JAM framework uses LLM-as-Judge for theory-agnostic personality inference

Personality recognition has long been shackled by rigid taxonomies like the Big Five, forcing models to fit predefined categories rather than discovering how traits actually manifest. A team of eight researchers from Malaysia, Japan, France, and Taiwan proposes a radical departure: JAM (Judge for Adaptive Metric-Alignment), a theory-agnostic framework that learns latent psychological structure directly from text. The core innovation is an Attention-Pooled Graph Prototypical Network that clusters textual embeddings into shared pseudo-facets, allowing the model to infer an individual's psychological profile without requiring theory-specific labels during training or inference. To handle the messiness of heterogeneous datasets annotated under different frameworks, JAM employs Cross-Theory Harmonization (CTH) with two complementary mechanisms: Human-Guided Linkage and Machine-Induced Consensus, which unify data without predefined labels. This lets JAM generalize across personality theories and even support low-resource theories that lack large tagged corpora.

To further tighten data quality, JAM integrates an LLM-as-a-Judge mechanism in two configurations: LLM-before-the-loop pre-screens ambiguous samples, while LLM-in-the-loop actively guides adaptive metric learning during training. This dual setup reduces noise from mislabeled or borderline examples—a common pitfall in subjective personality annotation. The framework achieved state-of-the-art cross-framework generalization in experiments, outperforming theory-constrained baselines on multiple personality benchmarks. The paper, published in IEEE Transactions on Affective Computing (2026), includes public code, model weights, and artifacts. JAM points toward a future where AI can understand personality flexibly, adapting to any cultural or contextual framework without being locked into Western psychological categories. For the AI industry, this is a step toward more nuanced, less biased user modeling—whether for chatbots, recommendation systems, or mental health tools.