CodeX FutureLaw 2026: Graph-Constrained LLMs

Why It Matters

Hybrid graph‑constrained LLMs promise reliable, explainable AI for high‑risk legal tasks, accelerating adoption while meeting regulatory transparency demands.

Key Takeaways

• Graph-constrained LLMs combine neural models with knowledge graphs for legal reasoning.
• Symbolic AI offers explainability and determinism missing in pure LLMs.
• Hart’s adjudication theory guides rule‑discretion split in hybrid architecture.
• Recreating Gardner’s 1987 knowledge base in Neo4j achieved 89% accuracy.
• Neural‑symbolic system outperformed baseline LLM by tenfold, enabling error tracing.

Summary

The CodeX FutureLaw 2026 talk introduced graph‑constrained large language models (LLMs) that fuse neural language processing with symbolic knowledge graphs to emulate legal reasoning. Presented by LSE PhD candidate Zarja Hude, the research targets the opacity and unreliability of pure LLMs in high‑stakes legal contexts, proposing a hybrid architecture grounded in jurisprudential theory.

Hude highlighted two core deficiencies of standalone LLMs: lack of explainability and inconsistent outputs across identical queries. By contrast, symbolic expert systems provide deterministic, traceable decisions but suffer from scalability challenges. Drawing on H.L.A. Hart’s balanced adjudication model—rules handle easy cases while judicial discretion fills interpretive gaps—she designed a system where a knowledge‑graph enforces rule‑based constraints and the LLM supplies contextual discretion.

The prototype rebuilt Anne Gardner’s 1987 contract‑analysis knowledge base in a Neo4j property graph, integrating it at critical reasoning stages. When tested on Gardner’s original contract scenario, the hybrid system achieved 89% correct reasoning versus under 9% for a pure LLM baseline, and it enabled granular error tracing unavailable in black‑box models. The work earned multiple awards, including best paper at the JURIX doctoral consortium.

These results suggest that neural‑symbolic integration can dramatically improve accuracy, transparency, and auditability of AI‑driven legal tools, paving the way for scalable, trustworthy applications in compliance, litigation support, and regulatory analysis.

Original Description

Pure LLMs get legal reasoning wrong 91% of the time. By combining them with knowledge graphs, that number flips to 89% accuracy — and that gap changes everything. In this talk, London School of Economics PhD researcher Zarja Hude reveals how grounding AI in centuries-old legal philosophy — specifically H.L.A. Hart's theory of easy vs. hard cases — unlocks a fundamentally more reliable approach to legal AI. Instead of letting language models hallucinate their way through complex legal logic, her neural-symbolic architecture uses knowledge graphs to constrain the AI to follow legal rules, only invoking the model's judgment where genuine interpretive discretion is required. The result: traceable, explainable, deterministic legal reasoning that standalone LLMs simply cannot deliver. If you think AI is already good enough for high-stakes legal work, this talk will make you think again.

Zarja Hude - PhD Candidate, London School of Economics

MC: Barclay Blair - DLA Piper - Senior Managing Director, AI Innovation Lead