Collective Intelligence or Collective Bias? Understanding and Designing Deliberation Among AI Agents
Dr Taha Mansouri
Applications accepted all year round
Self-Funded PhD Students Only
About the Project
Background and rationale. Multi-agent large language model (LLM) systems are increasingly used for reasoning, evaluation, simulation and decision support. However, adding more agents does not automatically create collective intelligence. Recent work shows that majority voting can explain much of the benefit attributed to debate, while conformity, identity cues, authority, persuasive style, confidence signalling and communication topology can push agent groups toward incorrect or prematurely…
narrow conclusions. Other studies show that diversity-aware selection, anonymisation, anti-conformity and structured confidence exchange can improve outcomes. The field therefore needs more than another fixed debate protocol. It needs an integrated account of how agent interaction works, how bias and useful correction emerge in the dialogue, and how systems should adapt to different tasks, models and users.
Aim. This PhD will develop a mixed-method science and engineering framework for understanding and improving deliberation among AI agents. The central question is deliberately broad but coherent: when does interaction produce useful collective reasoning, and when does it amplify bias, overconfidence, persuasion or diversity collapse? The student will investigate both agent-agent deliberation and, where appropriate, how people interpret or rely on apparently collective AI advice.
Research themes and questions. The project offers four connected strands. First, behavioural dynamics: which patterns resembling conformity, anchoring, authority bias, group polarisation, confirmation bias, social loafing or minority influence appear in AI collectives, and how stable are they across model families and domains? Second, reasoning process: which features of a transcript distinguish genuine error correction from imitation, rhetorical success or accidental convergence? Third, system design: how should roles, evidence access, identity, memory, speaking order, network structure, stopping rules and aggregation be adapted to preserve productive disagreement while controlling cost? Fourth, human-facing implications: how do consensus, dissent, confidence and argument presentation affect human trust, judgement and willingness to escalate a decision?
Mixed-method methodology. The computational strand will use controlled experiments across open and, where justified, closed models. Factorial interventions may vary model diversity, persona, task framing, evidence allocation, confidence visibility, identity or anonymisation, communication graph, debate rules and adversarial participants. Evaluation will include accuracy, calibration, robustness, fairness or cultural bias, semantic diversity, correction and harm rates, token use and latency. Statistical analysis may use mixed-effects models, network analysis, uncertainty estimation and causal inference where the design and assumptions support causal claims.
Qualitative and human-centred analysis. The qualitative strand will examine purposively sampled deliberation traces rather than reducing every interaction to a single score. The student may develop a transparent coding framework based on argumentation quality, evidence use, stance change, dissent, persuasion and failure mechanisms; combine thematic or discourse analysis with computational text analysis; and validate automated coding against human annotation and inter-rater agreement. Comparative case studies and process tracing will explain why similar numerical outcomes can arise through very different reasoning dynamics. An optional human-participant strand could study how users respond to multi-agent consensus and dissent, subject to ethics approval and feasible recruitment.
Design and evaluation. Findings from the empirical strands will guide new or adaptive deliberation mechanisms. Candidate directions include evidence-provenance displays, blind or anonymised review, protected minority reports, dynamic role assignment, selective communication, argument maps, confidence-aware aggregation, adversarial auditing, abstention and human escalation. The goal is not to claim one universal protocol. The project will identify which mechanisms work, for whom, under which task and model conditions, and at what computational or social cost. Strong baselines, pre-specified confirmatory tests, ablations and cross-model validation will be expected.
Expected contributions. Outputs may include an empirically grounded taxonomy of deliberation biases and correction mechanisms; an annotated corpus of agent interactions; an expanded benchmark and reproducible experiment platform; adaptive or bias-resilient deliberation algorithms; and practical guidance for auditing collective AI systems. The project is intended to generate publishable work across trustworthy AI, natural language processing, multi-agent systems and human-centred AI.
Candidate profile. Applicants should have strong analytical and programming skills and a background in machine learning, natural language processing, data science, human-computer interaction, psychology, statistics or a related field. Experience across every method is not expected. The project suits a candidate who is comfortable combining rigorous experiments with close analysis of how reasoning unfolds, and who wants room to shape a distinctive interdisciplinary thesis.
Funding Notes
To inquire about University of Salford funding schemes – including the Widening Participation Scholarship – visit this website: https://www.salford.ac.uk/doctoral-school/phd-studentships.
References
Choi, H. K., Zhu, X., and Li, S. (2025). Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? Advances in Neural Information Processing Systems 38.
Choi, M., Kim, K., Chae, S., and Baek, S. (2025). An Empirical Study of Group Conformity in Multi-Agent Systems. Findings of the Association for Computational Linguistics: ACL 2025, 5123-5139. doi:10.18653/v1/2025.findings-acl.265.
Ki, D., Rudinger, R., Zhou, T., and Carpuat, M. (2025). Multiple LLM Agents Debate for Equitable Cultural Alignment. Proceedings of ACL 2025, 24841-24877. doi:10.18653/v1/2025.acl-long.1210.
Choi, H. K., Zhu, J., and Li, S. (2026). When Identity Skews Debate: Anonymization for Bias-Reduced Multi-Agent Reasoning. Proceedings of ACL 2026, 14284-14311. doi:10.18653/v1/2026.acl-long.650.
Zhu, X., Zhang, C., Chi, Y., Stafford, T., Collier, N., and Vlachos, A. (2026). Demystifying Multi-Agent Debate: The Role of Confidence and Diversity. Findings of ACL 2026, 33909-33930. doi:10.18653/v1/2026.findings-acl.1694.
Ko, C., Shin, J., Song, H., Lee, H., Hwang, E. J., and Park, J. C. (2026). Social Dynamics as Critical Vulnerabilities that Undermine Objective Decision-Making in LLM Collectives. Proceedings of ACL 2026, 37865-37890. doi:10.18653/v1/2026.acl-long.1756.
Chen, N., Tong, Y., Yang, Y., He, Y., Zhang, X., Qingyun, Z., Wang, Q., and He, B. (2026). Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation. Findings of ACL 2026, 251-306. doi:10.18653/v1/2026.findings-acl.13.
Turkstra, F., Nabhani, S., and Al Khatib, K. (2026). ARGSBASE: A Multi-Agent Interface for Structured Human-AI Deliberation. Proceedings of EACL 2026, 563-574. doi:10.18653/v1/2026.eacl-demo.39.
This is a Preview Listing…
You must sign in to see the full job description, and to apply.
Manage / Upgrade this job to a Full Job Listing.
Find Your Best Opportunity
Tell them AcademicJobs.com sent you!