Towards Explainable Adjudicative Variance: Quantifying Judicial Discretion via Gated Multi-Task Learning

Legal outcome prediction often conflates objective case facts with the discretionary influence of individual judges. The authors investigate whether modern large language models (LLMs) can effectively disentangle these factors, specifically comparing two conditioning interfaces: generative supervised fine-tuning (where judge identity and legal taxonomy are injected as prompt tokens) versus a structured, differentiable architecture that treats these signals as dynamic control variables.

The authors introduce a Judge-Aware Gated Multi-Task Learning architecture. They define a fine-grained taxonomy of 11 Detailed Case Outcomes (DCO) to supervise the encoder, distinguishing between merit-based rulings and technical disposals. The model uses a Label-Wise Attention Network (LWAN) to map document embeddings into outcome-specific subspaces. Crucially, they implement a Gated Fusion mechanism that dynamically modulates the influence of judge identity embeddings based on the case's legal context, effectively acting as a switch that ignores judge identity for clear-cut cases while prioritizing it for discretionary decisions.

The study evaluates these methods on 13,937 UK Employment Tribunal decisions. The results show that prompt-based composition of contextual signals is sub-additive, meaning that forcing judge identity and taxonomic supervision through a single autoregressive channel limits the model's performance. In contrast, the proposed hybrid architecture—coupling a LoRA-adapted Gemma-4 encoder with the gated structured head—outperforms the strongest generative baseline by 5.1 points in Macro-F1. The gains are most pronounced in rare and ambiguous outcome classes, confirming that the choice of conditioning interface is more critical than model scale for identity-conditioned legal classification.