outperforming single-round assistance, multi-round interaction as the main driver, the conditional value of a second AI, and the persona and engagement effects. The limitations correctly state that only the abstract is available and that results may not extend beyond these tasks. No benefits or costs were invented.

a second AI added value when the single AI had room to improve

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct way for readers to reach the source.