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

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

Galat and Rizoiu's system for classifying psychological defense mechanisms in emotional support dialogues secured second place (F1 0.406) among 64 teams in the PsyDefDetect shared task. A core principle of their approach is that defense mechanisms are characterized by absence, such as missing affect or denied reality. This insight is translated into prompt-level clinical rules, which model an affect-cognition integration spectrum and contributed the largest F1 gain of +11.4 percentage points. The system's architecture features a multi-phase deliberative council of Gemini 2.5 agents, where advocates assess evidence strength for specific classes, achieving an F1 of 0.382 without fine-tuning. Despite this, the council exhibited a systematic "L7 attractor," leading to 59–80% incorrect predictions for minority classes. To address this, a targeted override ensemble, comprising three fine-tuned Qwen3.5 models, applies 16 overrides, adding +2.4 percentage points to the F1 score. This ensemble is managed by a multi-agent system including a builder, critic, and regression guard.

Key takeaway

For NLP Engineers developing systems for psychological text analysis, particularly when classifying nuanced or minority defense mechanisms, you should consider integrating absence-based reasoning into your prompt engineering. Your multi-agent architectures, like the Gemini 2.5 council, can achieve strong baseline performance, but be prepared to implement targeted override ensembles with fine-tuned models, such as Qwen3.5, to mitigate systematic biases against minority classes and significantly improve overall F1 scores.

Key insights

Defense mechanisms are defined by what is absent, which can be modeled for classification.

Principles

• Absence defines defense mechanisms.
• Multi-agent councils improve classification.
• Address minority class bias.

Method

A multi-phase deliberative council of Gemini 2.5 agents rates evidence, followed by a Qwen3.5 override ensemble selected by a builder, critic, and regression guard system.

In practice

• Encode absence as clinical rules.
• Use class-specific agent advocates.
• Apply targeted overrides for minority classes.

Topics

• Psychological Defense Mechanisms
• Multi-Agent Systems
• Absence-Based Reasoning
• Gemini 2.5
• Qwen3.5
• Minority Class Classification

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• AI Coding Agents in Social Science: Methodologically Diverse, Empirically Consistent, Interpretively Vulnerable — AI coding agents show methodological diversity but are vulnerable to interpretive bias, not estimation bias.
• Defending against Adaptive Prompt Injection Attacks via Reasoning-enabled Task Alignment — RETA defends against adaptive prompt injection by aligning LLM actions with user tasks via reasoning, outperforming pattern-based defenses.
• The Illusion of Multi-Agent Advantage — Automated Multi-Agent Systems often incur high costs for superficial complexity, failing to outperform simpler Single-Agent Systems.
• MetaCogAgent: A Metacognitive Multi-Agent LLM Framework with Self-Aware Task Delegation — Metacognitive multi-agent LLM systems improve task accuracy and efficiency by enabling self-assessment and adaptive delegation.
• Multi-Agent Prompt Engineering is Dead: NEW Proof Your Agents Aren't Talking — Multi-agent LLM prompt optimization is often unnecessary; focus on independent agent tuning and architectural harnesses.
• Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate — IMAD distills multi-agent debate into a single LLM, improving efficiency and enabling fine-grained behavioral control via agent-specific latent subspaces.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.