‹
Back
 
10:15 - 11:45
Location: Multi-Function Room 1 (19/F LAU)
Chair/s:
Emilie Tran
Discussant: King-wa FU

Lisa Lechner - Detecting Absence in Large Language Models: Missing Precedents and Omitted Provisions in Legal Text
Runyi Ma - Normative Rigidity vs. Probabilistic Fragility: Identifying the Phase Transition of Systematic Failure in Large Language Models via Structured Legal Logic
Emilie Tran, Eric Sautede - Governing the Lingua Franca of Power: Middle Eastern Agency in the Regulation of Large Language Models
James Wong - Generative AI in Policymaking: Bias, Limitations, and Implications for Policy Change
Submission 49
Normative Rigidity vs. Probabilistic Fragility: Identifying the Phase Transition of Systematic Failure in Large Language Models via Structured Legal Logic
Panel 2-Multi-Function Room 1 (19/F LAU)-02
Presented by: Runyi Ma
Runyi Ma 1, 2, 3, 4, Jinting Deng 5
1 Center for Digital Rule of Law, Peking University
2 Law School, Renmin University of China
3 Beijing Association for Digital Economy and Digital Governance Rule of Law Studies
4 Peking University’s Analytics Lab for Global Risk Politics
5 Discipline Inspection and Supervision School, Renmin University of China
This research addresses the conflict between the normative rigidity of legal reasoning and the probabilistic fragility of LLMs. We leverage a high-fidelity dataset of thousands of statutory rules across 300 case causes, manually transformed by legal experts into "Chain-Graph JSON" structures. Utilizing a rigorous methodological framework, we quantify legal complexity through multiple statistical and topological metrics, including nesting depth (D), node density, and branching factors. Our empirical stress tests across leading models (GPT, Gemini, Claude, DeepSeek) reveal a distinct phase transition: beyond a critical complexity threshold, the Logical Alignment Score suffers an abrupt, non-linear collapse, regressing models into "stochastic parrots" with structural hallucinations. To mitigate this, we establish the "LL-Turing" Benchmark and demonstrate that Logical Rails—structured protocols for model constraints—are an ontological necessity for ensuring determinism. This study pivots Computational Social Science (CSS) from a data-driven to a "structure-logic dual-driven" paradigm, providing a robust empirical template for global risk governance.