UTS at PsyDefDetect: Multi-Agent Councils and Absence-Based Reasoning for Defense Mechanism Classification

Researchers from UTS achieved second place in a psychological defense mechanism classification competition using a multi-agent AI system that identifies defense patterns through absence-based reasoning rather than presence detection. The system combines Gemini 2.5 agents with fine-tuned Qwen models to achieve an F1 score of 0.406, addressing critical biases in minority class prediction through structured ensemble methods.

This research advances psychological AI classification by inverting traditional detection logic. Defense mechanisms are fundamentally absences—missing emotions, blocked thoughts, denied reality—yet most machine learning systems optimize for presence signals. By encoding an affect-cognition integration spectrum as clinical rules, the UTS team added 11.4 percentage points to their F1 score, demonstrating that domain-specific reasoning frameworks outperform raw model scaling. The multi-agent architecture reflects a broader industry trend toward deliberative AI systems where specialized agents evaluate evidence strength rather than applying majority voting, reducing overconfidence in well-represented classes.

The research reveals a critical vulnerability in language models: the "L7 attractor" phenomenon where minority classes systematically collapse into majority class predictions with high confidence. This pattern appears across emotion-based classification tasks, suggesting the problem extends beyond this specific domain. The team's solution—a multi-agent builder-critic-regression-guard system that generated more F1 improvement in one iteration than eight prior attempts combined—demonstrates that structured oversight mechanisms can efficiently identify and correct systematic model failures.

For AI practitioners, this highlights the value of adversarial analysis within ensemble architectures. Rather than treating model disagreement as noise, the fine-tuned Qwen override system weaponizes disagreement through targeted interventions. The absence-based reasoning framework also has implications for clinical AI deployments where false confidence in wrong predictions poses greater risk than missed predictions. Future work should investigate whether this attractor phenomenon applies to other minority-class scenarios and whether similar structured override systems could improve performance across psychological, medical, and safety-critical domains.

• →Defense mechanisms are defined by absences rather than presences, requiring inverted classification logic that traditional models struggle to execute naturally.
• →Multi-agent systems with role-specific evaluation (advocates rating evidence strength) outperformed single-model approaches and achieved top-5 results without fine-tuning.
• →The L7 attractor reveals systematic bias where minority classes default to majority labels with high confidence, affecting at least 59-80% of minority predictions.
• →Structured multi-agent oversight (builder-critic-regression-guard) produced faster optimization gains than eight previous iterations, suggesting ensemble design matters more than scale.
• →Domain-specific clinical rule encoding added the largest single performance improvement, demonstrating that expert knowledge integration remains critical for specialized classification tasks.