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

Abstract: Legal outcome prediction must disentangle objective case facts from adjudicative context. Merit-based rulings rely on factual evidence while technical disposals may hinge on judicial discretion. We propose a Judge-Aware Gated Multi-Task Learning architecture that explicitly models this distinction. We introduce a fine-grained outcome taxonomy to supervise the encoder, enforcing a structural regularization that disentangles distinct semantic pathways. This granular legal curriculum enables our Gated Fusion mechanism to dynamically modulate reliance on judge identity. We evaluate our approach on 13,937 UK Employment Tribunal decisions. We benchmark our design against supervised fine-tuning (SFT) of a Gemma-4 26B-A4B backbone, in which judge identity and the taxonomy are injected as prompt tokens or autoregressive output targets. The two contextual signals compose only weakly when forced through a single autoregressive channel. In contrast, coupling a LoRA-adapted Gemma-4 encoder with our gated architecture defines a new state of the art on this benchmark while requiring an order of magnitude fewer trainable parameters than the generative SFT baselines, with gains concentrated on the most ambiguous and rarest outcome classes. Beyond accuracy, the architecture is interpretable; learned judge embeddings and calibration profiles localize the cases where adjudicative context drives the prediction. These results indicate that, for identity-conditioned classification of legal outcomes, the choice of conditioning interface dominates scale: differentiable structured composition yields more accurate, more parameter-efficient models than prompt-based composition over a substantially larger backbone.

• The paper demonstrates a novel gated multi-task learning model that separates factual legal reasoning from adjudicative discretion.
• It leverages a hybrid Gemma-4 encoder with judge identity embeddings to achieve a 5.1 macro-F1 point gain on rare and ambiguous outcomes.
• The study highlights that structured conditioning of judicial identity enhances both interpretability and parameter efficiency in legal AI.

Quantifying Judicial Discretion in Legal Outcome Prediction Using Gated Multi-Task Learning

Introduction

The challenge of predicting court rulings automatically in the legal domain requires modeling nuanced interactions between facts, adjudicative rules, and discretionary variance. The paper "Towards Explainable Adjudicative Variance: Quantifying Judicial Discretion via Gated Multi-Task Learning" (2606.27069) directly addresses the need for compositional models that separate fact-driven legal reasoning from procedural and judge-driven variance, with a specific focus on UK Employment Tribunal (UKET) decisions. The authors introduce a comprehensive legal outcome taxonomy and a judge-aware gated multi-task learning (MTL) architecture. Their results highlight the importance of the conditioning interface in achieving interpretable, accurate, and parameter-efficient legal judgment prediction, with strong performance on rare and ambiguous outcome classes.

Motivation and Problem Formalization

Legal Judgment Prediction (LJP) systems must account for the distinct roles of factual evidence and procedural rules. In the UKET context, outcomes hinge not only on merits but often on technical disposals, heavily influenced by judicial discretion. Existing LLM approaches tend to treat all signals as prompt/context tokens, which forces divergent legal and adjudicative pathways through a single generative output—leading to entangled, uninterpretable models and weak composition of judge identity and fine-grained outcomes.

The paper formalizes the task as a hierarchical classification problem, leveraging both a coarse-grained four-way label (GCO: Claimant Wins, Loses, Partly Wins, or Other) and an auxiliary fine-grained taxonomy (DCO) that decouples substantive legal resolution from technical/procedural terminators. The core scientific question is how to architect models such that judicial discretion is explicitly and explainably captured, allowing for controlled conditioning and counterfactual intervention.

Gated Multi-Task Architecture and Fine-Grained Outcome Taxonomy

The authors propose a hybrid discriminative model built on a LoRA-adapted Gemma-4 26B-A4B encoder. The judge identity signal is modeled as a learned embedding, fused with legal textual features via a dynamically activated gated mechanism. This Gated Fusion module consults judge identity only when outcome semantics require discretionary evaluation (e.g., ambiguous mixed outcomes or technical defaults), while learning to ignore it for factually clear or formulaic cases. The primary legal outcome is supervised at both the GCO and DCO levels, with the auxiliary multi-class DCO head providing structural regularization.

