Adaptive Trait Elicitation in Multi-Agent Systems for Improved Reasoning
Supervisor: Dr Debasis Ganguly
Industry Partner: Adobe Research India
School: Computing Science
Description:
This project proposes a new paradigm for multi-agent large language model (LLM) systems by integrating psychologically grounded personality traits into role-based agent architectures. While recent multi-agent systems have demonstrated improved reasoning through structured role differentiation (e.g., adversarial legal advocates, code reviewers, hypothesis critics), these systems typically assume behavioral homogeneity across agents. The proposed research hypothesizes that controlled behavioral diversity—implemented through explicit personality conditioning—can further enhance deliberative quality, robustness, and objective task performance.
The core innovation lies in combining functional role specialization with adaptive personality profiling. Each agent will be instantiated not only with a task role (e.g., judge, advocate, coder, reviewer) but also with a personality configuration derived from established psychological frameworks such as DISC or the Big Five. These traits are modelled as distributions over personality dimensions and are expected to influence reasoning style, argumentative strategy, and interaction dynamics. For example, pairing dominant advocates with highly conscientious judges in legal reasoning, or combining creative coders with meticulous reviewers in software development, may improve outcome reliability and depth of analysis.
The research will proceed in two major phases. First, we will investigate static trait elicitation mechanisms, including prompt-based conditioning, supervised fine-tuning, and reinforcement learning approaches, to produce psychologically validated agent archetypes. These archetypes will be evaluated using behavioural probing methodologies to assess trait consistency during task-driven interaction .
Second, the project will develop a data-driven adaptive trait selection framework. A supervised mapping will be learned to predict optimal personality distributions based on input instance characteristics and agent role. This mapping may be trained using downstream task rewards (e.g., legal judgment accuracy, code correctness) or proxy signals such as diversity of reasoning pathways and calibration measures . Reinforcement learning variants will be explored to optimise personality allocation dynamically within multi-agent deliberation loops.
Evaluation will span legal judgment prediction, multi-agent software engineering tasks, and complex reasoning benchmarks such as long-form question answering and scientific hypothesis testing . Outcomes will include a personality-aware multi-agent framework, a validated library of agent archetypes, and empirical insights into artificial collaboration and emergent team dynamics.