The DCO taxonomy, derived in consultation with legal experts, defines 11 functional outcome classes split between merit-based determinations and non-merit-based disposals, plus mixed and additional procedural classes. This taxonomy is extracted using LLM-assisted labeling at scale, and used as both supervision target and semantic prototype in the model’s label-wise attention mechanism.

Empirical Evaluation and Ablative Analysis

Experiments are conducted on a curated set of 13,937 UKET cases, with strict judge-stratified train/validation/test splits. Macro-F1 (with per-class breakouts) and Weighted-F1 are reported, focusing on the ability to recover rare/ambiguous outcomes (e.g., "Partly Wins", "Other") which are most susceptible to adjudicative variance.

Three model families are compared:

1. D-Track: A ModernBERT-based discriminative model with variable MTL and conditioning interfaces.
2. G-Track: A suite of generative Gemma-4 SFT baselines, with judge identity and DCO supervision supplied as prompt/output tokens in various configurations.
3. B-Track (Hybrid): The proposed architecture, using Gemma-4 as encoder with a structured, judge-aware discriminative head.

Key findings include:

• Simply prompting with judge identity and fine-grained label supervision is sub-additive in generative LLMs; their interaction via a single autoregressive channel collapses the intended compositional pathways, plateauing performance.
• By contrast, architecturally routing judge identity as a learnable embedding and supervising a parallel DCO head restores additive composition, yielding strong gains in parameter efficiency and interpretability.
• The hybrid B2 model attains 65.21 macro-F1, a +5.1 point improvement over the strongest generative SFT variant, with an order-of-magnitude fewer trainable parameters.
• Gains are concentrated on non-monolithic, rare, and fuzzy outcome classes—precisely where legal reasoning and procedural discretion intersect.

Interpretability and Probing Adjudicative Variance

Beyond accuracy, the architecture supports robust interpretability diagnostics. The model’s sensitivity to counterfactual judge swaps is measured via the KL divergence of output distributions when the judge embedding is perturbed. The structured head displays marked shifts in outcome predictions for cases where judicial discretion is known to be outcome-determinative, while maintaining invariance where facts dominate.

Latent judge embeddings, analyzed via t-SNE and clustering, recover semantically coherent behavioral territories: clusters align with merit-dominant judges, proceduralists, and specialists in various case-management functions. The gate activation provides a direct interpretable audit trail, showing when and how judge context influences model predictions.

Implications for Legal AI and Structured Modeling

The results underscore that in domains with intrinsic identity-conditioned outcome variance (e.g., judicial, clinical, demographic-aware contexts), the choice of conditioning interface—gradient-accessible parameterization versus prompt token—dominates the effect of scaling backbone models. This has both practical and theoretical implications:

• Parameter-Efficient Specialization: Prompt-based composition in LLMs cannot reliably surface the downstream effect of high-variance contextual factors; leveraging learnable embeddings within a multi-task structure not only improves accuracy but provides transparent surrogates for discretionary influence.
• Compositional Bottlenecks: Autoregressive generation by construction enforces output consistency, which can inadvertently collapse distinct reasoning pathways. Parallel structured heads and differentiable conditioning interfaces restore compositional additivity and robustness.
• Design Heuristics: Variables whose effect must be interpreted, counterfactually manipulated, or audited (e.g., judicial identity) are better handled upstream of the model’s output layer, avoiding the absorption of such signals into 'black box' context windows.

Limitations and Future Directions

The empirical scope is constrained to UKET cases, as required access to judge identities is not globally generalizable. The LLM-assigned DCO labels, while robust, are heuristic proxies, reflecting the practicalities of scaling legal annotation. Further transfer experiments across jurisdictions and LLM families, as well as the integration of richer behavioral and doctrinal features, are natural directions for extension.

Future work should examine the broader utility of gradient-parameterized identity signals, and the limits of explainable structured composition in more adversarial or less formulaic domains. Additionally, studying adversarial vulnerability in cases where factual and discretionary signals are entangled remains critical for responsible legal AI deployment.

Conclusion

This work decisively demonstrates that modeling adjudicative variance in legal outcomes—the quantification and audit of judicial discretion—demands explicit, structured, and differentiable composition of identity signals, rather than reliance on prompt-augmented generative LLMs. The proposed gated multi-task approach yields both strong empirical gains and transparent behavioral diagnostics, setting a new standard for AI applications in complex, actor-conditioned institutional domains.